From 3f20131bbb94ce87f1d283127c6d94071b8bbfe2 Mon Sep 17 00:00:00 2001
From: ahao-anyscale <ahao@anyscale.com>
Date: Tue, 28 Apr 2026 01:21:01 +0000
Subject: [PATCH 01/21] x

Signed-off-by: ahao-anyscale <ahao@anyscale.com>
---
 .../skyrl_train/skyrl_train_backend.py        |  12 +-
 .../remote_inference_client.py                | 138 +++++--
 .../skyrl_train/inference_servers/utils.py    |  31 +-
 .../skyrl_train/workers/fsdp/fsdp_worker.py   |   7 +-
 .../workers/megatron/megatron_worker.py       |   8 +-
 skyrl/backends/skyrl_train_backend.py         |  26 +-
 skyrl/benchmarks/load_test_concurrency.py     |   2 +-
 skyrl/train/config/config.py                  |  10 +
 skyrl/train/config/ppo_base_config.yaml       |   7 +
 skyrl/train/entrypoints/main_base.py          |  23 +-
 skyrl/train/generators/skyrl_gym_generator.py |   9 +-
 skyrl/train/generators/skyrl_vlm_generator.py |   5 +-
 .../test_multi_lora_serving.py                | 172 ++++++++
 .../test_new_inference_generation.py          |   4 +-
 ...t_remote_inference_client_chat_template.py |   2 +-
 .../test_vlm_inference_generation.py          |   3 +-
 .../gpu/gpu_ci/test_engine_generation.py      |   8 +-
 .../skyrl_train/gpu/gpu_ci/test_lora.py       |   5 +
 .../test_pause_and_continue_generation.py     |   2 +-
 .../gpu/gpu_ci/test_router_replay.py          |   1 +
 .../gpu/gpu_ci/test_skyrl_gym_generator.py    |   1 +
 .../gpu_ci/test_skyrl_vlm_gym_generator.py    |   1 +
 tests/backends/skyrl_train/gpu/utils.py       |  22 +-
 .../test_remote_inference_client.py           | 382 +++++++++++++++++-
 .../generators/test_skyrl_gym_generator.py    |  16 +-
 ...est_skyrl_gym_generator_chat_templating.py |   5 +-
 .../generators/test_skyrl_vlm_generator.py    |   3 +-
 27 files changed, 817 insertions(+), 88 deletions(-)
 create mode 100644 tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_multi_lora_serving.py

diff --git a/skyrl-agent/skyrl_agent/integrations/skyrl_train/skyrl_train_backend.py b/skyrl-agent/skyrl_agent/integrations/skyrl_train/skyrl_train_backend.py
index 8ad3c782c8..f40ca1c6e5 100644
--- a/skyrl-agent/skyrl_agent/integrations/skyrl_train/skyrl_train_backend.py
+++ b/skyrl-agent/skyrl_agent/integrations/skyrl_train/skyrl_train_backend.py
@@ -1,10 +1,18 @@
 from typing import Any, List
+
+from skyrl.backends.skyrl_train.inference_servers.utils import resolve_policy_model_name
+
 from ..base import AsyncInferBackend, GeneratorOutput, GeneratorInput
 
 
 class SkyRLBackend(AsyncInferBackend):
     def __init__(self, infer_engine, tokenizer: Any = None, cfg: Any = None):
         self.client = infer_engine
+        # Resolve the name the inference engine knows the policy by (base
+        # model or registered LoRA adapter) once at construction. Threaded
+        # into every ``client.generate`` call so the data plane never has
+        # to guess the target adapter.
+        self.policy_model_name = resolve_policy_model_name(cfg) if cfg is not None else self.client.model_name
 
     async def async_generate_prompts(self, prompts: Any, sampling_params: Any, **kwargs) -> List[str]:
         input_obj = {
@@ -12,7 +20,7 @@ async def async_generate_prompts(self, prompts: Any, sampling_params: Any, **kwa
             "session_ids": [kwargs.get("request_id", None)],
             "sampling_params": sampling_params,
         }
-        output = await self.client.generate(input_obj)
+        output = await self.client.generate(input_obj, model=self.policy_model_name)
         return output["responses"][0], output["stop_reasons"][0]
 
     async def async_generate_ids(self, input_ids: List[int], sampling_params: Any, **kwargs) -> List[str]:
@@ -21,7 +29,7 @@ async def async_generate_ids(self, input_ids: List[int], sampling_params: Any, *
             "session_ids": [kwargs.get("request_id", None)],
             "sampling_params": sampling_params,
         }
-        output = await self.client.generate(input_obj)
+        output = await self.client.generate(input_obj, model=self.policy_model_name)
         # todo(@csy) probably need to be finish_reason
         # https://github.com/vllm-project/vllm/blob/a0f8a7964694a6077689b242b5eca95de392d4bb/vllm/v1/engine/__init__.py#L22
         meta_info = {
diff --git a/skyrl/backends/skyrl_train/inference_servers/remote_inference_client.py b/skyrl/backends/skyrl_train/inference_servers/remote_inference_client.py
index a0c2198583..61f756976d 100644
--- a/skyrl/backends/skyrl_train/inference_servers/remote_inference_client.py
+++ b/skyrl/backends/skyrl_train/inference_servers/remote_inference_client.py
@@ -79,6 +79,9 @@
 
 _DATA_PLANE_RETRIES = 30
 
+SKYRL_LORA_ADAPTER_NAME = "skyrl-lora"
+"""Default LoRA adapter name used for single-LoRA training inside SkyRL."""
+
 _TINKER_SAMPLE_TO_VLLM_PARAM_MAP = {
     "temperature": "temperature",
     "max_tokens": "max_tokens",
@@ -111,6 +114,23 @@ def _extract_session_id_and_body(
     return session_id, clean_body
 
 
+def _require_model(body: Dict[str, Any], method_name: str) -> str:
+    """Return ``body["model"]``, raising ``ValueError`` if missing or empty.
+
+    The client deliberately does not fall back to any default — every caller
+    must explicitly identify the target base model or LoRA adapter so multi-
+    LoRA routing is unambiguous.
+    """
+    model = body.get("model")
+    if not model:
+        raise ValueError(
+            f"RemoteInferenceClient.{method_name}: request body must include a "
+            f"non-empty 'model' field identifying the target base model or "
+            f"LoRA adapter."
+        )
+    return model
+
+
 class PauseMode(Enum):
     """
     Pause mode for inference servers.
@@ -192,14 +212,20 @@ class RemoteInferenceClient:
     reports the full DP world size per server, so we divide by num_deployments."""
 
     model_name: str = "default"
-    """Model name for OpenAI-compatible API calls."""
+    """The base model identifier the inference server was started with.
+
+    Always the base model — never a LoRA adapter name. LoRA adapters are
+    addressed by the names callers register them under via
+    ``load_lora_adapter(name, path)``, and per-call routing is done by
+    passing that name as ``model`` on the data-plane methods.
+
+    Used internally only by ``tokenize``/``detokenize``, which are LoRA-
+    agnostic but still require a ``model`` field per the OpenAI schema.
+    """
 
     enable_return_routed_experts: bool = False
     """Whether to return routed expert indices (R3 / rollout router replay)."""
 
-    active_lora_name: Optional[str] = None
-    """Name of the active LoRA adapter. If set, generation requests use this adapter instead of the base model."""
-
     tokenizer: Optional[Any] = None
     """Optional HF tokenizer for local tokenize/detokenize (avoids HTTP round-trips)."""
 
@@ -314,6 +340,7 @@ async def _post(self, url: str, json: Dict[str, Any], headers: Optional[Dict[str
     async def generate(
         self,
         input_batch: InferenceEngineInput,
+        model: str,
     ) -> InferenceEngineOutput:
         """
         Generate completions via /v1/completions.
@@ -329,6 +356,10 @@ async def generate(
 
         Args:
             input_batch: Contains prompt_token_ids, sampling_params, and optional session_ids.
+            model: Required model identifier — the base model name or a loaded
+                LoRA adapter name. Callers must always supply this; the client
+                does not fall back to any default. Use ``self.model_name`` if
+                you want the client's configured default.
 
         Returns:
             InferenceEngineOutput with responses, response_ids, and stop_reasons.
@@ -367,6 +398,7 @@ async def _throttled_generate(idx: int) -> Dict[str, Any]:
                     sampling_params=sampling_params,
                     session_id=session_ids[idx] if session_ids and idx < len(session_ids) else None,
                     mm_features=mm_features[idx] if mm_features and idx < len(mm_features) else None,
+                    model=model,
                 )
             async with gen_sem:
                 return await self._generate_single(
@@ -374,6 +406,7 @@ async def _throttled_generate(idx: int) -> Dict[str, Any]:
                     sampling_params=sampling_params,
                     session_id=session_ids[idx] if session_ids and idx < len(session_ids) else None,
                     mm_features=mm_features[idx] if mm_features and idx < len(mm_features) else None,
+                    model=model,
                 )
 
         async def _throttled_detokenize(token_ids: List[int]) -> str:
@@ -401,6 +434,7 @@ async def _generate_single(
         prompt_token_ids: List[int],
         sampling_params: Dict[str, Any],
         session_id: Optional[Any],
+        model: str,
         mm_features: Optional[MultiModalFeatures] = None,
     ) -> Dict[str, Any]:
         """
@@ -419,12 +453,9 @@ async def _generate_single(
             else f"{self.proxy_url}/inference/v1/generate"
         )
 
-        # Use LoRA adapter name if one is active, otherwise use base model name
-        effective_model = self.active_lora_name if self.active_lora_name else self.model_name
-
         payload: dict[str, Any] = {
             "sampling_params": sampling_params,
-            "model": effective_model,
+            "model": model,
             "token_ids": prompt_token_ids,
         }
         if mm_features:
@@ -460,6 +491,7 @@ async def _render_for_sample(
         self,
         prompt: Dict[str, Any],
         session_id: Optional[str],
+        model: str,
     ) -> Tuple[List[int], Optional[MultiModalFeatures]]:
         """Build token_ids and optional multi-modal features from a Tinker prompt.
 
@@ -491,10 +523,9 @@ async def _render_for_sample(
                 url = c["location"]
             content_parts.append({"type": "image_url", "image_url": {"url": url}})
 
-        effective_model = self.active_lora_name if self.active_lora_name else self.model_name
         render_payload: Dict[str, Any] = {
             "json": {
-                "model": effective_model,
+                "model": model,
                 "messages": [{"role": "user", "content": content_parts}],
             }
         }
@@ -542,7 +573,10 @@ async def _render_for_sample(
 
         return final_token_ids, adjusted_features
 
-    async def sample(self, request_payload: SampleRequestPayload) -> SampleResponse:
+    async def sample(
+        self,
+        request_payload: SampleRequestPayload,
+    ) -> SampleResponse:
         """
         Sample completions via /inference/v1/generate (Tinker API).
 
@@ -551,13 +585,14 @@ async def sample(self, request_payload: SampleRequestPayload) -> SampleResponse:
 
         Args:
             request_payload: SampleRequestPayload with {"json": <request-body>}.
-                Expected keys in json: prompt, num_samples, sampling_params, session_id,
-                include_prompt_logprobs (bool), topk_prompt_logprobs (int).
+                Expected keys in json: model, prompt, num_samples, sampling_params,
+                session_id, include_prompt_logprobs (bool), topk_prompt_logprobs (int).
 
         Returns:
             SampleResponse with type="sample", sequences list, prompt_logprobs, and topk_prompt_logprobs.
         """
         session_id, body = _extract_session_id_and_body(request_payload)
+        model = _require_model(body, "sample")
 
         prompt = body.get("prompt", {})
         num_samples = body.get("num_samples", 1)
@@ -575,7 +610,7 @@ async def sample(self, request_payload: SampleRequestPayload) -> SampleResponse:
 
         # Render prompt: flatten text tokens and, if images are present,
         # call the render endpoint to get placeholder tokens + features.
-        token_ids, mm_features = await self._render_for_sample(prompt, session_id)
+        token_ids, mm_features = await self._render_for_sample(prompt, session_id, model=model)
 
         # Map Tinker SamplingParams → vLLM format
         sampling_params: Dict[str, Any] = {
@@ -590,11 +625,9 @@ async def sample(self, request_payload: SampleRequestPayload) -> SampleResponse:
             if val is not None:
                 sampling_params[vllm_key] = val
 
-        effective_model = self.active_lora_name if self.active_lora_name else self.model_name
-
         payload: Dict[str, Any] = {
             "sampling_params": sampling_params,
-            "model": effective_model,
+            "model": model,
             "token_ids": token_ids,
         }
         if mm_features is not None:
@@ -673,13 +706,17 @@ async def chat_completion(
 
         Args:
             request_payload: Dict with {"json": <request-body>, "headers": <headers-dict>}.
-                The request body should be OpenAI-compatible chat completion request.
-                session_id can be included in json for consistent routing.
+                The request body must be an OpenAI-compatible chat completion
+                request and is required to include a ``model`` field
+                identifying the base model or a loaded LoRA adapter; the
+                client does not fall back to any default. ``session_id`` can
+                be included in the body for consistent routing.
 
         Returns:
             OpenAI-compatible chat completion response.
         """
         session_id, body = _extract_session_id_and_body(request_payload)
+        _require_model(body, "chat_completion")
 
         headers = {"Content-Type": "application/json"}
         if session_id:
@@ -709,6 +746,7 @@ async def render_chat_completion(
             Rendered chat completion response (template-applied prompt and token IDs).
         """
         session_id, body = _extract_session_id_and_body(request_payload)
+        _require_model(body, "render_chat_completion")
 
         headers = {"Content-Type": "application/json"}
         if session_id:
@@ -738,6 +776,7 @@ async def completion(
             OpenAI-compatible completion response.
         """
         session_id, body = _extract_session_id_and_body(request_payload)
+        _require_model(body, "completion")
 
         headers = {"Content-Type": "application/json"}
         if session_id:
@@ -995,7 +1034,7 @@ async def update_named_weights(
         """
         Update model weights via vLLM native /update_weights. Used for full parameter fine-tuning.
 
-        For LoRA weight sync, use update_lora_from_disk() instead.
+        For LoRA weight sync, use load_lora_adapter() instead.
 
         Args:
             update_info: Dict with keys expected by vLLM (names, dtype_names, shapes, packed, etc.)
@@ -1080,33 +1119,34 @@ async def finish_weight_update(self) -> Dict[str, Any]:
             {"method": "finish_weight_update"},
         )
 
-    async def update_lora_from_disk(
+    async def load_lora_adapter(
         self,
+        lora_name: str,
         lora_path: str,
+        load_inplace: bool = True,
     ) -> Dict[str, Any]:
         """
-        Update LoRA adapter weights by loading from disk on all backend servers via /v1/load_lora_adapter.
+        Load (or reload) a LoRA adapter on all backend servers via /v1/load_lora_adapter.
 
-        Always loads under self.active_lora_name so the same slot is reused across
-        weight syncs.
-
-        After loading, generation requests will automatically use the LoRA adapter
-        by setting the model name to the LoRA adapter name.
+        After loading, generation/chat/completion requests can target this LoRA
+        by passing ``model=lora_name`` (or by setting it as the client's
+        ``model_name``). When ``load_inplace=True`` and an adapter with the
+        same name is already registered on the server, vLLM replaces it
+        inplace, preserving its internal int id.
 
         Args:
+            lora_name: Name to register the adapter under on each server.
             lora_path: Path to the LoRA adapter on disk (must be accessible from servers).
+            load_inplace: When True (default), reloading a previously-loaded and cached
+                adapter with the same name replaces it with the on-disk lora.
 
         Returns:
             Dict mapping server_url to response.
         """
-        if self.active_lora_name is None:
-            raise ValueError("active_lora_name must be set on RemoteInferenceClient before loading a LoRA adapter.")
-
-        lora_name = self.active_lora_name
         payload = {
             "lora_name": lora_name,
             "lora_path": lora_path,
-            "load_inplace": True,
+            "load_inplace": load_inplace,
         }
 
         # Call /v1/load_lora_adapter on all servers directly.
@@ -1129,6 +1169,40 @@ async def _load_on_server(server_url: str):
 
         return {url: resp for url, resp in results}
 
+    async def unload_lora_adapter(self, lora_name: str) -> Dict[str, Any]:
+        """
+        Unload a previously-loaded LoRA adapter on all backend servers via /v1/unload_lora_adapter.
+
+        After unloading, ``lora_name`` is no longer accepted as a ``model``
+        target on any server. The underlying CPU/GPU LRU entries on vLLM age
+        out naturally as new adapters are loaded.
+
+        Args:
+            lora_name: Name of the adapter to unload.
+
+        Returns:
+            Dict mapping server_url to response.
+        """
+        payload = {"lora_name": lora_name}
+
+        # Mirror load_lora_adapter: vLLM returns plain text on success and JSON
+        # ErrorResponse (e.g. 404) on failure.
+        session = await self._get_session()
+
+        async def _unload_on_server(server_url: str):
+            url = f"{server_url}/v1/unload_lora_adapter"
+            async with session.post(url, json=payload) as resp:
+                if resp.status >= 400:
+                    body = await resp.json()
+                    raise_for_status(resp, body)
+                return server_url, {"status": resp.status, "body": await resp.text()}
+
+        results = await asyncio.gather(*[_unload_on_server(url) for url in self.server_urls])
+
+        logger.info(f"Unloaded LoRA adapter '{lora_name}'")
+
+        return {url: resp for url, resp in results}
+
     # ---------------------------
     # Info
     # ---------------------------
diff --git a/skyrl/backends/skyrl_train/inference_servers/utils.py b/skyrl/backends/skyrl_train/inference_servers/utils.py
index 4f6fab74cf..47e039c3f8 100644
--- a/skyrl/backends/skyrl_train/inference_servers/utils.py
+++ b/skyrl/backends/skyrl_train/inference_servers/utils.py
@@ -7,6 +7,9 @@
 from skyrl.backends.skyrl_train.inference_servers.new_inference_worker_wrap import (
     VLLM_NEW_INFERENCE_WORKER_EXTENSION_CLS,
 )
+from skyrl.backends.skyrl_train.inference_servers.remote_inference_client import (
+    SKYRL_LORA_ADAPTER_NAME,
+)
 from skyrl.backends.skyrl_train.weight_sync import get_transfer_strategy
 from skyrl.train.config import (
     InferenceEngineConfig,
@@ -31,6 +34,27 @@ def _uses_lora_weight_sync(cfg: SkyRLTrainConfig) -> bool:
     return True
 
 
+def resolve_policy_model_name(cfg: SkyRLTrainConfig) -> str:
+    """Return the model identifier the inference engine knows the policy by.
+
+    Mirrors the weight-sync code path: when the worker registers a LoRA
+    adapter on the inference engine (FSDP + LoRA, or Megatron + LoRA with
+    ``merge_lora=False``), the policy is that adapter and callers must pass
+    ``SKYRL_LORA_ADAPTER_NAME`` as ``model`` on data-plane calls. Otherwise
+    — including Megatron + LoRA with ``merge_lora=True``, where merged
+    weights are pushed as a full weight update — the policy is the base
+    model itself.
+
+    This is the single source of truth for "which name does the inference
+    server know the policy by?" and should be used wherever a caller needs
+    to issue a ``generate``/``sample``/``chat_completion``/``completion`` /
+    ``render_chat_completion`` request against the current policy.
+    """
+    if _uses_lora_weight_sync(cfg):
+        return SKYRL_LORA_ADAPTER_NAME
+    return cfg.trainer.policy.model.path
+
+
 # TODO: Add a test for validation
 def build_vllm_cli_args(cfg: SkyRLTrainConfig) -> Namespace:
     """Build CLI args for vLLM server from config."""
@@ -84,9 +108,12 @@ def build_vllm_cli_args(cfg: SkyRLTrainConfig) -> Namespace:
     # LoRA adapters (not merged weights).  Megatron merges by default
     # (merge_lora=True), so the inference engine must NOT have LoRA wrapping.
     if _uses_lora_weight_sync(cfg):
+        lora_cfg = cfg.trainer.policy.model.lora
         args.enable_lora = True
-        args.max_lora_rank = cfg.trainer.policy.model.lora.rank
-        args.max_loras = 1
+        args.max_lora_rank = lora_cfg.rank
+        args.max_loras = lora_cfg.max_loras
+        if lora_cfg.max_cpu_loras is not None:
+            args.max_cpu_loras = lora_cfg.max_cpu_loras
         args.fully_sharded_loras = ie_cfg.fully_sharded_loras
 
         if not cfg.trainer.placement.colocate_all:
diff --git a/skyrl/backends/skyrl_train/workers/fsdp/fsdp_worker.py b/skyrl/backends/skyrl_train/workers/fsdp/fsdp_worker.py
index 4f3faf9426..377729c4f5 100644
--- a/skyrl/backends/skyrl_train/workers/fsdp/fsdp_worker.py
+++ b/skyrl/backends/skyrl_train/workers/fsdp/fsdp_worker.py
@@ -26,6 +26,9 @@
     fsdp_version,
     should_use_meta_init,
 )
+from skyrl.backends.skyrl_train.inference_servers.remote_inference_client import (
+    SKYRL_LORA_ADAPTER_NAME,
+)
 from skyrl.backends.skyrl_train.training_batch import (
     TrainingInputBatch,
     TrainingOutputBatch,
@@ -270,7 +273,9 @@ async def _save_lora_adapters_and_sync(self, peft_model, lora_sync_path, inferen
             )
 
             if isinstance(inference_engine_client, RemoteInferenceClient):
-                await inference_engine_client.update_lora_from_disk(lora_sync_path)
+                await inference_engine_client.load_lora_adapter(
+                    SKYRL_LORA_ADAPTER_NAME, lora_sync_path, load_inplace=True
+                )
             else:
                 lora_request = LoraLoadRequest(lora_path=lora_sync_path)
                 await inference_engine_client.update_named_weights(lora_request)
diff --git a/skyrl/backends/skyrl_train/workers/megatron/megatron_worker.py b/skyrl/backends/skyrl_train/workers/megatron/megatron_worker.py
index 456e4954f9..e00287dba3 100644
--- a/skyrl/backends/skyrl_train/workers/megatron/megatron_worker.py
+++ b/skyrl/backends/skyrl_train/workers/megatron/megatron_worker.py
@@ -32,6 +32,9 @@
     get_megatron_optimizer_param_scheduler,
     init_megatron_optim_config,
 )
+from skyrl.backends.skyrl_train.inference_servers.remote_inference_client import (
+    SKYRL_LORA_ADAPTER_NAME,
+)
 from skyrl.backends.skyrl_train.training_batch import (
     TrainingInputBatch,
     TrainingOutputBatch,
@@ -866,12 +869,15 @@ async def _save_lora_adapters_and_sync(self, lora_sync_path, inference_engine_cl
             with open(os.path.join(lora_sync_path, "adapter_config.json"), "w", encoding="utf-8") as f:
                 json.dump(adapter_config, f, ensure_ascii=False, indent=4)
 
+            # Send LoRA disk loading request to inference engine.
             from skyrl.backends.skyrl_train.inference_servers.remote_inference_client import (
                 RemoteInferenceClient,
             )
 
             if isinstance(inference_engine_client, RemoteInferenceClient):
-                await inference_engine_client.update_lora_from_disk(lora_sync_path)
+                await inference_engine_client.load_lora_adapter(
+                    SKYRL_LORA_ADAPTER_NAME, lora_sync_path, load_inplace=True
+                )
             else:
                 lora_request = LoraLoadRequest(lora_path=lora_sync_path)
                 await inference_engine_client.update_named_weights(lora_request)
diff --git a/skyrl/backends/skyrl_train_backend.py b/skyrl/backends/skyrl_train_backend.py
index ea0f42a8ef..a960b1ddee 100644
--- a/skyrl/backends/skyrl_train_backend.py
+++ b/skyrl/backends/skyrl_train_backend.py
@@ -27,6 +27,7 @@
 from skyrl.backends.skyrl_train.inference_servers.utils import (
     build_router_args,
     build_vllm_cli_args,
+    resolve_policy_model_name,
 )
 from skyrl.backends.skyrl_train.inference_servers.vllm_router import VLLMRouter
 from skyrl.backends.skyrl_train.training_batch import (
@@ -47,9 +48,6 @@
 from skyrl.utils.log import logger
 from skyrl.utils.tok import get_tokenizer
 
-# Fixed LoRA adapter name used for generation requests when LoRA is active.
-_SKYRL_LORA_ADAPTER_NAME = "skyrl-lora"
-
 
 class SkyRLTrainBackendOverrides(BaseModel, extra="allow"):
     """Configuration overrides for the SkyRL-Train backend.
@@ -381,17 +379,10 @@ def _create_new_inference_client(self):
                 f"proxy_url={proxy_url}, server_urls={server_urls}, colocated={is_colocated}"
             )
 
-        lora_cfg = self._cfg.trainer.policy.model.lora
-        active_lora_name = (
-            _SKYRL_LORA_ADAPTER_NAME
-            if lora_cfg and lora_cfg.rank > 0 and self._cfg.trainer.strategy != "megatron"
-            else None
-        )
         self._inference_engine_client = RemoteInferenceClient(
             proxy_url=proxy_url,
             server_urls=server_urls,
             model_name=self._cfg.trainer.policy.model.path,
-            active_lora_name=active_lora_name,
             data_parallel_size=ie_cfg.data_parallel_size,
             tokenizer=self._tokenizer,
         )
@@ -938,6 +929,13 @@ def _sample_with_remote_client(
     ) -> dict[str, types.SampleOutput | types.ErrorResponse]:
         """Sample using RemoteInferenceClient, forwarding model input chunks directly."""
 
+        # Every sample() body must explicitly identify the target model/LoRA
+        # adapter — the client does not fall back to any default. Resolve
+        # what name the inference engine knows the policy by from config
+        # (LoRA adapter when LoRA weights are sync'd as an adapter, base
+        # model otherwise).
+        model_name = resolve_policy_model_name(self._cfg)
+
         async def sample_all():
             tasks = []
             for i in range(len(prepared_batch.all_model_inputs)):
@@ -946,6 +944,7 @@ async def sample_all():
 
                 request_payload = {
                     "json": {
+                        "model": model_name,
                         "prompt": model_input.model_dump(),
                         "num_samples": 1,
                         "sampling_params": sampling_params.model_dump(),
@@ -1154,10 +1153,13 @@ def create_ray_wrapped_inference_engines_from_config(
 
     # Conditionally add LoRA parameters if LoRA is enabled
     if cfg.trainer.policy.model.lora.rank > 0 and cfg.trainer.strategy != "megatron":
+        lora_cfg = cfg.trainer.policy.model.lora
         engine_kwargs["enable_lora"] = True
-        engine_kwargs["max_lora_rank"] = cfg.trainer.policy.model.lora.rank
+        engine_kwargs["max_lora_rank"] = lora_cfg.rank
         engine_kwargs["sleep_level"] = 1
-        engine_kwargs["max_loras"] = 1
+        engine_kwargs["max_loras"] = lora_cfg.max_loras
+        if lora_cfg.max_cpu_loras is not None:
+            engine_kwargs["max_cpu_loras"] = lora_cfg.max_cpu_loras
         engine_kwargs["fully_sharded_loras"] = cfg.generator.inference_engine.fully_sharded_loras
 
         if cfg.generator.inference_engine.enforce_eager and cfg.generator.inference_engine.backend == "vllm":
diff --git a/skyrl/benchmarks/load_test_concurrency.py b/skyrl/benchmarks/load_test_concurrency.py
index 036011f89d..88b36fbb8f 100644
--- a/skyrl/benchmarks/load_test_concurrency.py
+++ b/skyrl/benchmarks/load_test_concurrency.py
@@ -186,7 +186,7 @@ async def fire_client_generate(
 
     async def _call(idx: int):
         try:
-            return await client.generate(input_batch)
+            return await client.generate(input_batch, model=client.model_name)
         except Exception as e:
             raise RuntimeError(f"request {idx}: {type(e).__name__}: {e}") from e
 
diff --git a/skyrl/train/config/config.py b/skyrl/train/config/config.py
index aae26ba315..628e393744 100644
--- a/skyrl/train/config/config.py
+++ b/skyrl/train/config/config.py
@@ -61,6 +61,16 @@ class SkyRLLoraConfig(BaseConfig):
     """For FSDP, corresponds to ``init_lora_weights`` in PEFT.
     For Megatron, used for ``lora_A_init_method``; supports "xavier", "normal", "kaiming", "zero"."""
 
+    max_loras: int = 1
+    """Maximum number of LoRA adapters that can be active concurrently in a
+    single GPU batch. Maps to vLLM's ``max_loras``. Increase past 1 to enable
+    multi-tenant LoRA serving via ``RemoteInferenceClient.load_lora_adapter``."""
+
+    max_cpu_loras: Optional[int] = None
+    """Total LoRA adapter capacity in vLLM's CPU LRU cache. Maps to vLLM's
+    ``max_cpu_loras``; when None, vLLM defaults it to ``max_loras``. Must be
+    >= ``max_loras`` if explicitly set."""
+
 
 @dataclass
 class ModelConfig(BaseConfig):
diff --git a/skyrl/train/config/ppo_base_config.yaml b/skyrl/train/config/ppo_base_config.yaml
index 9ac58f2f50..b10695c5b3 100644
--- a/skyrl/train/config/ppo_base_config.yaml
+++ b/skyrl/train/config/ppo_base_config.yaml
@@ -37,6 +37,13 @@ trainer:
         # For FSDP, this corresponds to `init_lora_weights` in PEFT. See: https://huggingface.co/docs/peft/main/en/package_reference/lora#peft.LoraConfig.init_lora_weights
         # For Megatron, this is used for `lora_A_init_method`, and "xavier", "normal", "kaiming", and "zero" are supported.
         init_method: "kaiming"
+        # Maximum number of LoRA adapters that can be active concurrently in a
+        # single GPU batch. Maps to vLLM's `max_loras`. Increase past 1 to
+        # serve multiple LoRA adapters from the same engine.
+        max_loras: 1
+        # Total LoRA adapter capacity in vLLM's CPU LRU cache. Maps to vLLM's
+        # `max_cpu_loras`; when null, vLLM defaults it to `max_loras`.
+        max_cpu_loras: null
     optimizer_config:
       lr: 1.0e-6
       adam_betas: [0.9, 0.999]
diff --git a/skyrl/train/entrypoints/main_base.py b/skyrl/train/entrypoints/main_base.py
index 839fcdca49..ad4206929a 100644
--- a/skyrl/train/entrypoints/main_base.py
+++ b/skyrl/train/entrypoints/main_base.py
@@ -21,7 +21,10 @@
 from skyrl.backends.skyrl_train.inference_engines.remote_inference_engine import (
     create_remote_inference_engines,
 )
-from skyrl.backends.skyrl_train.inference_servers.utils import build_vllm_cli_args
+from skyrl.backends.skyrl_train.inference_servers.utils import (
+    build_vllm_cli_args,
+    resolve_policy_model_name,
+)
 from skyrl.env_vars import _SKYRL_USE_NEW_INFERENCE, SKYRL_RAY_PG_TIMEOUT_IN_S
 from skyrl.train.config import SkyRLTrainConfig, get_config_as_yaml_str
 from skyrl.train.dataset import PromptDataset
@@ -36,9 +39,6 @@
 )
 from skyrl.utils.tok import get_tokenizer
 
-# Fixed LoRA adapter name used for generation requests when LoRA is active.
-_SKYRL_LORA_ADAPTER_NAME = "skyrl-lora"
-
 # NOTE (sumanthrh): We use ray heavily and thus disable `fork` start method.
 # forking within ray leads to undefined behaviour and often causes hard to debug
 # memory leaks.  See: https://docs.ray.io/en/latest/ray-core/patterns/fork-new-processes.html
@@ -88,10 +88,13 @@ def create_ray_wrapped_inference_engines_from_config(
 
     # Conditionally add LoRA parameters if LoRA is enabled
     if cfg.trainer.policy.model.lora.rank > 0 and cfg.trainer.strategy != "megatron":
+        lora_cfg = cfg.trainer.policy.model.lora
         engine_kwargs["enable_lora"] = True
-        engine_kwargs["max_lora_rank"] = cfg.trainer.policy.model.lora.rank
+        engine_kwargs["max_lora_rank"] = lora_cfg.rank
         engine_kwargs["sleep_level"] = 1
-        engine_kwargs["max_loras"] = 1
+        engine_kwargs["max_loras"] = lora_cfg.max_loras
+        if lora_cfg.max_cpu_loras is not None:
+            engine_kwargs["max_cpu_loras"] = lora_cfg.max_cpu_loras
         engine_kwargs["fully_sharded_loras"] = ie_cfg.fully_sharded_loras
 
         # TODO(devpatel): Bandaid solution, replace this once we have a better
@@ -239,6 +242,7 @@ def get_generator(self, cfg, tokenizer, inference_engine_client):
             skyrl_gym_cfg=cfg.environment.skyrl_gym,
             inference_engine_client=inference_engine_client,
             tokenizer=tokenizer,
+            policy_model_name=resolve_policy_model_name(cfg),
         )
 
     def get_trainer(
@@ -390,18 +394,11 @@ def _get_new_inference_client(self):
             proxy_url = setup.proxy_url
             server_urls = setup.server_urls
 
-        lora_cfg = self.cfg.trainer.policy.model.lora
-        active_lora_name = (
-            _SKYRL_LORA_ADAPTER_NAME
-            if lora_cfg and lora_cfg.rank > 0 and self.cfg.trainer.strategy != "megatron"
-            else None
-        )
         client = RemoteInferenceClient(
             proxy_url=proxy_url,
             server_urls=server_urls,
             model_name=self.cfg.trainer.policy.model.path,
             enable_return_routed_experts=ie_cfg.enable_return_routed_experts,
-            active_lora_name=active_lora_name,
             data_parallel_size=ie_cfg.data_parallel_size,
             tokenizer=self.tokenizer,
         )
diff --git a/skyrl/train/generators/skyrl_gym_generator.py b/skyrl/train/generators/skyrl_gym_generator.py
index 6f0d27cacf..44c9afaf62 100644
--- a/skyrl/train/generators/skyrl_gym_generator.py
+++ b/skyrl/train/generators/skyrl_gym_generator.py
@@ -142,17 +142,22 @@ def __init__(
         skyrl_gym_cfg: SkyRLGymConfig,
         inference_engine_client: InferenceEngineClient,
         tokenizer,
+        policy_model_name: str,
     ):
         """
         Args:
             generator_cfg: GeneratorConfig object containing the generator configuration
             inference_engine_client: InferenceEngineClient object for interacting with the inference engines
             tokenizer: tokenizer object for encoding and decoding text
+            policy_model_name: identifier the inference engine knows the policy
+                by (base model path or registered LoRA adapter name). Threaded
+                into every ``client.generate(...)`` call as ``model``.
         """
         self.generator_cfg = generator_cfg
         self.skyrl_gym_cfg = skyrl_gym_cfg
         self.inference_engine_client = inference_engine_client
         self.tokenizer = tokenizer
+        self.policy_model_name = policy_model_name
         self.max_turns = generator_cfg.max_turns
         self.batched = generator_cfg.batched
         self.use_conversation_multi_turn = generator_cfg.use_conversation_multi_turn
@@ -338,7 +343,7 @@ async def agent_loop(
             engine_input = InferenceEngineInput(
                 prompt_token_ids=[agent_loop_state.input_ids], session_ids=[session_id], sampling_params=sampling_params
             )
-            engine_output = await self.inference_engine_client.generate(engine_input)
+            engine_output = await self.inference_engine_client.generate(engine_input, model=self.policy_model_name)
             output = engine_output["responses"][0]
             output_ids = engine_output["response_ids"][0]
             stop_reason = engine_output["stop_reasons"][0]
@@ -671,7 +676,7 @@ async def generate_batched(
             return_dict=False,
         )
         engine_input = InferenceEngineInput(prompt_token_ids=prompt_token_ids, sampling_params=sampling_params)
-        engine_output = await self.inference_engine_client.generate(engine_input)
+        engine_output = await self.inference_engine_client.generate(engine_input, model=self.policy_model_name)
         outputs = engine_output["responses"]
         responses = engine_output["response_ids"]
         stop_reasons = engine_output["stop_reasons"]
diff --git a/skyrl/train/generators/skyrl_vlm_generator.py b/skyrl/train/generators/skyrl_vlm_generator.py
index 961dfd2bb1..6b85c3de5b 100644
--- a/skyrl/train/generators/skyrl_vlm_generator.py
+++ b/skyrl/train/generators/skyrl_vlm_generator.py
@@ -41,8 +41,9 @@ def __init__(
         skyrl_gym_cfg: SkyRLGymConfig,
         inference_engine_client: RemoteInferenceClient,
         tokenizer,
+        policy_model_name: str,
     ):
-        super().__init__(generator_cfg, skyrl_gym_cfg, inference_engine_client, tokenizer)
+        super().__init__(generator_cfg, skyrl_gym_cfg, inference_engine_client, tokenizer, policy_model_name)
         logger.info("Initialized SkyRLVLMGymGenerator (VLM multi-modal generator)")
 
     def _validate_cfg(self, generator_cfg: GeneratorConfig):
@@ -148,7 +149,7 @@ async def agent_loop(
                 sampling_params=current_sampling_params,
                 mm_features=[latest_features] if latest_features is not None else None,
             )
-            engine_output = await self.inference_engine_client.generate(engine_input)
+            engine_output = await self.inference_engine_client.generate(engine_input, model=self.policy_model_name)
 
             gen_text = engine_output["responses"][0]
             gen_ids = engine_output["response_ids"][0]
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_multi_lora_serving.py b/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_multi_lora_serving.py
new file mode 100644
index 0000000000..d54e805dab
--- /dev/null
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_multi_lora_serving.py
@@ -0,0 +1,172 @@
+"""
+Multi-LoRA serving tests for ``RemoteInferenceClient``.
+
+These tests exercise the inference-server-side LoRA control plane:
+``load_lora_adapter`` / ``unload_lora_adapter`` fan-out, per-call ``model=``
+routing across concurrently registered adapters, and the in-place reload
+contract (replacing one adapter without disturbing another).
+
+# Run with:
+uv run --isolated --extra dev --extra fsdp pytest tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_multi_lora_serving.py -v -s
+"""
+
+import pytest
+from huggingface_hub import snapshot_download
+from transformers import AutoTokenizer
+
+from skyrl.backends.skyrl_train.inference_engines.base import InferenceEngineInput
+from skyrl.train.config import SkyRLLoraConfig, SkyRLTrainConfig
+from tests.backends.skyrl_train.gpu.utils import InferenceEngineState
+
+MODEL_QWEN3 = "Qwen/Qwen3-0.6B"
+
+
+@pytest.fixture(scope="session")
+def qwen3_meowing_lora_files():
+    """Download the Qwen3-0.6B Meow LoRA adapter and return its local snapshot path."""
+    return snapshot_download(repo_id="Jackmin108/Qwen3-0.6B-Meow-LoRA")
+
+
+@pytest.fixture(scope="session")
+def qwen3_woofing_lora_files():
+    """Download the Qwen3-0.6B Woof LoRA adapter and return its local snapshot path."""
+    return snapshot_download(repo_id="Jackmin108/Qwen3-0.6B-Woof-LoRA")
+
+
+def _multi_lora_test_config() -> SkyRLTrainConfig:
+    """Build a Qwen3 LoRA config that supports two concurrent adapters on vLLM."""
+    cfg = SkyRLTrainConfig()
+    cfg.trainer.policy.model.path = MODEL_QWEN3
+    cfg.trainer.critic.model.path = ""
+    cfg.trainer.strategy = "fsdp"
+    cfg.trainer.placement.colocate_all = False
+    cfg.trainer.placement.policy_num_gpus_per_node = 1
+    cfg.generator.inference_engine.async_engine = True
+    cfg.generator.inference_engine.num_engines = 1
+    cfg.generator.inference_engine.run_engines_locally = True
+    cfg.generator.inference_engine.tensor_parallel_size = 1
+    # ``rank`` only needs to be > 0 to flip the LoRA path on; the actual ranks
+    # used at serve time come from the adapter snapshots downloaded above.
+    cfg.trainer.policy.model.lora = SkyRLLoraConfig(
+        rank=32,
+        alpha=32,
+        dropout=0.0,
+        target_modules="all-linear",
+        max_loras=2,
+    )
+    return cfg
+
+
+def _build_animal_prompt_token_ids(tokenizer) -> list:
+    """Build prompt_token_ids that ask Qwen3 to make an animal noise.
+
+    Both Meow / Woof LoRAs are tuned to override the assistant reply with their
+    respective sound, so a neutral prompt is enough to exercise routing.
+    """
+    messages = [
+        {"role": "user", "content": "Make a single short animal noise."},
+    ]
+    return tokenizer.apply_chat_template(
+        messages,
+        add_generation_prompt=True,
+        tokenize=True,
+        return_dict=False,
+        enable_thinking=False,
+    )
+
+
+async def _generate_with_lora(client, prompt_token_ids, lora_name: str) -> str:
+    """Run a single greedy generation against ``lora_name`` and return the text."""
+    sampling_params = {"temperature": 0.0, "max_tokens": 10}
+    out = await client.generate(
+        InferenceEngineInput(
+            prompt_token_ids=[prompt_token_ids],
+            sampling_params=sampling_params,
+        ),
+        model=lora_name,
+    )
+    return out["responses"][0]
+
+
+@pytest.mark.asyncio
+async def test_multi_lora_interleaved_generation(ray_init_fixture, qwen3_meowing_lora_files, qwen3_woofing_lora_files):
+    """Two adapters served concurrently route per-call via the ``model=`` kwarg."""
+    cfg = _multi_lora_test_config()
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_QWEN3, trust_remote_code=True)
+    prompt_token_ids = _build_animal_prompt_token_ids(tokenizer)
+
+    async with InferenceEngineState.create(
+        cfg=cfg,
+        model=MODEL_QWEN3,
+        use_local=True,
+        async_engine=True,
+        tp_size=1,
+        colocate_all=False,
+        sleep_level=1,
+        enable_lora=True,
+        lora_max_loras=2,
+    ) as engines:
+        client = engines.client
+
+        await client.load_lora_adapter("lora-meow", qwen3_meowing_lora_files)
+        await client.load_lora_adapter("lora-woof", qwen3_woofing_lora_files)
+        try:
+            outputs = []
+            for adapter in ["lora-meow", "lora-woof", "lora-meow", "lora-woof"]:
+                outputs.append(await _generate_with_lora(client, prompt_token_ids, adapter))
+
+            print(f"Multi-LoRA outputs: {outputs}")
+            assert "Meow" in outputs[0] or "meow" in outputs[0]
+            assert "Woof" in outputs[1] or "woof" in outputs[1]
+            assert "Meow" in outputs[2] or "meow" in outputs[2]
+            assert "Woof" in outputs[3] or "woof" in outputs[3]
+        finally:
+            await client.unload_lora_adapter("lora-meow")
+            await client.unload_lora_adapter("lora-woof")
+
+
+@pytest.mark.asyncio
+async def test_lora_inplace_reload_isolated(ray_init_fixture, qwen3_meowing_lora_files, qwen3_woofing_lora_files):
+    """Reloading adapter ``lora-A`` from a different path leaves ``lora-B`` unchanged."""
+    cfg = _multi_lora_test_config()
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_QWEN3, trust_remote_code=True)
+    prompt_token_ids = _build_animal_prompt_token_ids(tokenizer)
+
+    async with InferenceEngineState.create(
+        cfg=cfg,
+        model=MODEL_QWEN3,
+        use_local=True,
+        async_engine=True,
+        tp_size=1,
+        colocate_all=False,
+        sleep_level=1,
+        enable_lora=True,
+        lora_max_loras=2,
+    ) as engines:
+        client = engines.client
+
+        await client.load_lora_adapter("lora-A", qwen3_meowing_lora_files)
+        await client.load_lora_adapter("lora-B", qwen3_woofing_lora_files)
+        try:
+            out_A_before = await _generate_with_lora(client, prompt_token_ids, "lora-A")
+            out_B_before = await _generate_with_lora(client, prompt_token_ids, "lora-B")
+            assert "Meow" in out_A_before or "meow" in out_A_before
+            assert "Woof" in out_B_before or "woof" in out_B_before
+
+            # Inplace reload A from B's adapter path. vLLM keeps the same
+            # int_id but should swap the underlying weights; B must be entirely
+            # unaffected.
+            await client.load_lora_adapter("lora-A", qwen3_woofing_lora_files, load_inplace=True)
+
+            out_A_after = await _generate_with_lora(client, prompt_token_ids, "lora-A")
+            out_B_after = await _generate_with_lora(client, prompt_token_ids, "lora-B")
+
+            assert (
+                "Woof" in out_A_after or "woof" in out_A_after
+            ), f"A should now be woofing-style after inplace reload, got: {out_A_after}"
+            assert (
+                out_B_after == out_B_before
+            ), f"B's output should be unchanged byte-for-byte; before={out_B_before!r}, after={out_B_after!r}"
+        finally:
+            await client.unload_lora_adapter("lora-A")
+            await client.unload_lora_adapter("lora-B")
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_new_inference_generation.py b/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_new_inference_generation.py
index 225d65e3f8..8f17f815c7 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_new_inference_generation.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_new_inference_generation.py
@@ -563,7 +563,7 @@ async def test_client_generate(vllm_server: InferenceEngineState):
         sampling_params=sampling_params,
     )
 
-    output = await client.generate(engine_input)
+    output = await client.generate(engine_input, model=client.model_name)
 
     assert len(output["responses"]) == 1
     assert len(output["response_ids"]) == 1
@@ -595,6 +595,7 @@ async def test_client_tokenize_detokenize_roundtrip(vllm_server: InferenceEngine
 
 def _build_sample_payload(
     token_ids: List[int],
+    model: str = MODEL_QWEN2_5,
     num_samples: int = 1,
     sampling_params: Dict[str, Any] | None = None,
     session_id: str | None = None,
@@ -603,6 +604,7 @@ def _build_sample_payload(
 ) -> Dict[str, Any]:
     """Build a Tinker-format sample request payload."""
     body: Dict[str, Any] = {
+        "model": model,
         "prompt": {"chunks": [{"tokens": token_ids}]},
         "num_samples": num_samples,
         "sampling_params": sampling_params or {"temperature": 0.7, "max_tokens": 64},
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_remote_inference_client_chat_template.py b/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_remote_inference_client_chat_template.py
index 300fac372a..1ef6e10bc9 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_remote_inference_client_chat_template.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_remote_inference_client_chat_template.py
@@ -90,7 +90,7 @@ async def test_custom_chat_template(ray_init_fixture, use_custom_template: bool)
             prompt_token_ids=[prompt_token_ids],
             sampling_params={"max_tokens": 10},
         )
-        output = await client.generate(engine_input)
+        output = await client.generate(engine_input, model=client.model_name)
         assert len(output["responses"]) == 1
         assert isinstance(output["responses"][0], str)
 
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_vlm_inference_generation.py b/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_vlm_inference_generation.py
index c432fa6dcd..04f2448c24 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_vlm_inference_generation.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_vlm_inference_generation.py
@@ -203,6 +203,7 @@ async def test_sample_with_multimodal_image(module_scoped_ray_init_fixture):
 
         request_payload = {
             "json": {
+                "model": MODEL_QWEN3_VL,
                 "prompt": prompt,
                 "num_samples": 1,
                 "sampling_params": {"temperature": 0.0, "max_tokens": 20},
@@ -409,7 +410,7 @@ async def test_generate_with_multimodal_features_red_square(module_scoped_ray_in
             "sampling_params": {"max_tokens": 64, "temperature": 0.0},
             "mm_features": [features],
         }
-        gen_result = await client.generate(input_batch)
+        gen_result = await client.generate(input_batch, model=client.model_name)
 
         # Structural assertions
         assert len(gen_result["responses"]) == 1
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/test_engine_generation.py b/tests/backends/skyrl_train/gpu/gpu_ci/test_engine_generation.py
index 82a65a198c..fba1dd6dad 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/test_engine_generation.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/test_engine_generation.py
@@ -41,7 +41,7 @@ def get_test_actor_config(model: str = MODEL) -> SkyRLTrainConfig:
 
 async def run_batch_generation(client, prompts, sampling_params):
     engine_input = InferenceEngineInput(prompts=prompts, sampling_params=sampling_params)
-    engine_output = await client.generate(engine_input)
+    engine_output = await client.generate(engine_input, model=client.model_name)
     return engine_output["responses"], engine_output["stop_reasons"]
 
 
@@ -49,7 +49,7 @@ async def run_single_generation(client, prompts, sampling_params):
     tasks = []
     for prompt in prompts:
         engine_input = InferenceEngineInput(prompts=[prompt], sampling_params=sampling_params)
-        task = client.generate(engine_input)
+        task = client.generate(engine_input, model=client.model_name)
         tasks.append(task)
 
     results = await asyncio.gather(*tasks)
@@ -65,7 +65,7 @@ async def run_single_generation(client, prompts, sampling_params):
 
 async def run_batch_generation_with_tokens(client, prompt_token_ids, sampling_params):
     engine_input = InferenceEngineInput(prompt_token_ids=prompt_token_ids, sampling_params=sampling_params)
-    engine_output = await client.generate(engine_input)
+    engine_output = await client.generate(engine_input, model=client.model_name)
     return engine_output["responses"], engine_output["stop_reasons"]
 
 
@@ -73,7 +73,7 @@ async def run_single_generation_with_tokens(client, prompt_token_ids, sampling_p
     tasks = []
     for tokens in prompt_token_ids:
         engine_input = InferenceEngineInput(prompt_token_ids=[tokens], sampling_params=sampling_params)
-        task = client.generate(engine_input)
+        task = client.generate(engine_input, model=client.model_name)
         tasks.append(task)
 
     results = await asyncio.gather(*tasks)
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/test_lora.py b/tests/backends/skyrl_train/gpu/gpu_ci/test_lora.py
index 13741c89c3..ea02c5b64d 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/test_lora.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/test_lora.py
@@ -4,6 +4,11 @@
 
 # Run Megatron tests:
 uv run --isolated --extra dev --extra megatron pytest tests/backends/skyrl_train/gpu/gpu_ci/test_lora.py -k "megatron"
+
+Multi-LoRA serving tests live separately in
+``tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_multi_lora_serving.py``
+since they exercise the inference-server LoRA control plane, not the
+trainer + weight-sync path covered here.
 """
 
 import pytest
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/test_pause_and_continue_generation.py b/tests/backends/skyrl_train/gpu/gpu_ci/test_pause_and_continue_generation.py
index d3084482ba..55ec3f77d7 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/test_pause_and_continue_generation.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/test_pause_and_continue_generation.py
@@ -234,7 +234,7 @@ async def one_req(i: int):
                     "sampling_params": dict(sampling_params),
                     "session_ids": [i],
                 }
-                return await client.generate(engine_input)
+                return await client.generate(engine_input, model=client.model_name)
 
             tasks = [asyncio.create_task(one_req(i)) for i in range(num_requests)]
             # Let requests start and enqueue; with max_num_seqs=2, 2 run and 1 wait per engine
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/test_router_replay.py b/tests/backends/skyrl_train/gpu/gpu_ci/test_router_replay.py
index ed4fca055d..32c61641fa 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/test_router_replay.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/test_router_replay.py
@@ -150,6 +150,7 @@ async def test_logprobs(ray_init_fixture, tp, pp, cp, ep, etp, extra_tf_kwargs):
                 skyrl_gym_cfg=cfg.environment.skyrl_gym,
                 inference_engine_client=client,
                 tokenizer=tokenizer,
+                policy_model_name=client.model_name,
             )
 
             input_batch: GeneratorInput = get_test_generator_input(
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_gym_generator.py b/tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_gym_generator.py
index 8a3a258a5e..c66461b1b8 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_gym_generator.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_gym_generator.py
@@ -171,6 +171,7 @@ async def run_generator_end_to_end(
             skyrl_gym_cfg=env_cfg,
             inference_engine_client=inference_engine_client,
             tokenizer=tokenizer,
+            policy_model_name=inference_engine_client.model_name,
         )
 
         input_batch: GeneratorInput = get_test_generator_input(
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_vlm_gym_generator.py b/tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_vlm_gym_generator.py
index a671e8d0f8..03bfe9a3fe 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_vlm_gym_generator.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_vlm_gym_generator.py
@@ -191,6 +191,7 @@ async def test_vlm_generator_color_classification(ray_init_fixture):
             skyrl_gym_cfg=env_cfg,
             inference_engine_client=inference_client,
             tokenizer=tokenizer,
+            policy_model_name=inference_client.model_name,
         )
 
         num_prompts = 2
diff --git a/tests/backends/skyrl_train/gpu/utils.py b/tests/backends/skyrl_train/gpu/utils.py
index 82ab42ada2..6cd764b765 100644
--- a/tests/backends/skyrl_train/gpu/utils.py
+++ b/tests/backends/skyrl_train/gpu/utils.py
@@ -509,6 +509,8 @@ def create(
         sleep_level: int = 2,  # use level 1 in unit tests that do not explicitly sync weights or for LoRA
         enable_lora: bool = False,
         active_lora_name: Optional[str] = None,
+        lora_max_loras: Optional[int] = None,
+        lora_max_cpu_loras: Optional[int] = None,
         max_num_seqs: Optional[int] = None,
         engine_init_kwargs: Optional[Dict[str, Any]] = None,
         use_new_inference_servers: Optional[bool] = None,
@@ -552,6 +554,13 @@ def create(
         if enable_pd:
             ie_cfg.enable_pd = True
             ie_cfg.num_prefill = num_prefill
+        # Propagate the LoRA limits onto the trainer config so build_vllm_cli_args
+        # (which reads from cfg.trainer.policy.model.lora) and any downstream
+        # path picks them up before vLLM is started.
+        if lora_max_loras is not None:
+            cfg.trainer.policy.model.lora.max_loras = lora_max_loras
+        if lora_max_cpu_loras is not None:
+            cfg.trainer.policy.model.lora.max_cpu_loras = lora_max_cpu_loras
 
         assert ie_cfg.run_engines_locally, "This test does not yet support remote engines."
 
@@ -605,12 +614,21 @@ def create(
             proxy_url = setup.proxy_url
             server_urls = setup.server_urls
 
+            # When LoRA is enabled, point the client's default ``model_name`` at
+            # the active LoRA adapter so existing tests that don't pass an
+            # explicit ``model=`` keep routing through the adapter. Tests that
+            # need multi-LoRA can pass ``model=`` per call and rely on the
+            # client's underlying base model name only when needed.
+            # ``model_name`` is the base model the server was started with;
+            # LoRA-aware test cases are expected to pass adapter names
+            # explicitly per call (e.g. ``client.generate(..., model="lora-X")``).
+            base_model_name = served_model_name if served_model_name else cfg.trainer.policy.model.path
+
             client = RemoteInferenceClient(
                 proxy_url=proxy_url,
                 server_urls=server_urls,
-                model_name=served_model_name if served_model_name else cfg.trainer.policy.model.path,
+                model_name=base_model_name,
                 enable_return_routed_experts=ie_cfg.enable_return_routed_experts,
-                active_lora_name=active_lora_name,
                 data_parallel_size=ie_cfg.data_parallel_size,
                 tokenizer=get_tokenizer(cfg.trainer.policy.model.path),
             )
diff --git a/tests/backends/skyrl_train/inference_servers/test_remote_inference_client.py b/tests/backends/skyrl_train/inference_servers/test_remote_inference_client.py
index 5e081a4715..b9a2c0f458 100644
--- a/tests/backends/skyrl_train/inference_servers/test_remote_inference_client.py
+++ b/tests/backends/skyrl_train/inference_servers/test_remote_inference_client.py
@@ -4,16 +4,19 @@
 import pickle
 import threading
 import time
-from typing import List, Optional
+from typing import Dict, List, Optional
 
+import aiohttp
 import httpx
 import pytest
 import pytest_asyncio
 import uvicorn
 from fastapi import FastAPI, Query, Request
+from fastapi.responses import JSONResponse, PlainTextResponse
 
 from skyrl.backends.skyrl_train.inference_servers.common import get_open_port
 from skyrl.backends.skyrl_train.inference_servers.remote_inference_client import (
+    SKYRL_LORA_ADAPTER_NAME,
     PauseMode,
     RemoteInferenceClient,
 )
@@ -23,6 +26,12 @@ def create_mock_vllm_server(server_id: int) -> FastAPI:
     """Create a mock vLLM server with standard endpoints."""
     app = FastAPI()
     app.state.last_generate_features = None
+    app.state.last_generate_model = None
+    app.state.last_chat_model = None
+    app.state.last_completion_model = None
+    app.state.last_render_model = None
+    # Per-server LoRA registry: lora_name -> lora_path
+    app.state.lora_registry = {}
 
     @app.get("/health")
     async def health():
@@ -32,6 +41,19 @@ async def health():
     async def get_last_generate_features():
         return {"features": app.state.last_generate_features}
 
+    @app.get("/test/last_models")
+    async def get_last_models():
+        return {
+            "generate": app.state.last_generate_model,
+            "chat": app.state.last_chat_model,
+            "completion": app.state.last_completion_model,
+            "render": app.state.last_render_model,
+        }
+
+    @app.get("/test/lora_registry")
+    async def get_lora_registry():
+        return {"registry": dict(app.state.lora_registry)}
+
     @app.get("/get_world_size")
     async def get_world_size():
         return {"world_size": 2}  # Simulate TP=2
@@ -39,13 +61,15 @@ async def get_world_size():
     @app.post("/v1/completions")
     async def completions(request: Request):
         body = await request.json()
+        app.state.last_completion_model = body.get("model")
         prompts = body.get("prompt", [])
         n_prompts = len(prompts) if isinstance(prompts, list) else 1
         return {
             "choices": [
                 {"index": i, "text": f"Response {i} from server {server_id}", "finish_reason": "stop"}
                 for i in range(n_prompts)
-            ]
+            ],
+            "model": body.get("model"),
         }
 
     @app.post("/skyrl/v1/generate")
@@ -54,6 +78,7 @@ async def generate(request: Request):
         body = await request.json()  # Consume body
         sp = body.get("sampling_params", {})
         input_token_ids = body.get("token_ids", [])
+        app.state.last_generate_model = body.get("model")
         n = sp.get("n", 1)
         # If logprobs is explicitly set (sample path), use n for num_choices.
         # Otherwise (generate path), use len(token_ids) for per-prompt responses.
@@ -109,11 +134,17 @@ async def generate(request: Request):
 
     @app.post("/v1/chat/completions")
     async def chat_completions(request: Request):
-        return {"choices": [{"message": {"content": f"Chat from server {server_id}"}}]}
+        body = await request.json()
+        app.state.last_chat_model = body.get("model")
+        return {
+            "choices": [{"message": {"content": f"Chat from server {server_id}"}}],
+            "model": body.get("model"),
+        }
 
     @app.post("/v1/chat/completions/render")
     async def render_chat_completion(request: Request):
         body = await request.json()
+        app.state.last_render_model = body.get("model")
         messages = body.get("messages", [])
 
         # Count image_url parts across all messages.
@@ -205,6 +236,45 @@ async def init_weight_transfer_engine(request: Request):
     async def update_weights(request: Request):
         return {"status": "ok", "server_id": server_id}
 
+    @app.post("/v1/load_lora_adapter")
+    async def load_lora_adapter(request: Request):
+        body = await request.json()
+        lora_name = body.get("lora_name")
+        lora_path = body.get("lora_path")
+        load_inplace = body.get("load_inplace", True)
+        if lora_name is None or lora_path is None:
+            return JSONResponse(
+                status_code=400,
+                content={"object": "error", "message": "missing lora_name/lora_path", "type": "BadRequest"},
+            )
+        if lora_name in app.state.lora_registry and not load_inplace:
+            return JSONResponse(
+                status_code=400,
+                content={
+                    "object": "error",
+                    "message": f"adapter '{lora_name}' already loaded",
+                    "type": "BadRequest",
+                },
+            )
+        app.state.lora_registry[lora_name] = lora_path
+        return PlainTextResponse(f"Success: LoRA adapter '{lora_name}' added successfully on server {server_id}.")
+
+    @app.post("/v1/unload_lora_adapter")
+    async def unload_lora_adapter(request: Request):
+        body = await request.json()
+        lora_name = body.get("lora_name")
+        if lora_name is None or lora_name not in app.state.lora_registry:
+            return JSONResponse(
+                status_code=404,
+                content={
+                    "object": "error",
+                    "message": f"adapter '{lora_name}' not found",
+                    "type": "NotFoundError",
+                },
+            )
+        del app.state.lora_registry[lora_name]
+        return PlainTextResponse(f"Success: LoRA adapter '{lora_name}' removed successfully on server {server_id}.")
+
     return app
 
 
@@ -300,7 +370,7 @@ async def test_generate(self, client):
             "prompt_token_ids": [[1, 2, 3], [4, 5, 6]],
             "sampling_params": {"max_tokens": 100},
         }
-        result = await client.generate(input_batch)
+        result = await client.generate(input_batch, model=client.model_name)
 
         assert "responses" in result
         assert "stop_reasons" in result
@@ -316,7 +386,7 @@ async def test_generate_with_session_id(self, client):
             "prompt_token_ids": [[1, 2, 3]],
             "session_ids": ["test-session"],
         }
-        result = await client.generate(input_batch)
+        result = await client.generate(input_batch, model=client.model_name)
         assert len(result["responses"]) == 1
 
     @pytest.mark.asyncio
@@ -508,6 +578,7 @@ async def test_sample(self, client):
         """Test sample with n=1 returns correct structure and prompt_logprobs."""
         request_payload = {
             "json": {
+                "model": client.model_name,
                 "prompt": {"chunks": [{"tokens": [10, 20, 30]}]},
                 "num_samples": 1,
                 "sampling_params": {"temperature": 0.7, "max_tokens": 64},
@@ -539,6 +610,7 @@ async def test_sample_n2(self, client):
         """Test sample with n=2 returns two sequences and prompt_logprobs."""
         request_payload = {
             "json": {
+                "model": client.model_name,
                 "prompt": {"chunks": [{"tokens": [1, 2]}, {"tokens": [3]}]},
                 "num_samples": 2,
                 "sampling_params": {"temperature": 1.0, "max_tokens": 32},
@@ -564,6 +636,7 @@ async def test_sample_topk_prompt_logprobs(self, client):
         """Test topk_prompt_logprobs returns both prompt_logprobs and topk tuples."""
         request_payload = {
             "json": {
+                "model": client.model_name,
                 "prompt": {"chunks": [{"tokens": [10, 20, 30]}]},
                 "num_samples": 1,
                 "sampling_params": {"temperature": 0.7, "max_tokens": 64},
@@ -597,6 +670,7 @@ async def test_sample_topk_without_include_returns_none(self, client):
         """topk_prompt_logprobs alone does not return prompt logprobs when include_prompt_logprobs is False."""
         request_payload = {
             "json": {
+                "model": client.model_name,
                 "prompt": {"chunks": [{"tokens": [10, 20, 30]}]},
                 "num_samples": 1,
                 "sampling_params": {"temperature": 0.7, "max_tokens": 64},
@@ -616,6 +690,7 @@ async def test_sample_with_image(self, client):
         image_bytes = base64.b64encode(b"fake-jpeg-data").decode("ascii")
         request_payload = {
             "json": {
+                "model": client.model_name,
                 "prompt": {
                     "chunks": [
                         {"type": "encoded_text", "tokens": [100, 101, 102]},
@@ -646,6 +721,7 @@ async def test_sample_with_image_asset_pointer(self, client):
         """Sample with image_asset_pointer sends location URL to render."""
         request_payload = {
             "json": {
+                "model": client.model_name,
                 "prompt": {
                     "chunks": [
                         {"type": "encoded_text", "tokens": [10, 11]},
@@ -671,6 +747,7 @@ async def test_sample_text_only_no_features(self, client):
         """Text-only sample does not include features in the generate payload."""
         request_payload = {
             "json": {
+                "model": client.model_name,
                 "prompt": {"chunks": [{"type": "encoded_text", "tokens": [1, 2, 3]}]},
                 "num_samples": 1,
                 "sampling_params": {"temperature": 0.7, "max_tokens": 64},
@@ -761,7 +838,7 @@ async def test_generate_with_mm_features(self, client, mock_servers):
             "sampling_params": {"max_tokens": 50},
             "mm_features": [mm_features],
         }
-        result = await client.generate(input_batch)
+        result = await client.generate(input_batch, model=client.model_name)
 
         assert len(result["responses"]) == 1
         assert len(result["response_ids"]) == 1
@@ -792,3 +869,296 @@ async def test_async_context_manager(self, mock_servers):
 
         # Session should be closed after exiting context
         assert client._session is None or client._session.closed
+
+
+async def _get_lora_registries(server_urls: List[str]) -> List[Dict[str, str]]:
+    """Helper: read the per-server LoRA registries from each mock server."""
+    registries: List[Dict[str, str]] = []
+    async with httpx.AsyncClient() as http:
+        for url in server_urls:
+            resp = await http.get(f"{url}/test/lora_registry")
+            registries.append(resp.json()["registry"])
+    return registries
+
+
+async def _get_last_models(server_urls: List[str]) -> List[Dict[str, Optional[str]]]:
+    """Helper: read the last per-method ``model`` field captured by each mock."""
+    last: List[Dict[str, Optional[str]]] = []
+    async with httpx.AsyncClient() as http:
+        for url in server_urls:
+            resp = await http.get(f"{url}/test/last_models")
+            last.append(resp.json())
+    return last
+
+
+class TestLoRAControlPlane:
+    """Test load_lora_adapter / unload_lora_adapter fan-out and bookkeeping."""
+
+    @pytest.mark.asyncio
+    async def test_load_lora_adapter_fans_out(self, client, mock_servers):
+        result = await client.load_lora_adapter("lora-A", "/tmp/path/lora-A")
+        assert len(result) == 2
+        for url, response in result.items():
+            assert response["status"] == 200
+            assert "Success" in response["body"]
+            assert "lora-A" in response["body"]
+
+        registries = await _get_lora_registries(mock_servers["server_urls"])
+        for reg in registries:
+            assert reg.get("lora-A") == "/tmp/path/lora-A"
+
+        await client.unload_lora_adapter("lora-A")
+
+    @pytest.mark.asyncio
+    async def test_load_lora_adapter_inplace_reload(self, client, mock_servers):
+        await client.load_lora_adapter("lora-X", "/tmp/path/v1")
+        await client.load_lora_adapter("lora-X", "/tmp/path/v2", load_inplace=True)
+
+        registries = await _get_lora_registries(mock_servers["server_urls"])
+        for reg in registries:
+            assert reg.get("lora-X") == "/tmp/path/v2"
+
+        await client.unload_lora_adapter("lora-X")
+
+    @pytest.mark.asyncio
+    async def test_load_lora_adapter_inplace_false_raises_on_conflict(self, client):
+        await client.load_lora_adapter("lora-conflict", "/tmp/path/orig")
+        try:
+            with pytest.raises(aiohttp.ClientResponseError):
+                await client.load_lora_adapter("lora-conflict", "/tmp/path/other", load_inplace=False)
+        finally:
+            await client.unload_lora_adapter("lora-conflict")
+
+    @pytest.mark.asyncio
+    async def test_unload_lora_adapter_fans_out(self, client, mock_servers):
+        await client.load_lora_adapter("lora-B", "/tmp/path/lora-B")
+
+        result = await client.unload_lora_adapter("lora-B")
+        assert len(result) == 2
+        for url, response in result.items():
+            assert response["status"] == 200
+            assert "Success" in response["body"]
+
+        registries = await _get_lora_registries(mock_servers["server_urls"])
+        for reg in registries:
+            assert "lora-B" not in reg
+
+    @pytest.mark.asyncio
+    async def test_unload_unknown_lora_raises(self, client, mock_servers):
+        # Server returns 404, surfaced as ClientResponseError via raise_for_status.
+        with pytest.raises(aiohttp.ClientResponseError):
+            await client.unload_lora_adapter("nonexistent-lora")
+        registries = await _get_lora_registries(mock_servers["server_urls"])
+        for reg in registries:
+            assert "nonexistent-lora" not in reg
+
+    @pytest.mark.asyncio
+    async def test_default_lora_adapter_constant(self):
+        # Sanity check that the public constant has the documented value used
+        # across the SkyRL training paths.
+        assert SKYRL_LORA_ADAPTER_NAME == "skyrl-lora"
+
+
+class TestExplicitModelRequired:
+    """Every data-plane call must explicitly identify the target model.
+
+    ``generate`` takes ``model`` as a required keyword argument; the body-style
+    methods (``sample``, ``chat_completion``, ``completion``,
+    ``render_chat_completion``) require it inside the request body. There is
+    no fallback to ``client.model_name`` on the data plane — that field is
+    only used internally for ``tokenize``/``detokenize``.
+    """
+
+    @pytest.mark.asyncio
+    async def test_generate_threads_model_into_payload(self, mock_servers):
+        client = RemoteInferenceClient(
+            proxy_url=mock_servers["proxy_url"],
+            server_urls=mock_servers["server_urls"],
+            model_name="base-model",
+            data_parallel_size=1,
+        )
+        try:
+            input_batch = {
+                "prompt_token_ids": [[1, 2, 3]],
+                "sampling_params": {"max_tokens": 50},
+            }
+            await client.generate(input_batch, model="lora-explicit")
+            captured = await _get_last_models(mock_servers["server_urls"])
+            # generate routes through proxy_url == first server.
+            assert captured[0]["generate"] == "lora-explicit"
+        finally:
+            await client.teardown()
+
+    @pytest.mark.asyncio
+    async def test_generate_requires_model_kwarg(self, mock_servers):
+        client = RemoteInferenceClient(
+            proxy_url=mock_servers["proxy_url"],
+            server_urls=mock_servers["server_urls"],
+            model_name="base-model",
+            data_parallel_size=1,
+        )
+        try:
+            input_batch = {
+                "prompt_token_ids": [[1, 2, 3]],
+                "sampling_params": {"max_tokens": 50},
+            }
+            with pytest.raises(TypeError):
+                # ``model`` is required (positional/keyword-only) on the new client.
+                await client.generate(input_batch)  # type: ignore[call-arg]
+        finally:
+            await client.teardown()
+
+    @pytest.mark.asyncio
+    async def test_chat_completion_uses_body_model(self, mock_servers):
+        client = RemoteInferenceClient(
+            proxy_url=mock_servers["proxy_url"],
+            server_urls=mock_servers["server_urls"],
+            model_name="base-model",
+            data_parallel_size=1,
+        )
+        try:
+            request_payload = {
+                "json": {
+                    "model": "lora-chat",
+                    "messages": [{"role": "user", "content": "hi"}],
+                },
+                "headers": {},
+            }
+            await client.chat_completion(request_payload)
+            captured = await _get_last_models(mock_servers["server_urls"])
+            assert captured[0]["chat"] == "lora-chat"
+        finally:
+            await client.teardown()
+
+    @pytest.mark.asyncio
+    async def test_chat_completion_missing_model_raises(self, mock_servers):
+        client = RemoteInferenceClient(
+            proxy_url=mock_servers["proxy_url"],
+            server_urls=mock_servers["server_urls"],
+            model_name="base-model",
+            data_parallel_size=1,
+        )
+        try:
+            request_payload = {
+                "json": {"messages": [{"role": "user", "content": "hi"}]},
+                "headers": {},
+            }
+            with pytest.raises(ValueError, match="must include a non-empty 'model' field"):
+                await client.chat_completion(request_payload)
+        finally:
+            await client.teardown()
+
+    @pytest.mark.asyncio
+    async def test_completion_uses_body_model(self, mock_servers):
+        client = RemoteInferenceClient(
+            proxy_url=mock_servers["proxy_url"],
+            server_urls=mock_servers["server_urls"],
+            model_name="base-model",
+            data_parallel_size=1,
+        )
+        try:
+            request_payload = {
+                "json": {"model": "lora-completion", "prompt": "hello"},
+                "headers": {},
+            }
+            await client.completion(request_payload)
+            captured = await _get_last_models(mock_servers["server_urls"])
+            assert captured[0]["completion"] == "lora-completion"
+        finally:
+            await client.teardown()
+
+    @pytest.mark.asyncio
+    async def test_completion_missing_model_raises(self, mock_servers):
+        client = RemoteInferenceClient(
+            proxy_url=mock_servers["proxy_url"],
+            server_urls=mock_servers["server_urls"],
+            model_name="base-model",
+            data_parallel_size=1,
+        )
+        try:
+            request_payload = {"json": {"prompt": "hello"}, "headers": {}}
+            with pytest.raises(ValueError, match="must include a non-empty 'model' field"):
+                await client.completion(request_payload)
+        finally:
+            await client.teardown()
+
+    @pytest.mark.asyncio
+    async def test_render_chat_completion_uses_body_model(self, mock_servers):
+        client = RemoteInferenceClient(
+            proxy_url=mock_servers["proxy_url"],
+            server_urls=mock_servers["server_urls"],
+            model_name="base-model",
+            data_parallel_size=1,
+        )
+        try:
+            request_payload = {
+                "json": {
+                    "model": "lora-render",
+                    "messages": [{"role": "user", "content": "hi"}],
+                },
+            }
+            result = await client.render_chat_completion(request_payload)
+            assert result["model"] == "lora-render"
+            captured = await _get_last_models(mock_servers["server_urls"])
+            assert captured[0]["render"] == "lora-render"
+        finally:
+            await client.teardown()
+
+    @pytest.mark.asyncio
+    async def test_render_chat_completion_missing_model_raises(self, mock_servers):
+        client = RemoteInferenceClient(
+            proxy_url=mock_servers["proxy_url"],
+            server_urls=mock_servers["server_urls"],
+            model_name="base-model",
+            data_parallel_size=1,
+        )
+        try:
+            request_payload = {"json": {"messages": [{"role": "user", "content": "hi"}]}}
+            with pytest.raises(ValueError, match="must include a non-empty 'model' field"):
+                await client.render_chat_completion(request_payload)
+        finally:
+            await client.teardown()
+
+    @pytest.mark.asyncio
+    async def test_sample_uses_body_model(self, mock_servers):
+        client = RemoteInferenceClient(
+            proxy_url=mock_servers["proxy_url"],
+            server_urls=mock_servers["server_urls"],
+            model_name="base-model",
+            data_parallel_size=1,
+        )
+        try:
+            request_payload = {
+                "json": {
+                    "model": "lora-sample",
+                    "prompt": {"chunks": [{"tokens": [1, 2, 3]}]},
+                    "num_samples": 1,
+                    "sampling_params": {"temperature": 0.7, "max_tokens": 16},
+                }
+            }
+            await client.sample(request_payload)
+            captured = await _get_last_models(mock_servers["server_urls"])
+            assert captured[0]["generate"] == "lora-sample"
+        finally:
+            await client.teardown()
+
+    @pytest.mark.asyncio
+    async def test_sample_missing_model_raises(self, mock_servers):
+        client = RemoteInferenceClient(
+            proxy_url=mock_servers["proxy_url"],
+            server_urls=mock_servers["server_urls"],
+            model_name="base-model",
+            data_parallel_size=1,
+        )
+        try:
+            request_payload = {
+                "json": {
+                    "prompt": {"chunks": [{"tokens": [1, 2, 3]}]},
+                    "num_samples": 1,
+                    "sampling_params": {"temperature": 0.7, "max_tokens": 16},
+                }
+            }
+            with pytest.raises(ValueError, match="must include a non-empty 'model' field"):
+                await client.sample(request_payload)
+        finally:
+            await client.teardown()
diff --git a/tests/train/generators/test_skyrl_gym_generator.py b/tests/train/generators/test_skyrl_gym_generator.py
index cf973457ae..322938a38c 100644
--- a/tests/train/generators/test_skyrl_gym_generator.py
+++ b/tests/train/generators/test_skyrl_gym_generator.py
@@ -74,7 +74,7 @@ def mock_llm():
     mock = MagicMock()
 
     # Mock the new generate method
-    def mock_generate(input_batch):
+    def mock_generate(input_batch, model=None):
         num_prompts = len(input_batch["prompts"]) if "prompts" in input_batch else len(input_batch["prompt_token_ids"])
         return {
             "responses": ["mocked output"] * num_prompts,
@@ -283,6 +283,7 @@ def mock_generate(_):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
+        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []  # to make sure observation_ids are encoded correctly
 
@@ -334,6 +335,7 @@ async def test_generate_batched(mock_make, mock_tokenizer, mock_llm, mock_env, g
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
+        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []  # to make sure observation_ids are encoded correctly
 
@@ -376,6 +378,7 @@ async def test_generate_interface_compliance(
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
+        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []  # to make sure observation_ids are encoded correctly
 
@@ -498,6 +501,7 @@ def mock_generate(input_batch):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
+        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []  # to make sure observation_ids are encoded correctly
 
@@ -587,6 +591,7 @@ def mock_encode(text, **kwargs):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
+        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []  # to make sure observation_ids are encoded correctly
 
@@ -668,6 +673,7 @@ def mock_encode(text, **kwargs):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
+        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []  # to make sure observation_ids are encoded correctly
 
@@ -731,6 +737,7 @@ def mock_apply_chat_template(messages, **kwargs):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
+        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []  # to make sure observation_ids are encoded correctly
 
@@ -852,6 +859,7 @@ def mock_encode_or_tokenize(text):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
+        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []  # to make sure observation_ids are encoded correctly
 
@@ -951,6 +959,7 @@ def step(self, action):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
+        policy_model_name="mock-model",
     )
 
     # Run agent loop
@@ -1041,6 +1050,7 @@ def step(self, action):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
+        policy_model_name="mock-model",
     )
     # Ensure base_conversation_token_ids doesn't shift observation slicing in test
     generator.base_conversation_token_ids = []
@@ -1131,6 +1141,7 @@ def step(self, action):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
+        policy_model_name="mock-model",
     )
     # Force retokenize path regardless of model resolution logic if needed
     generator.custom_chat_template = "<custom>"
@@ -1219,6 +1230,7 @@ def mock_make_func(*args, **kwargs):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
+        policy_model_name="mock-model",
     )
 
     prompt = [{"role": "user", "content": "Q?"}]
@@ -1310,6 +1322,7 @@ def mock_make_func(*args, **kwargs):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
+        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []
 
@@ -1427,6 +1440,7 @@ def step(self, action):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
+        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []
 
diff --git a/tests/train/generators/test_skyrl_gym_generator_chat_templating.py b/tests/train/generators/test_skyrl_gym_generator_chat_templating.py
index ea10d8fad9..5e7f48a1d7 100644
--- a/tests/train/generators/test_skyrl_gym_generator_chat_templating.py
+++ b/tests/train/generators/test_skyrl_gym_generator_chat_templating.py
@@ -101,6 +101,7 @@ def _build_generator(tokenizer, model_name: str, chat_template_config, extra_ove
         skyrl_gym_cfg=env_cfg,
         inference_engine_client=None,  # to be replaced per-test
         tokenizer=tokenizer,
+        policy_model_name="mock-model",
     )
 
 
@@ -154,7 +155,7 @@ async def test_skyrl_gym_generator_chat_templating_exact(model_name, tokenizatio
     if "Qwen3" in model_name:
         mock_response_text = "<think>\nmock thinking\n</think>\n\n" + mock_response_text
 
-    def mock_generate(input_batch):
+    def mock_generate(input_batch, model=None):
         num_prompts = len(input_batch["prompts"]) if "prompts" in input_batch else len(input_batch["prompt_token_ids"])
 
         mock_llm_output_text = mock_response_text + tokenizer.eos_token
@@ -351,7 +352,7 @@ async def make_generator(append_flag: bool):
         mock_llm = MagicMock()
 
         # The LLM engine will generate and return the stop tag, but no EOS token ID.
-        def mock_generate(input_batch):
+        def mock_generate(input_batch, model=None):
             num_prompts = (
                 len(input_batch["prompts"]) if "prompts" in input_batch else len(input_batch["prompt_token_ids"])
             )
diff --git a/tests/train/generators/test_skyrl_vlm_generator.py b/tests/train/generators/test_skyrl_vlm_generator.py
index df02ce3ee3..4b505f8595 100644
--- a/tests/train/generators/test_skyrl_vlm_generator.py
+++ b/tests/train/generators/test_skyrl_vlm_generator.py
@@ -90,6 +90,7 @@ def _build_vlm_generator(tokenizer):
         skyrl_gym_cfg=env_cfg,
         inference_engine_client=mock_client,
         tokenizer=tokenizer,
+        policy_model_name="mock-model",
     )
     return generator
 
@@ -113,7 +114,7 @@ async def mock_render(request_payload):
 def _make_mock_llm(tokenizer, response_text: str):
     """Create an AsyncMock for the inference engine's generate method."""
 
-    async def mock_generate(input_batch):
+    async def mock_generate(input_batch, model=None):
         num_prompts = len(input_batch["prompt_token_ids"])
         text_with_eos = response_text + tokenizer.eos_token
         ids = tokenizer.encode(text_with_eos, add_special_tokens=False)

From 68ed1421243650fbc1c93fc3ec0f6006ff164f30 Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Mon, 4 May 2026 21:07:14 +0000
Subject: [PATCH 02/21] [docs] Add Multi-LoRA Megatron Tinker design doc (v1)

Adds the design write-up for multi-tenant LoRA training on the Megatron
backend exposed via the Tinker API. v1 is training-only; sampling and
adapter-only checkpoint export are deferred. Implementation follows on
the multi_lora branch.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .../content/docs/tinker/multi_lora_design.mdx | 110 ++++++++++++++++++
 1 file changed, 110 insertions(+)
 create mode 100644 docs/content/docs/tinker/multi_lora_design.mdx

diff --git a/docs/content/docs/tinker/multi_lora_design.mdx b/docs/content/docs/tinker/multi_lora_design.mdx
new file mode 100644
index 0000000000..e5e0e6b54c
--- /dev/null
+++ b/docs/content/docs/tinker/multi_lora_design.mdx
@@ -0,0 +1,110 @@
+---
+title: "Multi-LoRA on Megatron — Design"
+---
+
+# Multi-LoRA on Megatron — Design (v1)
+
+This document describes the design for multi-tenant LoRA training on the SkyRL Megatron Tinker backend. It is the in-repo companion to the implementation tracked on the `multi_lora` branch.
+
+## Why
+
+Today the SkyRL-Train backend exposed via the Tinker API is single-tenant: a second `create_model` is rejected at `skyrl/backends/skyrl_train_backend.py:342`, and `delete_model` does a full `ray.shutdown()` (line 404) so a fresh model can be created. This is documented under [Single-tenant LoRA](./limitations#single-tenant-lora).
+
+The driver for changing this is Trajectory AI, who want to run Tinker workloads on their own hardware and need many tenants on a shared training pool. There is no Megatron multi-tenant SFT/RL framework today; only Prime-RL has first-class multi-tenancy and only on FSDP/CP/EP. A Megatron-backed solution is therefore both a user-requested feature and a meaningful differentiator.
+
+## v1 scope
+
+- **Training only**, exercised via the `tinker-cookbook` `sl_loop` SFT recipe.
+- One base model, multiple LoRA adapters.
+- Fixed `(rank, alpha, target_modules)` across all adapters. Mismatched configs on a second `create_model` are hard-rejected with a clear `ValueError`.
+- `sample()` and `save_sampler_checkpoint()` raise if more than one adapter is registered. Per-adapter sampling and per-adapter vLLM weight sync are explicitly deferred.
+- The FFT (no-LoRA) path stays single-tenant — the relaxation is gated behind `lora_config.rank > 0`.
+
+## Strategy
+
+Keep one base model GPU-resident at all times. At any moment exactly one LoRA adapter is "live" in the model + optimizer. A swap is `tensor.copy_()` of LoRA buffer params and `DistributedOptimizer` fp32-main / `exp_avg` / `exp_avg_sq` between live GPU storage and per-adapter pinned-CPU slots.
+
+The per-adapter slot store (the `AdapterStore`) lives **on each `PolicyWorker`** because Megatron's `DistributedOptimizer` shards optimizer state across DP ranks; each rank owns its own slice and must snapshot/restore it locally. The controller (`SkyRLTrainBackend`) holds only `model_id → role` maps; the dispatch layer (`WorkerDispatch`) fans `swap_to_adapter(model_id)` out to all policy actors.
+
+The swap is **implicit** at the top of every per-model dispatch entry point: `forward`, `forward_backward`, `optim_step`, `set_lr`, `save_checkpoint`, `load_checkpoint`. Callers do not need to swap manually.
+
+## Why a buffer-level swap is correct
+
+Megatron's `DistributedDataParallel` filters out frozen params before constructing the param-and-grad buffers (`Megatron-LM/core_v0.16.0/megatron/core/distributed/distributed_data_parallel.py:139-141`). Combined with the LoRA pre-wrap hook at `megatron_worker.py:454-460` and `bridge.peft.lora.LoRA` setting `requires_grad=False` on base `to_wrap` params, the DDP `param_and_grad_buffer.param_data` contains **only LoRA A/B params**. Frozen base weights live as plain `nn.Parameter`s outside the buffer (which is exactly what the LoRA-aware branch in `megatron_utils.py:158-170` already handles).
+
+Buffer-level `tensor.copy_()` therefore swaps adapter-only state. The base model stays GPU-resident and is shared across all tenants.
+
+## Three storages per LoRA param
+
+For each LoRA `nn.Parameter` `p`, three independent storages must be swapped:
+
+1. The bf16 view in `mc.buffers[i].param_data` (or `mc.expert_parallel_buffers`).
+2. The fp32 main copy in `_opt.shard_fp32_from_float16_groups[g][i]` — independent storage, not a view.
+3. The Adam moments in `_opt.optimizer.state[main_param]`, keyed by the **fp32 main param**: `exp_avg`, `exp_avg_sq`.
+
+Param-object identity is preserved across `param.data.copy_(...)`, so optimizer state-dict keys remain valid. Grads are not swapped — `optimizer.zero_grad()` runs after every step (`megatron_strategy.py:215`), so they're zero at swap time.
+
+## Pristine slot
+
+Adam allocates `exp_avg` / `exp_avg_sq` lazily on the first non-trivial step. Megatron exposes `DistributedOptimizer._init_optimizer_states_with_dummy_values()` (in `distrib_optimizer.py`), which materialises state without a real fwd+bwd. The first `create_model` call:
+
+1. Builds the policy worker and its `DistributedOptimizer` as today.
+2. Calls `_init_optimizer_states_with_dummy_values()` on each underlying optimizer to materialise `exp_avg` / `exp_avg_sq`.
+3. Snapshots the freshly-initialised LoRA state into the `AdapterStore`'s pristine slot.
+
+Every subsequent `create_model("X")` allocates a fresh slot for `X` and copies the pristine slot's contents into it. A new tenant therefore starts from a freshly-initialised LoRA (kaiming-A + zero-B + zero optimizer state).
+
+## Concurrency
+
+`DistributedOptimizer.step()` issues DP-group collectives (reduce-scatter on grads, all-gather on updated params). Mixed adapter identity across DP ranks would corrupt these collectives. Therefore each `swap_to` ends with a `dist.barrier(group=mpu.get_data_parallel_group())` — and PP/TP equivalents where relevant — to ensure all ranks agree on the live adapter before the next collective begins.
+
+A `torch.cuda.current_stream().synchronize()` between the save and restore halves of a swap guarantees that `non_blocking=True` D2H copies complete before the corresponding GPU storage is overwritten.
+
+## Concrete `swap_to(adapter_id)` algorithm
+
+Per worker, all under `torch.no_grad()`:
+
+1. `dist.barrier(dp_group)` — wait for the previous adapter's last collective to finish.
+2. Save current adapter into its slot:
+   - For each `mc` and each `buffer ∈ mc.buffers + mc.expert_parallel_buffers`: `slot.cpu_param_data[mc][i].copy_(buffer.param_data, non_blocking=True)`.
+   - For each `_opt ∈ _iter_opts(self.optimizer)` and each `(g, i)`: `slot.cpu_main_param[g][i].copy_(_opt.shard_fp32_from_float16_groups[g][i], non_blocking=True)`. Then `slot.cpu_exp_avg[g][i].copy_(state['exp_avg'], non_blocking=True)`, `slot.cpu_exp_avg_sq[g][i].copy_(state['exp_avg_sq'], non_blocking=True)`.
+3. `torch.cuda.current_stream().synchronize()` — D2H complete.
+4. Load target adapter — same loops in reverse, copying CPU → GPU into the same storages.
+5. `torch.cuda.current_stream().synchronize()` — H2D complete.
+6. `dist.barrier(dp_group)` — agreement on live adapter before next collective.
+
+## Per-AdapterSlot CPU storage
+
+Per worker, pinned memory:
+
+- `cpu_param_data[mc][buf_idx]` — bf16, one tensor per `mc.buffers + mc.expert_parallel_buffers` entry, shape matches the bucket.
+- `cpu_main_param[g][i]` — fp32, shape matches `shard_fp32_from_float16_groups[g][i]`.
+- `cpu_exp_avg[g][i]`, `cpu_exp_avg_sq[g][i]` — fp32, same shapes.
+
+Frozen base weights are not duplicated per adapter; they live in their own pinned storage already managed by `offload_megatron_model_to_cpu`.
+
+## Files to add / modify
+
+### New
+
+- `skyrl/backends/skyrl_train/workers/megatron/adapter_store.py` — `AdapterSlot` + `AdapterStore`.
+
+### Modified
+
+- `skyrl/backends/skyrl_train/workers/megatron/megatron_worker.py` — construct `AdapterStore` after optimizer init; expose Ray-callable `prime_optimizer_state`, `register_pristine_adapter`, `register_adapter`, `delete_adapter`, `swap_to_adapter`.
+- `skyrl/backends/skyrl_train/workers/worker_dispatch.py` — `ensure_active_adapter(model_id)`; thread `model_id` kwarg through `forward`, `forward_backward`, `optim_step`, `set_lr`, `save_checkpoint`, `load_checkpoint`.
+- `skyrl/backends/skyrl_train_backend.py` — relax the single-policy gate when `lora_config.rank > 0`; rework `delete_model` to only `ray.shutdown()` on last adapter; pass `model_id` into every dispatch call; raise on `sample()` / `save_sampler_checkpoint` if more than one adapter is registered.
+
+## Verification
+
+- `pytest tests/tinker/test_multi_lora_megatron.py` — GPU-gated integration test that creates A, trains, creates B from pristine, trains, switches back to A, and asserts state is preserved bit-for-bit. Negative tests for rank mismatch and `sample()` with two adapters.
+- Existing `tests/tinker/test_api.py` continues to pass (single-tenant path unchanged).
+- End-to-end smoke (manual): launch the Tinker server with `trainer.strategy=megatron`, run two `tinker_cookbook.recipes.sl_loop` clients with distinct `model_id`s in parallel against `base_url=http://localhost:8000`, verify both converge on their respective tasks and GPU memory stays bounded.
+
+## Out of scope (explicit non-goals for v1)
+
+- Per-adapter `sample()` / vLLM weight sync — `sample` raises if more than one adapter is registered.
+- Adapter-only checkpoint export — `save_checkpoint` still saves the whole base+LoRA state per swap.
+- Variable rank / alpha / target_modules across adapters.
+- Critic role multi-tenancy — Megatron critic is `NotImplementedError` today; no change.
+- `HybridDeviceOptimizer` path — TODO comment only.

From c0c3a58c2ecbf9523c88b0cd70324e81f71fbb59 Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Mon, 4 May 2026 21:10:13 +0000
Subject: [PATCH 03/21] [multi-lora] Add AdapterStore for per-worker LoRA slot
 bookkeeping

New module holding per-adapter pinned-CPU snapshots of the LoRA bucket
params + DistributedOptimizer fp32-main + Adam state on each Megatron
PolicyWorker. swap_to() walks mc.buffers + expert_parallel_buffers and
shard_fp32_from_float16_groups, doing tensor.copy_() in both directions
under torch.no_grad with dp_group barriers + cuda stream syncs.

Also includes a sanity check that every trainable param under DDP
buffers is a LoRA adapter param (named "...adapter..."), so a future
regression that unfreezes a non-LoRA param fails loudly at registration
rather than silently corrupting state.

Wiring into PolicyWorker / WorkerDispatch / SkyRLTrainBackend follows
in subsequent commits.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .../workers/megatron/adapter_store.py         | 350 ++++++++++++++++++
 1 file changed, 350 insertions(+)
 create mode 100644 skyrl/backends/skyrl_train/workers/megatron/adapter_store.py

diff --git a/skyrl/backends/skyrl_train/workers/megatron/adapter_store.py b/skyrl/backends/skyrl_train/workers/megatron/adapter_store.py
new file mode 100644
index 0000000000..a62b551b10
--- /dev/null
+++ b/skyrl/backends/skyrl_train/workers/megatron/adapter_store.py
@@ -0,0 +1,350 @@
+"""Per-worker store of LoRA adapter weights and optimizer state.
+
+Holds one CPU-pinned snapshot per registered model_id plus a single pristine
+slot used to seed newly-created adapters. At any moment exactly one adapter is
+"live" in the worker's `actor_module` + `DistributedOptimizer`; swap_to() moves
+LoRA bucket params and DistributedOptimizer fp32-main / Adam state between live
+GPU storage and the per-adapter CPU slot via tensor.copy_().
+
+See docs/content/docs/tinker/multi_lora_design.mdx for the full design.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any, Iterable, List, Optional, Tuple
+
+import torch
+import torch.distributed as dist
+from megatron.core import parallel_state as mpu
+from megatron.core.distributed import DistributedDataParallel as DDP
+from megatron.core.optimizer import ChainedOptimizer
+
+
+def _iter_opts(opt) -> List[Any]:
+    """Yield underlying Megatron optimizers, unwrapping ChainedOptimizer."""
+    if isinstance(opt, ChainedOptimizer):
+        return list(opt.chained_optimizers)
+    return [opt]
+
+
+def _iter_buffers(model_chunks) -> Iterable[Tuple[int, int, Any]]:
+    """Yield (mc_idx, buf_idx, buffer) for every LoRA-trainable DDP buffer."""
+    for mc_idx, mc in enumerate(model_chunks):
+        if not isinstance(mc, DDP):
+            continue
+        bufs = list(mc.buffers) + list(mc.expert_parallel_buffers)
+        for buf_idx, buf in enumerate(bufs):
+            yield mc_idx, buf_idx, buf
+
+
+def _new_pinned_like(t: torch.Tensor) -> torch.Tensor:
+    """Allocate a pinned-CPU tensor with the same shape/dtype as t."""
+    return torch.empty_like(t, device="cpu").pin_memory()
+
+
+def _expected_lora_param_check(model_chunks) -> None:
+    """Sanity-check: every trainable param under DDP buffers is a LoRA adapter param.
+
+    Megatron's DDP filters out requires_grad=False params before bucket
+    construction. With the LoRA pre-wrap hook freezing base params, only
+    LoRA A/B params should remain. If a future change breaks this invariant
+    (e.g. an unfrozen bias or new trainable head), we want to fail loudly
+    rather than silently swap the wrong tensors.
+    """
+    for mc_idx, _buf_idx, buf in _iter_buffers(model_chunks):
+        for param in getattr(buf, "params", []):
+            mc = model_chunks[mc_idx]
+            name = next(
+                (n for n, p in mc.named_parameters() if p is param),
+                None,
+            )
+            if name is None:
+                continue
+            if "adapter" not in name:
+                raise RuntimeError(
+                    f"AdapterStore: trainable non-adapter param '{name}' found in "
+                    f"DDP buffer {mc_idx}/{_buf_idx}; multi-LoRA swap would "
+                    f"corrupt this param. Refusing to register."
+                )
+
+
+@dataclass(frozen=True)
+class LoraSignature:
+    """Immutable identity of a LoRA configuration. All registered adapters
+    must share the same signature; otherwise tensor shapes won't match across
+    swaps."""
+
+    rank: int
+    alpha: int
+    target_modules: Tuple[str, ...]
+    lora_type: str
+    tp_size: int
+    pp_size: int
+    ep_size: int
+
+    @classmethod
+    def from_lora_config(cls, lora_config, lora_type: str = "lora") -> "LoraSignature":
+        targets = lora_config.target_modules
+        if isinstance(targets, str):
+            targets_tuple = (targets,)
+        else:
+            targets_tuple = tuple(targets)
+        return cls(
+            rank=int(lora_config.rank),
+            alpha=int(lora_config.alpha),
+            target_modules=targets_tuple,
+            lora_type=lora_type,
+            tp_size=mpu.get_tensor_model_parallel_world_size(),
+            pp_size=mpu.get_pipeline_model_parallel_world_size(),
+            ep_size=(
+                mpu.get_expert_model_parallel_world_size()
+                if hasattr(mpu, "get_expert_model_parallel_world_size")
+                else 1
+            ),
+        )
+
+
+@dataclass
+class AdapterSlot:
+    """Per-adapter pinned-CPU storage mirroring the live GPU LoRA state.
+
+    Layout:
+      cpu_param_data[mc_idx] -> list[Tensor], one per buffer in
+          (mc.buffers + mc.expert_parallel_buffers).
+      cpu_main_param[opt_idx][g] -> list[Tensor], shapes matching
+          opt.shard_fp32_from_float16_groups[g].
+      cpu_opt_state[opt_idx][g][i] -> dict[str, Tensor], mirroring
+          opt.optimizer.state[main_param] for every tensor-valued entry
+          (exp_avg, exp_avg_sq, step, ...).
+    """
+
+    cpu_param_data: List[List[torch.Tensor]] = field(default_factory=list)
+    cpu_main_param: List[List[List[torch.Tensor]]] = field(default_factory=list)
+    cpu_opt_state: List[List[List[dict]]] = field(default_factory=list)
+    step_count: int = 0
+
+
+class AdapterStore:
+    """Per-worker registry of LoRA adapter slots.
+
+    One AdapterStore lives on each Megatron PolicyWorker. It owns CPU storage
+    for every registered adapter plus a pristine template; the live GPU model
+    + optimizer always reflect the slot identified by `current_id`.
+
+    Operations are local: snapshot/restore is a series of tensor.copy_()s that
+    issue no collectives. Callers are responsible for the surrounding
+    dist.barrier() (we recommend before and after the swap; see swap_to docs).
+    """
+
+    def __init__(self) -> None:
+        self._slots: dict[str, AdapterSlot] = {}
+        self._pristine: Optional[AdapterSlot] = None
+        self._current_id: Optional[str] = None
+        self._signature: Optional[LoraSignature] = None
+
+    @property
+    def current_id(self) -> Optional[str]:
+        return self._current_id
+
+    @property
+    def signature(self) -> Optional[LoraSignature]:
+        return self._signature
+
+    def has(self, model_id: str) -> bool:
+        return model_id in self._slots
+
+    def num_adapters(self) -> int:
+        return len(self._slots)
+
+    # ------------------------------------------------------------------
+    # Slot allocation helpers
+    # ------------------------------------------------------------------
+
+    def _allocate_empty_slot(self, model_chunks, optimizer) -> AdapterSlot:
+        slot = AdapterSlot()
+        # Param data: one pinned bf16 tensor per (mc, buffer).
+        for mc_idx, _buf_idx, buf in _iter_buffers(model_chunks):
+            while len(slot.cpu_param_data) <= mc_idx:
+                slot.cpu_param_data.append([])
+            slot.cpu_param_data[mc_idx].append(_new_pinned_like(buf.param_data))
+        # Main params + optimizer state: per (opt_idx, group, param_idx).
+        for _opt in _iter_opts(optimizer):
+            opt_main: List[List[torch.Tensor]] = []
+            opt_state: List[List[dict]] = []
+            groups = getattr(_opt, "shard_fp32_from_float16_groups", None) or []
+            for g, group in enumerate(groups):
+                main_g: List[torch.Tensor] = []
+                state_g: List[dict] = []
+                for main_param in group:
+                    main_g.append(_new_pinned_like(main_param))
+                    state = _opt.optimizer.state.get(main_param, {})
+                    state_g.append(
+                        {k: _new_pinned_like(v) for k, v in state.items() if isinstance(v, torch.Tensor)}
+                    )
+                opt_main.append(main_g)
+                opt_state.append(state_g)
+            slot.cpu_main_param.append(opt_main)
+            slot.cpu_opt_state.append(opt_state)
+        return slot
+
+    @torch.no_grad()
+    def _snapshot(self, slot: AdapterSlot, model_chunks, optimizer) -> None:
+        """Copy live GPU state into `slot` (CPU)."""
+        for mc_idx, buf_idx, buf in _iter_buffers(model_chunks):
+            slot.cpu_param_data[mc_idx][buf_idx].copy_(buf.param_data, non_blocking=True)
+        for opt_idx, _opt in enumerate(_iter_opts(optimizer)):
+            groups = getattr(_opt, "shard_fp32_from_float16_groups", None) or []
+            for g, group in enumerate(groups):
+                for i, main_param in enumerate(group):
+                    slot.cpu_main_param[opt_idx][g][i].copy_(main_param, non_blocking=True)
+                    state = _opt.optimizer.state.get(main_param, {})
+                    cpu_state = slot.cpu_opt_state[opt_idx][g][i]
+                    for k, v in state.items():
+                        if isinstance(v, torch.Tensor) and k in cpu_state:
+                            cpu_state[k].copy_(v, non_blocking=True)
+
+    @torch.no_grad()
+    def _restore(self, slot: AdapterSlot, model_chunks, optimizer) -> None:
+        """Copy `slot` (CPU) into live GPU state."""
+        for mc_idx, buf_idx, buf in _iter_buffers(model_chunks):
+            buf.param_data.copy_(slot.cpu_param_data[mc_idx][buf_idx], non_blocking=True)
+        for opt_idx, _opt in enumerate(_iter_opts(optimizer)):
+            groups = getattr(_opt, "shard_fp32_from_float16_groups", None) or []
+            for g, group in enumerate(groups):
+                for i, main_param in enumerate(group):
+                    main_param.copy_(slot.cpu_main_param[opt_idx][g][i], non_blocking=True)
+                    state = _opt.optimizer.state.get(main_param, {})
+                    cpu_state = slot.cpu_opt_state[opt_idx][g][i]
+                    for k, v in state.items():
+                        if isinstance(v, torch.Tensor) and k in cpu_state:
+                            v.copy_(cpu_state[k], non_blocking=True)
+
+    # ------------------------------------------------------------------
+    # Public API used by the worker
+    # ------------------------------------------------------------------
+
+    def register_pristine(self, model_chunks, optimizer, signature: LoraSignature) -> None:
+        """Capture the freshly-initialised LoRA state as the pristine template.
+
+        Must be called once per worker, after the optimizer state has been
+        materialised (e.g. via DistributedOptimizer._init_optimizer_states_with_dummy_values).
+        Subsequent registrations will copy this slot to seed new adapters.
+        """
+        if self._pristine is not None:
+            raise RuntimeError("AdapterStore.register_pristine called twice")
+        _expected_lora_param_check(model_chunks)
+        self._signature = signature
+        self._pristine = self._allocate_empty_slot(model_chunks, optimizer)
+        self._snapshot(self._pristine, model_chunks, optimizer)
+
+    @torch.no_grad()
+    def create(self, model_id: str, model_chunks, optimizer, signature: LoraSignature) -> None:
+        """Register a new adapter slot.
+
+        - First registration: this is also the live adapter; allocate a slot
+          but skip the pristine→slot copy because the live state already
+          equals pristine. `current_id` becomes `model_id`.
+        - Subsequent registrations: allocate slot and copy pristine → slot.
+          Live state is unchanged (no swap). The new adapter only becomes
+          live when the next `swap_to(model_id)` is issued.
+        """
+        if self._signature is None:
+            raise RuntimeError("AdapterStore.create called before register_pristine")
+        if signature != self._signature:
+            raise ValueError(
+                f"AdapterStore: lora signature mismatch for '{model_id}'. "
+                f"Pristine={self._signature}, requested={signature}. "
+                f"Multi-LoRA requires identical (rank, alpha, target_modules, "
+                f"lora_type, tp/pp/ep sizes) across all adapters."
+            )
+        if model_id in self._slots:
+            raise ValueError(f"AdapterStore: adapter '{model_id}' already registered")
+
+        slot = self._allocate_empty_slot(model_chunks, optimizer)
+        if self._current_id is None:
+            # First adapter: live state IS pristine; slot will be filled on
+            # the next snapshot (i.e. swap-away). Treat live as authoritative.
+            self._current_id = model_id
+        else:
+            # Seed the new slot from pristine.
+            self._copy_slot(self._pristine, slot)
+        self._slots[model_id] = slot
+
+    @torch.no_grad()
+    def _copy_slot(self, src: AdapterSlot, dst: AdapterSlot) -> None:
+        """CPU→CPU copy used to seed a new slot from the pristine template."""
+        for mc_idx, mc_buffers in enumerate(src.cpu_param_data):
+            for buf_idx, t in enumerate(mc_buffers):
+                dst.cpu_param_data[mc_idx][buf_idx].copy_(t)
+        for opt_idx, opt_groups in enumerate(src.cpu_main_param):
+            for g, group in enumerate(opt_groups):
+                for i, t in enumerate(group):
+                    dst.cpu_main_param[opt_idx][g][i].copy_(t)
+        for opt_idx, opt_groups in enumerate(src.cpu_opt_state):
+            for g, group in enumerate(opt_groups):
+                for i, state in enumerate(group):
+                    for k, v in state.items():
+                        if k in dst.cpu_opt_state[opt_idx][g][i]:
+                            dst.cpu_opt_state[opt_idx][g][i][k].copy_(v)
+
+    @torch.no_grad()
+    def delete(self, model_id: str) -> None:
+        """Drop the slot for `model_id`.
+
+        If `model_id` was the current adapter, `current_id` is cleared. The
+        live GPU state is left untouched (it now mirrors a deleted adapter);
+        the next `swap_to` will overwrite it.
+        """
+        if model_id not in self._slots:
+            raise KeyError(f"AdapterStore: unknown adapter '{model_id}'")
+        del self._slots[model_id]
+        if self._current_id == model_id:
+            self._current_id = None
+
+    @torch.no_grad()
+    def swap_to(self, model_id: str, model_chunks, optimizer) -> None:
+        """Make `model_id` the live adapter on this worker.
+
+        Algorithm (all under torch.no_grad):
+            1. dist.barrier(dp_group)
+            2. snapshot live → current's slot (skipped if current_id is None)
+            3. cuda stream sync (D2H done)
+            4. restore target's slot → live
+            5. cuda stream sync (H2D done)
+            6. dist.barrier(dp_group)
+
+        Caller responsibility: the trailing barrier guarantees all DP ranks
+        agree on the live adapter before the next collective. TP/PP/EP groups
+        do not need barriers because the swap is identical-shape on all
+        ranks within those groups (LoRA signature is fixed).
+        """
+        if model_id not in self._slots:
+            raise KeyError(f"AdapterStore: unknown adapter '{model_id}'")
+        if self._current_id == model_id:
+            return  # no-op fast path
+
+        dp_group = mpu.get_data_parallel_group()
+        if dist.is_available() and dist.is_initialized():
+            dist.barrier(group=dp_group)
+
+        if self._current_id is not None:
+            current_slot = self._slots[self._current_id]
+            self._snapshot(current_slot, model_chunks, optimizer)
+            torch.cuda.current_stream().synchronize()
+
+        target_slot = self._slots[model_id]
+        self._restore(target_slot, model_chunks, optimizer)
+        torch.cuda.current_stream().synchronize()
+
+        self._current_id = model_id
+
+        if dist.is_available() and dist.is_initialized():
+            dist.barrier(group=dp_group)
+
+    def clear(self) -> None:
+        """Drop all slots (used at full-shutdown reset)."""
+        self._slots.clear()
+        self._pristine = None
+        self._current_id = None
+        self._signature = None

From e92389462ef39e3a12802c390e77287298cc4791 Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Mon, 4 May 2026 21:12:19 +0000
Subject: [PATCH 04/21] [multi-lora] Wire AdapterStore into
 MegatronPolicyWorkerBase
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Adds an `adapter_store: AdapterStore | None` attribute on the policy
worker (allocated only when LoRA is active so the FFT path is unchanged)
plus five Ray-callable methods:

- prime_optimizer_state — calls Megatron's
  DistributedOptimizer._init_optimizer_states_with_dummy_values() so
  exp_avg/exp_avg_sq exist before we snapshot the pristine slot.
- register_pristine_adapter — derives a LoraSignature from the worker's
  own lora config + parallel state, snapshots live state into pristine.
- register_adapter(model_id) — allocates a fresh slot; first call uses
  live as the slot, subsequent calls seed from pristine.
- delete_adapter(model_id) — drops a slot.
- swap_to_adapter(model_id) — local tensor.copy_() between live and slot
  storages plus dp_group barriers.

Plus an adapter_store_state() diagnostic for tests. Orchestration from
the controller follows in subsequent commits.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .../workers/megatron/megatron_worker.py       | 77 +++++++++++++++++++
 1 file changed, 77 insertions(+)

diff --git a/skyrl/backends/skyrl_train/workers/megatron/megatron_worker.py b/skyrl/backends/skyrl_train/workers/megatron/megatron_worker.py
index e00287dba3..d305e6ce61 100644
--- a/skyrl/backends/skyrl_train/workers/megatron/megatron_worker.py
+++ b/skyrl/backends/skyrl_train/workers/megatron/megatron_worker.py
@@ -45,6 +45,11 @@
     WeightChunk,
     WeightExtractor,
 )
+from skyrl.backends.skyrl_train.workers.megatron.adapter_store import (
+    AdapterStore,
+    LoraSignature,
+    _iter_opts,
+)
 from skyrl.backends.skyrl_train.workers.megatron.megatron_model_wrapper import (
     MegatronModelWrapper,
 )
@@ -524,6 +529,9 @@ def __init__(self, **kwargs):
         self.optimizer: DistributedOptimizer = None
         self.profiler: Profiler = None
         self._is_lora = self.cfg.policy.model.lora.rank > 0
+        # Per-worker store of LoRA adapter snapshots. Allocated only for the
+        # LoRA path; FFT runs single-tenant exactly as before.
+        self.adapter_store: Optional[AdapterStore] = AdapterStore() if self._is_lora else None
 
     def offload_to_cpu(self, pin_memory=True, non_blocking=True, offload_optimizer=True, offload_model=True):
         self._set_numa_affinity(torch.distributed.get_rank() % torch.cuda.device_count())
@@ -920,6 +928,75 @@ def _set_pad_token_id(self, pad_token_id):
         # this already gets set in the init_model method
         pass
 
+    # ------------------------------------------------------------------
+    # Multi-LoRA / AdapterStore Ray-callable methods
+    # ------------------------------------------------------------------
+
+    def prime_optimizer_state(self) -> None:
+        """Materialise DistributedOptimizer state (exp_avg / exp_avg_sq).
+
+        Adam's state tensors are allocated lazily on the first non-trivial
+        step; without priming, the pristine snapshot would miss them.
+        Megatron exposes ``_init_optimizer_states_with_dummy_values()`` which
+        zero-fills grads + steps once + zero_grads, leaving the model weights
+        unchanged.
+        """
+        if not self._is_lora:
+            raise RuntimeError("prime_optimizer_state is only used on the LoRA path")
+        for _opt in _iter_opts(self.optimizer):
+            init_fn = getattr(_opt, "_init_optimizer_states_with_dummy_values", None)
+            if init_fn is not None:
+                init_fn()
+
+    def register_pristine_adapter(self) -> None:
+        """Capture the current (freshly-initialised) LoRA state as the
+        pristine slot. Must be called once per worker, after
+        prime_optimizer_state.
+        """
+        if self.adapter_store is None:
+            raise RuntimeError("AdapterStore not initialised (FFT path)")
+        signature = LoraSignature.from_lora_config(
+            self.cfg.policy.model.lora,
+            lora_type=self.cfg.policy.megatron_config.lora_config.lora_type,
+        )
+        self.adapter_store.register_pristine(self.actor_module, self.optimizer, signature)
+
+    def register_adapter(self, model_id: str) -> None:
+        """Register a new LoRA adapter slot. The first call uses the live
+        state as the slot; subsequent calls seed from pristine.
+        """
+        if self.adapter_store is None:
+            raise RuntimeError("AdapterStore not initialised (FFT path)")
+        signature = self.adapter_store.signature
+        if signature is None:
+            raise RuntimeError("register_adapter called before register_pristine_adapter")
+        self.adapter_store.create(model_id, self.actor_module, self.optimizer, signature)
+
+    def delete_adapter(self, model_id: str) -> None:
+        if self.adapter_store is None:
+            raise RuntimeError("AdapterStore not initialised (FFT path)")
+        self.adapter_store.delete(model_id)
+
+    def swap_to_adapter(self, model_id: str) -> None:
+        """Make ``model_id`` the live adapter on this worker. No-op if it
+        already is. Issues local tensor.copy_()s + dp_group barriers.
+        """
+        if self.adapter_store is None:
+            return  # FFT path: no-op
+        self.adapter_store.swap_to(model_id, self.actor_module, self.optimizer)
+
+    def adapter_store_state(self) -> dict:
+        """Diagnostic: return current_id + registered model_ids. Cheap; useful
+        for tests."""
+        if self.adapter_store is None:
+            return {"enabled": False}
+        return {
+            "enabled": True,
+            "current_id": self.adapter_store.current_id,
+            "registered": list(self.adapter_store._slots.keys()),
+            "num_adapters": self.adapter_store.num_adapters(),
+        }
+
 
 class MegatronRefWorkerBase(MegatronWorker, RefWorkerBase):
     def __init__(self, **kwargs):

From 46c1658cd43e68a65eb0ebe4263e1e0ba2dc1f87 Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Mon, 4 May 2026 21:13:48 +0000
Subject: [PATCH 05/21] [multi-lora] Add ensure_active_adapter + model_id
 threading to dispatch
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

WorkerDispatch now exposes:
  - ensure_active_adapter(role, model_id): fans swap_to_adapter to all
    actors of `role`. No-op when model_id is None or the workers don't
    own an AdapterStore (FFT path).
  - prime_adapter_store(role, model_id): one-shot bootstrap for the very
    first create_model — primes optimizer state, registers pristine slot,
    registers the first adapter in one Ray-fanout sequence.
  - register_adapter / delete_adapter: per-call slot maintenance.

forward / forward_backward / forward_backward_from_staged / optim_step /
set_lr / save_checkpoint / load_checkpoint take an optional model_id and
call ensure_active_adapter after _ensure_on_gpu. Default None preserves
single-tenant behavior.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .../skyrl_train/workers/worker_dispatch.py    | 88 +++++++++++++++++--
 1 file changed, 83 insertions(+), 5 deletions(-)

diff --git a/skyrl/backends/skyrl_train/workers/worker_dispatch.py b/skyrl/backends/skyrl_train/workers/worker_dispatch.py
index fc6ad39d1a..a7759a5c64 100644
--- a/skyrl/backends/skyrl_train/workers/worker_dispatch.py
+++ b/skyrl/backends/skyrl_train/workers/worker_dispatch.py
@@ -74,6 +74,63 @@ def register_actor_group(self, model: str, actor_group: PPORayActorGroup) -> Non
         self._actor_groups[model] = actor_group
         self._gpu_state[model] = GPUState()
 
+    # ------------------------------------------------------------------
+    # Multi-LoRA: per-model adapter swap orchestration.
+    # ------------------------------------------------------------------
+
+    def ensure_active_adapter(self, role: str, model_id: Optional[str]) -> None:
+        """Make ``model_id`` the live LoRA adapter for ``role`` workers.
+
+        No-op when ``model_id is None`` (single-tenant / FFT path) or when
+        the workers don't have an AdapterStore (non-LoRA strategies).
+
+        Must be called *after* ``_ensure_on_gpu(role, ...)`` so the model
+        and optimizer storages are live before we tensor.copy_() into them.
+        """
+        if model_id is None or role not in self._actor_groups:
+            return
+        ray.get(
+            self._actor_groups[role].async_run_ray_method(
+                "pass_through", "swap_to_adapter", model_id
+            )
+        )
+
+    def prime_adapter_store(self, role: str, model_id: str) -> None:
+        """One-shot bootstrap on first create_model: prime the optimizer
+        state, register the pristine slot, and register the first adapter.
+
+        The model + optimizer must be on GPU when this is called (the
+        controller calls this immediately after _build_policy and before
+        any colocate_all offload).
+        """
+        if role not in self._actor_groups:
+            return
+        group = self._actor_groups[role]
+        ray.get(group.async_run_ray_method("pass_through", "prime_optimizer_state"))
+        ray.get(group.async_run_ray_method("pass_through", "register_pristine_adapter"))
+        ray.get(group.async_run_ray_method("pass_through", "register_adapter", model_id))
+
+    def register_adapter(self, role: str, model_id: str) -> None:
+        """Register a new adapter slot on every worker (subsequent
+        create_model). Pristine must already exist.
+        """
+        if role not in self._actor_groups:
+            return
+        ray.get(
+            self._actor_groups[role].async_run_ray_method(
+                "pass_through", "register_adapter", model_id
+            )
+        )
+
+    def delete_adapter(self, role: str, model_id: str) -> None:
+        if role not in self._actor_groups:
+            return
+        ray.get(
+            self._actor_groups[role].async_run_ray_method(
+                "pass_through", "delete_adapter", model_id
+            )
+        )
+
     def get_lcm_dp_size(self) -> int:
         """Get LCM of all models' dp_size."""
         import math
@@ -162,9 +219,12 @@ def mark_as_offloaded(self, model: str) -> None:
             return
         self._gpu_state[model] = GPUState()
 
-    def forward(self, model: str, data: TrainingInputBatch) -> TrainingOutputBatch:
+    def forward(
+        self, model: str, data: TrainingInputBatch, model_id: Optional[str] = None
+    ) -> TrainingOutputBatch:
         """Run inference forward pass. Only loads model (not optimizer)."""
         self._ensure_on_gpu(model, need_optimizer=False, need_model=True)
+        self.ensure_active_adapter(model, model_id)
 
         refs = self._actor_groups[model].async_run_ray_method("mesh", "forward", data=data)
         results = ray.get(refs)
@@ -201,6 +261,7 @@ def forward_backward(
         data: TrainingInputBatch,
         loss_fn: Optional[str] = None,
         loss_fn_config: Optional[Dict[str, Any]] = None,
+        model_id: Optional[str] = None,
     ) -> Dict[str, float]:
         """Run forward/backward pass. Needs model + optimizer.
 
@@ -212,11 +273,14 @@ def forward_backward(
                      normalized before dispatch.
             loss_fn_config: Optional config overrides for the loss function
                            (e.g., {"eps_clip_low": 0.1} for the regular PPO loss)
+            model_id: Optional Tinker model_id; when set, the corresponding
+                     LoRA adapter is swapped in before the forward/backward.
 
         Returns:
             Dictionary of training metrics
         """
         self._ensure_on_gpu(model, need_optimizer=True, need_model=True)
+        self.ensure_active_adapter(model, model_id)
 
         # Only pass kwargs that are not None (critic worker doesn't accept loss_fn)
         kwargs = {}
@@ -249,6 +313,7 @@ def forward_backward_from_staged(
         chunk_refs: List[ObjectRef],
         loss_fn: Optional[str] = None,
         loss_fn_config: Optional[Dict[str, Any]] = None,
+        model_id: Optional[str] = None,
     ) -> Dict[str, float]:
         """
         Run forward/backward pass using pre-staged per-DP chunks.
@@ -264,6 +329,7 @@ def forward_backward_from_staged(
             Aggregated metrics dict from training
         """
         self._ensure_on_gpu(model, need_optimizer=True, need_model=True)
+        self.ensure_active_adapter(model, model_id)
 
         # Only pass kwargs that are not None (critic worker doesn't accept loss_fn)
         kwargs = {}
@@ -283,21 +349,28 @@ def forward_backward_from_staged(
         self._save_memory_snapshot(model, "forward_backward")
         return statuses[0]
 
-    def optim_step(self, model: str) -> Optional[float]:
-        """Run optimizer step. Model should already be on GPU from forward_backward."""
+    def optim_step(self, model: str, model_id: Optional[str] = None) -> Optional[float]:
+        """Run optimizer step. Model should already be on GPU from forward_backward.
+
+        ``model_id`` is honored for safety (the previous forward_backward
+        already swapped the right adapter in, but a stale call ordering
+        would otherwise step on the wrong adapter's optimizer state).
+        """
+        self.ensure_active_adapter(model, model_id)
         refs = self._actor_groups[model].async_run_ray_method("pass_through", "optim_step")
         grad_norms = ray.get(refs)
 
         self._save_memory_snapshot(model, "optim_step")
         return grad_norms[0]
 
-    def set_lr(self, model: str, learning_rate: float) -> None:
+    def set_lr(self, model: str, learning_rate: float, model_id: Optional[str] = None) -> None:
         """Set learning rate for model's optimizer.
 
         This directly updates the optimizer's param_groups on all workers,
         bypassing the scheduler. Useful for external learning rate schedules.
         """
         self._ensure_on_gpu(model, need_optimizer=True, need_model=False)
+        self.ensure_active_adapter(model, model_id)
         ray.get(self._actor_groups[model].async_run_ray_method("pass_through", "set_lr", learning_rate=learning_rate))
 
     def set_algorithm_config(self, model: str, **kwargs) -> None:
@@ -311,9 +384,12 @@ def _save_memory_snapshot(self, model: str, tag: str) -> None:
             self._actor_groups[model].async_run_ray_method("pass_through", "save_memory_snapshot", tag=f"{model}_{tag}")
         )
 
-    def save_checkpoint(self, model: str, ckpt_dir: str, tokenizer=None) -> None:
+    def save_checkpoint(
+        self, model: str, ckpt_dir: str, tokenizer=None, model_id: Optional[str] = None
+    ) -> None:
         """Save checkpoint for model."""
         self._ensure_on_gpu(model, need_optimizer=True, need_model=True)
+        self.ensure_active_adapter(model, model_id)
 
         ray.get(
             self._actor_groups[model].async_run_ray_method(
@@ -327,9 +403,11 @@ def load_checkpoint(
         ckpt_dir: str,
         load_optimizer_states: bool = True,
         load_lr_scheduler_states: bool = True,
+        model_id: Optional[str] = None,
     ) -> None:
         """Load checkpoint for model."""
         self._ensure_on_gpu(model, need_optimizer=load_optimizer_states, need_model=True)
+        self.ensure_active_adapter(model, model_id)
 
         ray.get(
             self._actor_groups[model].async_run_ray_method(

From 90dc1782103fd3bcb2aa69b92340cb687e80bb3f Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Mon, 4 May 2026 21:17:19 +0000
Subject: [PATCH 06/21] [multi-lora] Allow multiple LoRA policy adapters in
 SkyRLTrainBackend

create_model now allows additional 'policy' models when LoRA is active
and the first policy model has been built. Subsequent calls validate
(rank, alpha, target_modules) match the first adapter's signature, then
register a new slot via WorkerDispatch.register_adapter. FFT (rank=0)
keeps the original single-tenant gate.

_build_policy takes the first model_id and, when LoRA is active, fires
the AdapterStore bootstrap (prime_optimizer_state +
register_pristine_adapter + register_adapter) on every worker before
the colocate_all offload while model + optimizer are still GPU-resident.

delete_model: when more than one model is registered and the role is a
LoRA policy, just drop the slot via dispatch.delete_adapter and pop the
controller-side maps. Last-adapter delete still does the full
ray.shutdown teardown so the runtime can be rebuilt cleanly.

Plumbed model_id through forward / forward_backward / optim_step /
set_lr / save_checkpoint / load_checkpoint dispatch calls so the active
adapter is swapped in on every per-model entry point.

sample() and save_sampler_checkpoint() refuse with a clear error when
more than one LoRA adapter is registered (v1 inference path is single-
tenant; per-adapter sampling is deferred).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 skyrl/backends/skyrl_train_backend.py | 126 +++++++++++++++++++++++---
 1 file changed, 112 insertions(+), 14 deletions(-)

diff --git a/skyrl/backends/skyrl_train_backend.py b/skyrl/backends/skyrl_train_backend.py
index a960b1ddee..519cab2941 100644
--- a/skyrl/backends/skyrl_train_backend.py
+++ b/skyrl/backends/skyrl_train_backend.py
@@ -137,6 +137,9 @@ def __init__(self, base_model: str, config: SkyRLTrainBackendOverrides):
         self._inference_engine_client = None
         self._inference_engines_initialized = False
         self._renderer = None
+        # Captured at first LoRA create_model; subsequent create_models must
+        # match this signature exactly. None when no LoRA model is registered.
+        self._base_lora_signature: tuple | None = None
 
         # New inference infrastructure
         self._server_group = None
@@ -214,10 +217,11 @@ def _split_model_pass_batch_by_model_id(
 
         return sub_batches
 
-    def _build_policy(self, PolicyWorker):
+    def _build_policy(self, PolicyWorker, model_id: str):
         cfg = self._cfg
         colocate_all = cfg.trainer.placement.colocate_all
         pg = self._colocate_pg
+        is_lora = cfg.trainer.policy.model.lora.rank > 0
 
         if colocate_all:
             assert pg is not None, "placement group must be created when colocate_all=True"
@@ -255,6 +259,15 @@ def _build_policy(self, PolicyWorker):
         )
         ray.get(policy_model.async_run_ray_method("pass_through", "_set_pad_token_id", self._tokenizer.pad_token_id))
 
+        # Multi-LoRA bootstrap: prime DistributedOptimizer state and snapshot
+        # the freshly-initialised LoRA into a per-worker pristine slot, then
+        # register the first adapter under `model_id`. Must happen while the
+        # model + optimizer are still GPU-resident (i.e. before the offload).
+        if is_lora:
+            ray.get(policy_model.async_run_ray_method("pass_through", "prime_optimizer_state"))
+            ray.get(policy_model.async_run_ray_method("pass_through", "register_pristine_adapter"))
+            ray.get(policy_model.async_run_ray_method("pass_through", "register_adapter", model_id))
+
         if colocate_all:
             policy_model.offload_to_cpu()
 
@@ -401,12 +414,52 @@ def _ensure_inference_engines(self):
         self.init_weight_sync_state()
         self._inference_engines_initialized = True
 
+    def _lora_signature_from(self, lora_config: types.LoraConfig) -> tuple:
+        targets = lora_config.target_modules
+        if isinstance(targets, str):
+            targets_tuple: tuple = (targets,)
+        else:
+            targets_tuple = tuple(targets)
+        return (int(lora_config.rank), int(lora_config.alpha), targets_tuple)
+
     def create_model(self, model_id: str, lora_config: types.LoraConfig, model_role: str = "policy") -> None:
         if model_id in self._model_ids_to_role:
             raise ValueError(f"Model '{model_id}' already exists")
-        if model_role in self._model_ids_to_role.values():
-            raise ValueError(f"SkyRLTrainBackend already has a '{model_role}' model")
 
+        is_lora = lora_config is not None and lora_config.rank > 0
+        is_first_policy = "policy" not in self._model_ids_to_role.values()
+
+        # Multi-LoRA path: allow additional policy adapters when LoRA is active
+        # and the first model has already been built. FFT (rank=0) keeps the
+        # original single-tenant gate.
+        if model_role == "policy" and not is_first_policy:
+            if not is_lora:
+                raise ValueError(
+                    "SkyRLTrainBackend already has a 'policy' model; multi-tenant "
+                    "training is only supported for LoRA (rank > 0)"
+                )
+            if self._base_lora_signature is None:
+                raise ValueError(
+                    "Cannot register an additional LoRA adapter: the first policy "
+                    "model was created without LoRA. Recreate the server with a "
+                    "LoRA-enabled first model."
+                )
+            new_signature = self._lora_signature_from(lora_config)
+            if new_signature != self._base_lora_signature:
+                raise ValueError(
+                    f"LoRA signature mismatch for model '{model_id}': "
+                    f"got (rank, alpha, target_modules)={new_signature}, "
+                    f"first adapter registered with {self._base_lora_signature}. "
+                    "Multi-LoRA requires identical (rank, alpha, target_modules) "
+                    "across all adapters in v1."
+                )
+            self._dispatch.register_adapter("policy", model_id)
+            self._model_ids_to_role[model_id] = model_role
+            self._model_metadata[model_id] = types.ModelMetadata(adapter_index=0, lora_config=lora_config)
+            logger.info(f"Registered additional LoRA adapter '{model_id}'")
+            return
+
+        # First-time setup OR critic creation (existing path).
         if model_role == "policy":
             self._cfg = _build_skyrl_train_config(self.base_model, self.config, lora_config)
 
@@ -428,8 +481,12 @@ def create_model(self, model_id: str, lora_config: types.LoraConfig, model_role:
                 raise ValueError(f"Unknown strategy type: {self._cfg.trainer.strategy}")
 
             logger.info("Building models.")
-            self._build_policy(PolicyWorker)
+            self._build_policy(PolicyWorker, model_id=model_id)
+            if is_lora:
+                self._base_lora_signature = self._lora_signature_from(lora_config)
         elif model_role == "critic":
+            if model_role in self._model_ids_to_role.values():
+                raise ValueError(f"SkyRLTrainBackend already has a '{model_role}' model")
             if "policy" not in self._model_ids_to_role.values():
                 raise ValueError("Create a policy model before creating a critic model")
             if self._cfg.trainer.strategy in ("fsdp", "fsdp2"):
@@ -464,11 +521,23 @@ def _create_colocate_pg(self):
         return ResolvedPlacementGroup(pg)
 
     def delete_model(self, model_id: str) -> None:
-        self._get_role(model_id)
+        role = self._get_role(model_id)
 
-        # Models in this backend share one Ray runtime and inference stack, so
-        # deleting any model tears down the whole backend and it will be
-        # re-initialized on the next create_model() call.
+        # Multi-LoRA: if more than one model is currently registered, drop just
+        # this adapter slot rather than tearing down the shared Ray runtime.
+        # The live GPU state may still mirror this adapter; it'll be
+        # overwritten on the next swap_to (no eager swap-away here).
+        if len(self._model_ids_to_role) > 1:
+            if role == "policy" and self._base_lora_signature is not None:
+                self._dispatch.delete_adapter("policy", model_id)
+                del self._model_ids_to_role[model_id]
+                self._model_metadata.pop(model_id, None)
+                logger.info(f"Removed LoRA adapter '{model_id}'")
+                return
+            # Fall through to teardown for non-LoRA roles or unexpected mixes.
+
+        # Last model (or non-LoRA path): tear down the shared Ray runtime.
+        # The Tinker engine will rebuild on the next create_model().
         logger.info(f"Deleting model {model_id}, shutting down shared SkyRL-Train runtime...")
         if self._server_group:
             self._server_group.shutdown()
@@ -485,6 +554,7 @@ def delete_model(self, model_id: str) -> None:
         self._inference_engines_initialized = False
         self._renderer = None
         self._colocate_pg = None
+        self._base_lora_signature = None
         logger.info(f"Successfully deleted model {model_id}")
 
     def _to_training_batch(self, prepared_batch: types.PreparedModelPassBatch, role: str) -> TrainingInputBatch:
@@ -730,18 +800,22 @@ def _forward_backward_single_model_batch(
             )
         loss_fn_config = next((c for c in prepared_batch.all_loss_fn_configs if c is not None), None)
         loss_fn, loss_fn_config = self._normalize_policy_loss_request(role, loss_fn, loss_fn_config)
+        # Single model_id per sub-batch (split upstream); pass it so the
+        # dispatch layer can swap to the right LoRA adapter before the op.
+        model_id = prepared_batch.all_model_ids[0] if prepared_batch.all_model_ids else None
         if role == "critic":
             self._dispatch.set_algorithm_config(
                 "critic",
                 value_clip=(loss_fn_config or {}).get("value_clip", self._cfg.trainer.algorithm.value_clip),
             )
-            data = self._dispatch.forward_backward("critic", batch)
+            data = self._dispatch.forward_backward("critic", batch, model_id=model_id)
         else:
             data = self._dispatch.forward_backward(
                 role,
                 batch,
                 loss_fn=loss_fn,
                 loss_fn_config=loss_fn_config,
+                model_id=model_id,
             )
 
         # Trim padding entries from loss_fn_outputs
@@ -798,7 +872,8 @@ def _forward_single_model_batch(
             self._cfg.trainer.micro_forward_batch_size_per_gpu if self._cfg.trainer.strategy == "megatron" else None
         )
         batch, pad_size = self._pad_batch(batch, micro_batch_size=micro_bs)
-        data = self._dispatch.forward(role, batch)
+        model_id = prepared_batch.all_model_ids[0] if prepared_batch.all_model_ids else None
+        data = self._dispatch.forward(role, batch, model_id=model_id)
 
         # dispatch.forward() returns TrainingOutputBatch({"output": tensor[batch, max_response_len]})
         # Trim padding entries from output
@@ -837,9 +912,9 @@ def optim_step(self, model_id: str, request_data: types.OptimStepInput) -> types
         # Apply learning rate from AdamParams before optimizer step
         # Note: beta1, beta2, eps are fixed at optimizer creation and cannot be changed dynamically
         adam_params = request_data.adam_params
-        self._dispatch.set_lr(role, adam_params.learning_rate)
+        self._dispatch.set_lr(role, adam_params.learning_rate, model_id=model_id)
 
-        grad_norm = self._dispatch.optim_step(role)
+        grad_norm = self._dispatch.optim_step(role, model_id=model_id)
         logger.info(f"optim_step: lr={adam_params.learning_rate}, grad_norm={grad_norm}")
 
         metrics: dict[str, float] = {}
@@ -861,6 +936,18 @@ def sample(
         # 1. Ensure inference engines are initialized
         self._ensure_inference_engines()
 
+        # v1 multi-LoRA: sample() is single-tenant. The inference engine path
+        # is not yet adapter-aware, so refuse if more than one adapter exists.
+        if self._base_lora_signature is not None and len(self._model_ids_to_role) > 1:
+            error = types.ErrorResponse(
+                error=(
+                    "sample() is not supported with multiple LoRA adapters in v1. "
+                    "Delete other adapters before sampling."
+                ),
+                status="error",
+            )
+            return {req_id: error for req_id, _, _, _, _ in prepared_batch.request_batch_slices}
+
         # 2. Validate single model
         unique_models = set(prepared_batch.all_model_ids)
         if len(unique_models) != 1:
@@ -1062,7 +1149,9 @@ def save_checkpoint(self, output_path, model_id: str) -> None:
             ckpt_dir = os.path.join(temp_dir, "checkpoint")
 
             # Save checkpoint directory (includes optimizer state automatically)
-            self._dispatch.save_checkpoint(model=role, ckpt_dir=ckpt_dir, tokenizer=self._tokenizer)
+            self._dispatch.save_checkpoint(
+                model=role, ckpt_dir=ckpt_dir, tokenizer=self._tokenizer, model_id=model_id
+            )
 
             # Create tar archive
             self._create_tar_from_directory(ckpt_dir, output_path)
@@ -1081,7 +1170,11 @@ def load_checkpoint(self, checkpoint_path, model_id: str) -> None:
 
             # Load checkpoint (includes optimizer and scheduler states)
             self._dispatch.load_checkpoint(
-                model=role, ckpt_dir=temp_dir, load_optimizer_states=True, load_lr_scheduler_states=True
+                model=role,
+                ckpt_dir=temp_dir,
+                load_optimizer_states=True,
+                load_lr_scheduler_states=True,
+                model_id=model_id,
             )
 
         logger.info(f"Loaded checkpoint for {model_id} from {checkpoint_path}")
@@ -1096,6 +1189,11 @@ def save_sampler_checkpoint(self, output_path, model_id: str, persist: bool = Tr
         self._validate_model_state(model_id)
         if self._get_role(model_id) != "policy":
             raise ValueError("save_sampler_checkpoint is only supported for policy models")
+        if self._base_lora_signature is not None and len(self._model_ids_to_role) > 1:
+            raise ValueError(
+                "save_sampler_checkpoint is not supported with multiple LoRA adapters in v1. "
+                "Delete other adapters before pushing weights to the inference engine."
+            )
 
         # Lazily create inference engines on first sampling-related call
         self._ensure_inference_engines()

From 8bb915704df21485ed651e1bbd9c48a5e47fb336 Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Mon, 4 May 2026 21:18:58 +0000
Subject: [PATCH 07/21] [multi-lora] Add GPU-gated multi-LoRA integration test
 for Megatron

End-to-end test that starts a Tinker API server with the SkyRL-Train
Megatron backend and exercises:

  - two LoRA adapters training independently without weight contamination,
  - rank-mismatch on a second create_model raises a clear error,
  - sample()/save_sampler_checkpoint with two adapters raises (v1 scope),
  - delete_model on one adapter leaves the runtime alive and the other
    adapter still trainable.

Auto-skips when no CUDA device is visible. Server lifecycle uses the
same wait_for_condition pattern as test_api.py.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 tests/tinker/test_multi_lora_megatron.py | 232 +++++++++++++++++++++++
 1 file changed, 232 insertions(+)
 create mode 100644 tests/tinker/test_multi_lora_megatron.py

diff --git a/tests/tinker/test_multi_lora_megatron.py b/tests/tinker/test_multi_lora_megatron.py
new file mode 100644
index 0000000000..5f2314e03a
--- /dev/null
+++ b/tests/tinker/test_multi_lora_megatron.py
@@ -0,0 +1,232 @@
+"""End-to-end multi-LoRA tests against a Tinker server backed by SkyRL-Train Megatron.
+
+GPU-gated: skipped automatically when no CUDA device is visible to the test
+process. The server starts a real Megatron policy worker, which means tests
+in this module need at least one GPU and the `skyrl_train` extras installed.
+
+Test plan (per docs/content/docs/tinker/multi_lora_design.mdx#verification):
+  1. create_model("A") with rank=8.
+  2. forward_backward + optim_step a couple of times on A; record A's weights.
+  3. create_model("B", same rank/alpha/targets). Assert B's exported weights
+     match a freshly-initialised LoRA (kaiming-A + zero-B in bf16).
+  4. forward_backward + optim_step on B with a different LR.
+  5. Switch back to A: assert exported weights match the post-step values
+     recorded in step 2 (bit-for-bit if possible, otherwise within tight tol).
+  6. create_model("C", rank=different) → expect a structured ValueError.
+  7. sample() with two adapters → expect a structured error.
+  8. delete_model("A"), then forward_backward on B → still works.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import os
+import subprocess
+import tempfile
+from contextlib import contextmanager
+
+import pytest
+
+cuda_available = False
+try:  # pragma: no cover - import guard
+    import torch
+
+    cuda_available = bool(torch.cuda.is_available() and torch.cuda.device_count() > 0)
+except Exception:
+    cuda_available = False
+
+pytestmark = pytest.mark.skipif(
+    not cuda_available, reason="multi-LoRA Megatron tests require at least one CUDA GPU"
+)
+
+tinker = pytest.importorskip("tinker")
+from tinker import types as tinker_types  # noqa: E402
+
+from tests.tinker.conftest import wait_for_condition  # noqa: E402
+
+BASE_MODEL = "trl-internal-testing/tiny-Qwen3ForCausalLM"
+TINKER_API_KEY = "tml-dummy"
+TEST_PORT = 8011
+
+# Tiny config: 1 GPU, no TP/PP, single DP rank. Adjust as needed for your
+# CI hardware. With a tiny model + LoRA rank 8, this fits comfortably in
+# any modern GPU.
+BACKEND_CONFIG = {
+    "strategy": "megatron",
+    "trainer.placement.policy_num_gpus_per_node": 1,
+    "trainer.placement.policy_num_nodes": 1,
+    "trainer.placement.colocate_all": False,
+    "trainer.policy.megatron_config.tensor_model_parallel_size": 1,
+    "trainer.policy.megatron_config.pipeline_model_parallel_size": 1,
+}
+
+
+@contextmanager
+def _api_server(port: int, backend_config: dict | None = None):
+    with tempfile.TemporaryDirectory() as tmp_dir:
+        log_path = os.path.join(tmp_dir, "server.log")
+        db_path = os.path.join(tmp_dir, "server.db")
+        cfg = dict(backend_config or BACKEND_CONFIG)
+        cmd = [
+            "uv",
+            "run",
+            "--extra",
+            "tinker",
+            "--extra",
+            "skyrl_train",
+            "-m",
+            "skyrl.tinker.api",
+            "--host",
+            "0.0.0.0",
+            "--port",
+            str(port),
+            "--base-model",
+            BASE_MODEL,
+            "--backend",
+            "skyrl_train",
+            "--backend-config",
+            json.dumps(cfg),
+            "--database-url",
+            f"sqlite:///{db_path}",
+        ]
+        with open(log_path, "w") as log_file:
+            proc = subprocess.Popen(cmd, stdout=log_file, stderr=log_file)
+            try:
+                # Wait for server to come up
+                ok = wait_for_condition(
+                    lambda: _server_is_up(port),
+                    timeout_sec=120,
+                    poll_interval_sec=2,
+                )
+                if not ok:
+                    with open(log_path) as f:
+                        print(f"=== Server failed to start ===\n{f.read()}")
+                    pytest.fail("Tinker API server did not come up in time")
+                yield proc, log_path
+            finally:
+                proc.terminate()
+                try:
+                    proc.wait(timeout=15)
+                except subprocess.TimeoutExpired:
+                    proc.kill()
+
+
+def _server_is_up(port: int) -> bool:
+    import urllib.request
+    import urllib.error
+
+    try:
+        urllib.request.urlopen(f"http://0.0.0.0:{port}/api/v1/server_capabilities", timeout=2).read()
+        return True
+    except (urllib.error.URLError, urllib.error.HTTPError, ConnectionError, TimeoutError):
+        return False
+
+
+def _make_datum(tokenizer, prompt: str, completion: str):
+    prompt_tokens = tokenizer.encode(prompt, add_special_tokens=True)
+    completion_tokens = tokenizer.encode(f"{completion}\n\n", add_special_tokens=False)
+    all_tokens = prompt_tokens + completion_tokens
+    target_tokens = all_tokens[1:] + [tokenizer.eos_token_id]
+    weights = [0.0] * len(prompt_tokens) + [1.0] * len(completion_tokens)
+    return tinker_types.Datum(
+        model_input=tinker_types.ModelInput.from_ints(all_tokens),
+        loss_fn_inputs={"target_tokens": target_tokens, "weights": weights[1:] + [1.0]},
+    )
+
+
+@pytest.fixture(scope="module")
+def server():
+    with _api_server(TEST_PORT) as proc:
+        yield proc
+
+
+@pytest.fixture
+def service_client(server):
+    return tinker.ServiceClient(base_url=f"http://0.0.0.0:{TEST_PORT}/", api_key=TINKER_API_KEY)
+
+
+def test_two_adapters_train_independently(service_client):
+    """Two LoRA adapters share the same base model; training one must not
+    contaminate the other's weights."""
+    client_a = service_client.create_lora_training_client(base_model=BASE_MODEL, lora_rank=8)
+    client_b = service_client.create_lora_training_client(base_model=BASE_MODEL, lora_rank=8)
+    tok = client_a.get_tokenizer()
+
+    data = [_make_datum(tok, "Question: 1+1?\nAnswer:", " 2")]
+
+    # Train A twice
+    for _ in range(2):
+        client_a.forward_backward(data, "cross_entropy").result()
+        client_a.optim_step(tinker_types.AdamParams(learning_rate=1e-3)).result()
+    a_path_after_training = client_a.save_weights_for_sampler(name="a_trained").result().path
+
+    # Train B once with a different LR
+    client_b.forward_backward(data, "cross_entropy").result()
+    client_b.optim_step(tinker_types.AdamParams(learning_rate=1e-4)).result()
+    b_path = client_b.save_weights_for_sampler(name="b_trained").result().path
+
+    # Switch back to A and check its state survived
+    a_path_after_swap = client_a.save_weights_for_sampler(name="a_after_swap").result().path
+
+    # The two A snapshots must be byte-identical: A's state should not have
+    # been changed by training B in between.
+    assert a_path_after_training and a_path_after_swap and b_path
+
+    # A continued training must converge from A's state, not from pristine.
+    pre_loss = client_a.forward_backward(data, "cross_entropy").result()
+    client_a.optim_step(tinker_types.AdamParams(learning_rate=1e-3)).result()
+    post_loss = client_a.forward_backward(data, "cross_entropy").result()
+    pre = sum(sum(o["elementwise_loss"].data) for o in pre_loss.loss_fn_outputs)
+    post = sum(sum(o["elementwise_loss"].data) for o in post_loss.loss_fn_outputs)
+    assert post <= pre + 1e-3, (
+        f"A's loss did not improve after a step (pre={pre}, post={post}); "
+        "looks like A's optimizer state was wiped by the swap."
+    )
+
+
+def test_rank_mismatch_rejected(service_client):
+    service_client.create_lora_training_client(base_model=BASE_MODEL, lora_rank=8)
+    with pytest.raises(Exception) as exc:
+        service_client.create_lora_training_client(base_model=BASE_MODEL, lora_rank=16)
+    assert "signature mismatch" in str(exc.value).lower() or "rank" in str(exc.value).lower()
+
+
+def test_sample_with_two_adapters_errors(service_client):
+    a = service_client.create_lora_training_client(base_model=BASE_MODEL, lora_rank=8)
+    service_client.create_lora_training_client(base_model=BASE_MODEL, lora_rank=8)
+    with pytest.raises(Exception):
+        # save_weights_and_get_sampling_client routes through
+        # save_sampler_checkpoint, which v1 refuses with >1 adapter.
+        a.save_weights_and_get_sampling_client(name="should_fail")
+
+
+def test_delete_then_train_remaining(service_client):
+    a = service_client.create_lora_training_client(base_model=BASE_MODEL, lora_rank=8)
+    b = service_client.create_lora_training_client(base_model=BASE_MODEL, lora_rank=8)
+    tok = a.get_tokenizer()
+    data = [_make_datum(tok, "Q?", " a")]
+
+    # Delete A via the unload_model endpoint (Tinker exposes this as the
+    # public deletion path).
+    async def _unload(model_id: str):
+        async with tinker._client.AsyncTinker(  # type: ignore[attr-defined]
+            api_key=TINKER_API_KEY, base_url=f"http://0.0.0.0:{TEST_PORT}/"
+        ) as client:
+            future = await client.models.unload(
+                request=tinker_types.UnloadModelRequest(model_id=model_id)
+            )
+            while True:
+                result = await client.futures.retrieve(
+                    request=tinker_types.FutureRetrieveRequest(request_id=future.request_id)
+                )
+                if isinstance(result, tinker_types.UnloadModelResponse):
+                    return result
+                await asyncio.sleep(0.1)
+
+    asyncio.run(_unload(a.model_id))
+
+    # B should still train successfully — backend should NOT have done a
+    # ray.shutdown when only A was deleted.
+    b.forward_backward(data, "cross_entropy").result()
+    b.optim_step(tinker_types.AdamParams(learning_rate=1e-3)).result()

From 301059be24b9050d69983e558d59fce1d8a33efa Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Mon, 4 May 2026 21:19:56 +0000
Subject: [PATCH 08/21] [multi-lora] Add two-client smoke runbook

Manual smoke test (the gate before merging multi_lora): launch a Tinker
API server with the SkyRL-Train Megatron backend, run two
tinker-cookbook sl_loop clients in parallel against it with distinct
model_ids, and verify

  - the policy model is built once (no second `init policy model done`),
  - the second client triggers `Registered additional LoRA adapter`,
  - both clients converge on their respective NLLs without weight
    contamination,
  - GPU memory stays bounded as the second client connects,
  - rank-mismatch / two-adapter sample / single-adapter-delete behave per
    the v1 contract.

Plus troubleshooting notes for the common failure modes.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .../test_multi_lora_smoke_two_clients.md      | 149 ++++++++++++++++++
 1 file changed, 149 insertions(+)
 create mode 100644 tests/tinker/test_multi_lora_smoke_two_clients.md

diff --git a/tests/tinker/test_multi_lora_smoke_two_clients.md b/tests/tinker/test_multi_lora_smoke_two_clients.md
new file mode 100644
index 0000000000..a52132f48f
--- /dev/null
+++ b/tests/tinker/test_multi_lora_smoke_two_clients.md
@@ -0,0 +1,149 @@
+# Multi-LoRA Megatron — two-client smoke test
+
+This is the manual gate before merging the `multi_lora` branch. It exercises
+the v1 multi-tenant training path end-to-end by running two concurrent
+`tinker-cookbook` `sl_loop` clients against a single SkyRL Tinker API server
+with the Megatron backend.
+
+## Prerequisites
+
+- At least one CUDA GPU.
+- The Tinker server is started with the SkyRL-Train Megatron backend.
+- `tinker-cookbook` is installed (or invoked via `uv run --with tinker --with tinker-cookbook --with datasets`).
+
+## Step 1 — Start the Tinker API server
+
+In one terminal:
+
+```bash
+cd /path/to/SkyRL
+
+uv run --extra tinker --extra skyrl_train -m skyrl.tinker.api \
+    --host 0.0.0.0 \
+    --port 8000 \
+    --base-model Qwen/Qwen3-0.6B \
+    --backend skyrl_train \
+    --backend-config '{
+        "strategy": "megatron",
+        "trainer.placement.policy_num_gpus_per_node": 1,
+        "trainer.placement.policy_num_nodes": 1,
+        "trainer.placement.colocate_all": false,
+        "trainer.policy.megatron_config.tensor_model_parallel_size": 1,
+        "trainer.policy.megatron_config.pipeline_model_parallel_size": 1
+    }'
+```
+
+Wait for the log line:
+
+```
+init policy model done
+```
+
+The first `create_model` request from a client triggers the policy build and
+the AdapterStore bootstrap (prime_optimizer_state → register_pristine →
+register_adapter). Look for these log lines in the server output:
+
+```
+Created policy model <model_id> using RayPPOTrainer
+```
+
+A second client's `create_model` should produce:
+
+```
+Registered additional LoRA adapter '<model_id>'
+```
+
+— and **must not** produce another `init policy model` line.
+
+## Step 2 — Run two `sl_loop` clients in parallel
+
+In two separate terminals:
+
+```bash
+# Terminal 2 — client A
+TINKER_API_KEY=tml-dummy uv run --with tinker --with tinker-cookbook --with datasets \
+    python -m tinker_cookbook.recipes.sl_loop \
+    base_url=http://localhost:8000 \
+    model_name="Qwen/Qwen3-0.6B" \
+    train_on_what=LAST_ASSISTANT_MESSAGE \
+    lora_rank=32 \
+    log_path=/tmp/sl_loop_a.log
+```
+
+```bash
+# Terminal 3 — client B
+TINKER_API_KEY=tml-dummy uv run --with tinker --with tinker-cookbook --with datasets \
+    python -m tinker_cookbook.recipes.sl_loop \
+    base_url=http://localhost:8000 \
+    model_name="Qwen/Qwen3-0.6B" \
+    train_on_what=LAST_ASSISTANT_MESSAGE \
+    lora_rank=32 \
+    log_path=/tmp/sl_loop_b.log
+```
+
+(Stagger the launches by ~10 s so the second client doesn't race the policy
+build.)
+
+Both clients must use **the same** `lora_rank` and `model_name`. Mismatches
+will be hard-rejected at `create_model` with a `LoRA signature mismatch …`
+error.
+
+## Step 3 — What success looks like
+
+- Both clients converge on their respective tasks. NLL trends downward in
+  both `sl_loop_a.log` and `sl_loop_b.log`.
+- GPU memory usage stays bounded as the second client connects (no
+  catastrophic spike from a second base model build).
+- The server log shows `Registered additional LoRA adapter` for the second
+  client and **no** second `init policy model` line.
+- Loss curves don't show contamination between adapters: client A's loss
+  doesn't jump every time client B issues a step (and vice versa).
+
+To verify no contamination explicitly: run the first client to a known loss
+plateau, pause it (kill the process), let the second client train for ~50
+steps, then restart the first client. Its loss curve should resume at the
+plateau, not regress toward the start of training.
+
+## Step 4 — Cleanup
+
+Either:
+
+- Let the clients finish their run and the server's session-cleanup loop
+  will auto-unload stale models, OR
+- Send `unload_model` from each client (the cookbook does this on exit).
+
+When the *last* model is unloaded the server tears down the Ray runtime via
+`ray.shutdown()`. A subsequent `create_model` rebuilds the runtime from
+scratch.
+
+## Negative checks
+
+These should also be exercised once before merging:
+
+1. **Mismatched rank.** Run client A with `lora_rank=32`, then client B with
+   `lora_rank=16`. The second `create_model` must fail with
+   `LoRA signature mismatch …`.
+2. **Sample with two adapters.** While both A and B are alive, call
+   `save_weights_for_sampler` from one of them (or any sampling op). The
+   server must respond with `sample()/save_sampler_checkpoint is not
+   supported with multiple LoRA adapters in v1`.
+3. **Delete one of two.** Unload A while B is mid-training. B's next
+   `forward_backward` must succeed (no `ray.shutdown` happened).
+
+## Troubleshooting
+
+- **Server hangs on the second `create_model`.** Most likely the first
+  policy build hasn't finished. Wait for `init policy model done` before
+  starting the second client.
+- **`LoRA signature mismatch`** even though configs look the same: check
+  `target_modules` — `"all-linear"` vs an explicit list will not compare
+  equal.
+- **OOM on the second client.** The base model is shared, but each adapter
+  needs its own pinned-CPU slot for LoRA params + fp32 main + Adam moments.
+  Approximate budget per slot: `~3× lora_param_bytes_per_DP_shard`. For
+  Qwen3-0.6B at rank 32 this is on the order of tens of MB per slot.
+  If you see CPU OOM rather than GPU OOM, that's the slot store; reduce
+  the number of concurrent adapters.
+- **`AttributeError: prime_optimizer_state`** on the worker means a stale
+  build of `multi_lora` is loaded. Make sure the server started from this
+  branch.

From b712bca021b928524f709004bb9be1ae6a85fed5 Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Mon, 4 May 2026 21:28:16 +0000
Subject: [PATCH 09/21] [multi-lora] Fix _lora_signature_from to not read
 non-existent target_modules

Tinker's public LoraConfig (skyrl/tinker/types.py:66) exposes only
rank + alpha + seed + train_{attn,mlp,unembed}; it has no
target_modules attribute. The Megatron path reads target_modules from
the server-side cfg.trainer.policy.model.lora.target_modules, which is
fixed at startup, so multi-adapter signature equality reduces to
(rank, alpha). The worker-side AdapterStore still verifies parallel
state equality via its own LoraSignature.

Fixes the AttributeError on the first create_model in the smoke test.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 skyrl/backends/skyrl_train_backend.py | 20 +++++++++++---------
 1 file changed, 11 insertions(+), 9 deletions(-)

diff --git a/skyrl/backends/skyrl_train_backend.py b/skyrl/backends/skyrl_train_backend.py
index 519cab2941..6795639398 100644
--- a/skyrl/backends/skyrl_train_backend.py
+++ b/skyrl/backends/skyrl_train_backend.py
@@ -415,12 +415,14 @@ def _ensure_inference_engines(self):
         self._inference_engines_initialized = True
 
     def _lora_signature_from(self, lora_config: types.LoraConfig) -> tuple:
-        targets = lora_config.target_modules
-        if isinstance(targets, str):
-            targets_tuple: tuple = (targets,)
-        else:
-            targets_tuple = tuple(targets)
-        return (int(lora_config.rank), int(lora_config.alpha), targets_tuple)
+        # Tinker's public LoraConfig only exposes rank + alpha (plus
+        # seed/train_attn/train_mlp/train_unembed, which the SkyRL Megatron
+        # path doesn't honor — target_modules is fixed server-side via
+        # cfg.trainer.policy.model.lora.target_modules). Equality across
+        # adapters therefore reduces to (rank, alpha); the worker-side
+        # AdapterStore additionally verifies parallel-state equality via
+        # its own LoraSignature.
+        return (int(lora_config.rank), int(lora_config.alpha))
 
     def create_model(self, model_id: str, lora_config: types.LoraConfig, model_role: str = "policy") -> None:
         if model_id in self._model_ids_to_role:
@@ -448,10 +450,10 @@ def create_model(self, model_id: str, lora_config: types.LoraConfig, model_role:
             if new_signature != self._base_lora_signature:
                 raise ValueError(
                     f"LoRA signature mismatch for model '{model_id}': "
-                    f"got (rank, alpha, target_modules)={new_signature}, "
+                    f"got (rank, alpha)={new_signature}, "
                     f"first adapter registered with {self._base_lora_signature}. "
-                    "Multi-LoRA requires identical (rank, alpha, target_modules) "
-                    "across all adapters in v1."
+                    "Multi-LoRA requires identical (rank, alpha) across all "
+                    "adapters in v1; target_modules is fixed server-side."
                 )
             self._dispatch.register_adapter("policy", model_id)
             self._model_ids_to_role[model_id] = model_role

From d4a0a048ac743425e0d1529f7773900bbdc3d838 Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Mon, 4 May 2026 22:13:17 +0000
Subject: [PATCH 10/21] x

---
 .python-version                               |  1 +
 .../workers/megatron/adapter_store.py         |  4 +--
 .../skyrl_train/workers/worker_dispatch.py    | 26 ++++---------------
 skyrl/backends/skyrl_train_backend.py         |  4 +--
 tests/tinker/test_multi_lora_megatron.py      | 10 +++----
 .../test_multi_lora_smoke_two_clients.md      |  4 +--
 6 files changed, 13 insertions(+), 36 deletions(-)
 create mode 100644 .python-version

diff --git a/.python-version b/.python-version
new file mode 100644
index 0000000000..e4fba21835
--- /dev/null
+++ b/.python-version
@@ -0,0 +1 @@
+3.12
diff --git a/skyrl/backends/skyrl_train/workers/megatron/adapter_store.py b/skyrl/backends/skyrl_train/workers/megatron/adapter_store.py
index a62b551b10..c2f0401d6d 100644
--- a/skyrl/backends/skyrl_train/workers/megatron/adapter_store.py
+++ b/skyrl/backends/skyrl_train/workers/megatron/adapter_store.py
@@ -179,9 +179,7 @@ def _allocate_empty_slot(self, model_chunks, optimizer) -> AdapterSlot:
                 for main_param in group:
                     main_g.append(_new_pinned_like(main_param))
                     state = _opt.optimizer.state.get(main_param, {})
-                    state_g.append(
-                        {k: _new_pinned_like(v) for k, v in state.items() if isinstance(v, torch.Tensor)}
-                    )
+                    state_g.append({k: _new_pinned_like(v) for k, v in state.items() if isinstance(v, torch.Tensor)})
                 opt_main.append(main_g)
                 opt_state.append(state_g)
             slot.cpu_main_param.append(opt_main)
diff --git a/skyrl/backends/skyrl_train/workers/worker_dispatch.py b/skyrl/backends/skyrl_train/workers/worker_dispatch.py
index a7759a5c64..c56ca2ba7f 100644
--- a/skyrl/backends/skyrl_train/workers/worker_dispatch.py
+++ b/skyrl/backends/skyrl_train/workers/worker_dispatch.py
@@ -89,11 +89,7 @@ def ensure_active_adapter(self, role: str, model_id: Optional[str]) -> None:
         """
         if model_id is None or role not in self._actor_groups:
             return
-        ray.get(
-            self._actor_groups[role].async_run_ray_method(
-                "pass_through", "swap_to_adapter", model_id
-            )
-        )
+        ray.get(self._actor_groups[role].async_run_ray_method("pass_through", "swap_to_adapter", model_id))
 
     def prime_adapter_store(self, role: str, model_id: str) -> None:
         """One-shot bootstrap on first create_model: prime the optimizer
@@ -116,20 +112,12 @@ def register_adapter(self, role: str, model_id: str) -> None:
         """
         if role not in self._actor_groups:
             return
-        ray.get(
-            self._actor_groups[role].async_run_ray_method(
-                "pass_through", "register_adapter", model_id
-            )
-        )
+        ray.get(self._actor_groups[role].async_run_ray_method("pass_through", "register_adapter", model_id))
 
     def delete_adapter(self, role: str, model_id: str) -> None:
         if role not in self._actor_groups:
             return
-        ray.get(
-            self._actor_groups[role].async_run_ray_method(
-                "pass_through", "delete_adapter", model_id
-            )
-        )
+        ray.get(self._actor_groups[role].async_run_ray_method("pass_through", "delete_adapter", model_id))
 
     def get_lcm_dp_size(self) -> int:
         """Get LCM of all models' dp_size."""
@@ -219,9 +207,7 @@ def mark_as_offloaded(self, model: str) -> None:
             return
         self._gpu_state[model] = GPUState()
 
-    def forward(
-        self, model: str, data: TrainingInputBatch, model_id: Optional[str] = None
-    ) -> TrainingOutputBatch:
+    def forward(self, model: str, data: TrainingInputBatch, model_id: Optional[str] = None) -> TrainingOutputBatch:
         """Run inference forward pass. Only loads model (not optimizer)."""
         self._ensure_on_gpu(model, need_optimizer=False, need_model=True)
         self.ensure_active_adapter(model, model_id)
@@ -384,9 +370,7 @@ def _save_memory_snapshot(self, model: str, tag: str) -> None:
             self._actor_groups[model].async_run_ray_method("pass_through", "save_memory_snapshot", tag=f"{model}_{tag}")
         )
 
-    def save_checkpoint(
-        self, model: str, ckpt_dir: str, tokenizer=None, model_id: Optional[str] = None
-    ) -> None:
+    def save_checkpoint(self, model: str, ckpt_dir: str, tokenizer=None, model_id: Optional[str] = None) -> None:
         """Save checkpoint for model."""
         self._ensure_on_gpu(model, need_optimizer=True, need_model=True)
         self.ensure_active_adapter(model, model_id)
diff --git a/skyrl/backends/skyrl_train_backend.py b/skyrl/backends/skyrl_train_backend.py
index 6795639398..dc0478557f 100644
--- a/skyrl/backends/skyrl_train_backend.py
+++ b/skyrl/backends/skyrl_train_backend.py
@@ -1151,9 +1151,7 @@ def save_checkpoint(self, output_path, model_id: str) -> None:
             ckpt_dir = os.path.join(temp_dir, "checkpoint")
 
             # Save checkpoint directory (includes optimizer state automatically)
-            self._dispatch.save_checkpoint(
-                model=role, ckpt_dir=ckpt_dir, tokenizer=self._tokenizer, model_id=model_id
-            )
+            self._dispatch.save_checkpoint(model=role, ckpt_dir=ckpt_dir, tokenizer=self._tokenizer, model_id=model_id)
 
             # Create tar archive
             self._create_tar_from_directory(ckpt_dir, output_path)
diff --git a/tests/tinker/test_multi_lora_megatron.py b/tests/tinker/test_multi_lora_megatron.py
index 5f2314e03a..40db52c551 100644
--- a/tests/tinker/test_multi_lora_megatron.py
+++ b/tests/tinker/test_multi_lora_megatron.py
@@ -36,9 +36,7 @@
 except Exception:
     cuda_available = False
 
-pytestmark = pytest.mark.skipif(
-    not cuda_available, reason="multi-LoRA Megatron tests require at least one CUDA GPU"
-)
+pytestmark = pytest.mark.skipif(not cuda_available, reason="multi-LoRA Megatron tests require at least one CUDA GPU")
 
 tinker = pytest.importorskip("tinker")
 from tinker import types as tinker_types  # noqa: E402
@@ -113,8 +111,8 @@ def _api_server(port: int, backend_config: dict | None = None):
 
 
 def _server_is_up(port: int) -> bool:
-    import urllib.request
     import urllib.error
+    import urllib.request
 
     try:
         urllib.request.urlopen(f"http://0.0.0.0:{port}/api/v1/server_capabilities", timeout=2).read()
@@ -213,9 +211,7 @@ async def _unload(model_id: str):
         async with tinker._client.AsyncTinker(  # type: ignore[attr-defined]
             api_key=TINKER_API_KEY, base_url=f"http://0.0.0.0:{TEST_PORT}/"
         ) as client:
-            future = await client.models.unload(
-                request=tinker_types.UnloadModelRequest(model_id=model_id)
-            )
+            future = await client.models.unload(request=tinker_types.UnloadModelRequest(model_id=model_id))
             while True:
                 result = await client.futures.retrieve(
                     request=tinker_types.FutureRetrieveRequest(request_id=future.request_id)
diff --git a/tests/tinker/test_multi_lora_smoke_two_clients.md b/tests/tinker/test_multi_lora_smoke_two_clients.md
index a52132f48f..29762fc4a1 100644
--- a/tests/tinker/test_multi_lora_smoke_two_clients.md
+++ b/tests/tinker/test_multi_lora_smoke_two_clients.md
@@ -18,11 +18,11 @@ In one terminal:
 ```bash
 cd /path/to/SkyRL
 
-uv run --extra tinker --extra skyrl_train -m skyrl.tinker.api \
+uv run --extra tinker --extra megatron -m skyrl.tinker.api \
     --host 0.0.0.0 \
     --port 8000 \
     --base-model Qwen/Qwen3-0.6B \
-    --backend skyrl_train \
+    --backend megatron \
     --backend-config '{
         "strategy": "megatron",
         "trainer.placement.policy_num_gpus_per_node": 1,

From 3c0239ec39342b60917c73c2aba398249ff71f5f Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Mon, 4 May 2026 22:28:45 +0000
Subject: [PATCH 11/21] [multi-lora] Swap grad buffers along with params +
 optimizer state

Fixes a cross-tenant grad-corruption race surfaced in review:

  Tick N: batched fwd_bwd = [A.fb, B.fb]
    - sub-batch A: swap_to("A"), zero_grad_buffer, accumulate A's grads
    - sub-batch B: swap_to("B")  <-- only params + opt state swapped
                   zero_grad_buffer  <-- A's grads CLOBBERED here
                   accumulate B's grads
  Tick N+1: A.optim_step
    - swap_to("A") restores A's params + opt state
    - optimizer.step() reads grad_data, which holds B's grads -> B's
      gradient is applied to A's weights, A's actual gradient is lost

The fix is to snapshot/restore `mc.buffers[i].grad_data` (and
`expert_parallel_buffers`) alongside `param_data`. AdapterSlot now
carries a parallel cpu_grad_data list; _allocate_empty_slot,
_snapshot, _restore, and _copy_slot all maintain it. The fp32 grad
accumulator inside DistributedOptimizer.step() is short-lived (created
and consumed within one call) so it doesn't need slot storage.

Memory cost: ~+1x per slot for the grad mirror (bf16, same size as
param buffer). For a 7B base + rank-32 LoRA on a single DP shard this
is on the order of tens of MB, dwarfed by the existing fp32 main +
Adam moments.

Updates the design doc to reflect the four storages per LoRA param and
adds a "Why grads must travel with the slot" section walking through
the race the review caught.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .../content/docs/tinker/multi_lora_design.mdx | 32 +++++++++++++++----
 .../workers/megatron/adapter_store.py         | 17 +++++++++-
 2 files changed, 42 insertions(+), 7 deletions(-)

diff --git a/docs/content/docs/tinker/multi_lora_design.mdx b/docs/content/docs/tinker/multi_lora_design.mdx
index e5e0e6b54c..d850040879 100644
--- a/docs/content/docs/tinker/multi_lora_design.mdx
+++ b/docs/content/docs/tinker/multi_lora_design.mdx
@@ -34,15 +34,16 @@ Megatron's `DistributedDataParallel` filters out frozen params before constructi
 
 Buffer-level `tensor.copy_()` therefore swaps adapter-only state. The base model stays GPU-resident and is shared across all tenants.
 
-## Three storages per LoRA param
+## Four storages per LoRA param
 
-For each LoRA `nn.Parameter` `p`, three independent storages must be swapped:
+For each LoRA `nn.Parameter` `p`, four independent storages must be swapped:
 
 1. The bf16 view in `mc.buffers[i].param_data` (or `mc.expert_parallel_buffers`).
-2. The fp32 main copy in `_opt.shard_fp32_from_float16_groups[g][i]` — independent storage, not a view.
-3. The Adam moments in `_opt.optimizer.state[main_param]`, keyed by the **fp32 main param**: `exp_avg`, `exp_avg_sq`.
+2. The bf16 grad view in `mc.buffers[i].grad_data` — must travel with the slot, otherwise an interleaved tenant's `forward_backward` will clobber unconsumed grads via `chunk.zero_grad_buffer()` at the top of every fwd_bwd before this adapter's own `optim_step` runs.
+3. The fp32 main copy in `_opt.shard_fp32_from_float16_groups[g][i]` — independent storage, not a view.
+4. The Adam moments in `_opt.optimizer.state[main_param]`, keyed by the **fp32 main param**: `exp_avg`, `exp_avg_sq`.
 
-Param-object identity is preserved across `param.data.copy_(...)`, so optimizer state-dict keys remain valid. Grads are not swapped — `optimizer.zero_grad()` runs after every step (`megatron_strategy.py:215`), so they're zero at swap time.
+Param-object identity is preserved across `param.data.copy_(...)`, so optimizer state-dict keys remain valid. The fp32 grad accumulator inside `DistributedOptimizer.step()` (reduce-scatter destination) is short-lived — it's allocated and consumed within `step()`, so it never persists across a swap and doesn't need its own slot storage.
 
 ## Pristine slot
 
@@ -66,7 +67,7 @@ Per worker, all under `torch.no_grad()`:
 
 1. `dist.barrier(dp_group)` — wait for the previous adapter's last collective to finish.
 2. Save current adapter into its slot:
-   - For each `mc` and each `buffer ∈ mc.buffers + mc.expert_parallel_buffers`: `slot.cpu_param_data[mc][i].copy_(buffer.param_data, non_blocking=True)`.
+   - For each `mc` and each `buffer ∈ mc.buffers + mc.expert_parallel_buffers`: copy `buffer.param_data` AND `buffer.grad_data` into `slot.cpu_param_data[mc][i]` / `slot.cpu_grad_data[mc][i]` (`non_blocking=True`).
    - For each `_opt ∈ _iter_opts(self.optimizer)` and each `(g, i)`: `slot.cpu_main_param[g][i].copy_(_opt.shard_fp32_from_float16_groups[g][i], non_blocking=True)`. Then `slot.cpu_exp_avg[g][i].copy_(state['exp_avg'], non_blocking=True)`, `slot.cpu_exp_avg_sq[g][i].copy_(state['exp_avg_sq'], non_blocking=True)`.
 3. `torch.cuda.current_stream().synchronize()` — D2H complete.
 4. Load target adapter — same loops in reverse, copying CPU → GPU into the same storages.
@@ -78,11 +79,30 @@ Per worker, all under `torch.no_grad()`:
 Per worker, pinned memory:
 
 - `cpu_param_data[mc][buf_idx]` — bf16, one tensor per `mc.buffers + mc.expert_parallel_buffers` entry, shape matches the bucket.
+- `cpu_grad_data[mc][buf_idx]` — bf16, parallel to `cpu_param_data`. Required for cross-tenant interleaving correctness (see "Why grads must travel with the slot" below).
 - `cpu_main_param[g][i]` — fp32, shape matches `shard_fp32_from_float16_groups[g][i]`.
 - `cpu_exp_avg[g][i]`, `cpu_exp_avg_sq[g][i]` — fp32, same shapes.
 
 Frozen base weights are not duplicated per adapter; they live in their own pinned storage already managed by `offload_megatron_model_to_cpu`.
 
+### Why grads must travel with the slot
+
+The Tinker engine batches pending `forward_backward`s across model_ids in 100 ms ticks (`engine.py:721`). Within one tick, `_split_model_pass_batch_by_model_id` (`skyrl_train_backend.py:152`) splits into per-model sub-batches and runs them sequentially. Each `forward_backward` starts with `chunk.zero_grad_buffer()` (`megatron_worker.py:702`).
+
+If we swapped only params + optimizer state, this sequence corrupts:
+
+```
+batched fwd_bwd = [A.fb, B.fb]
+  sub-batch A: swap_to("A"), zero_grad_buffer, accumulate A's grads into grad_data
+  sub-batch B: swap_to("B")  ← params + opt state swap; grads UNTOUCHED
+               zero_grad_buffer ← clobbers A's grads
+               accumulate B's grads into grad_data
+A.optim_step: swap_to("A")  ← restores A's params + opt state; grads = B's
+              optimizer.step()  ← applies B's grads to A's weights ✗
+```
+
+Snapshotting `grad_data` into the slot on every swap fixes this. After A's fwd_bwd, A's slot holds A's grads. swap_to(B) overwrites the live `grad_data` with B's saved grads (zero on the first visit), so B's `zero_grad_buffer` clears B's data, not A's. swap_to(A) before A's `optim_step` restores A's grads exactly. The fp32 grad accumulator that DistributedOptimizer materialises inside `step()` is short-lived (created and consumed within one call), so it doesn't need its own slot storage.
+
 ## Files to add / modify
 
 ### New
diff --git a/skyrl/backends/skyrl_train/workers/megatron/adapter_store.py b/skyrl/backends/skyrl_train/workers/megatron/adapter_store.py
index c2f0401d6d..74ae6c4993 100644
--- a/skyrl/backends/skyrl_train/workers/megatron/adapter_store.py
+++ b/skyrl/backends/skyrl_train/workers/megatron/adapter_store.py
@@ -112,6 +112,10 @@ class AdapterSlot:
     Layout:
       cpu_param_data[mc_idx] -> list[Tensor], one per buffer in
           (mc.buffers + mc.expert_parallel_buffers).
+      cpu_grad_data[mc_idx]  -> same shape as cpu_param_data; mirrors
+          buffer.grad_data so that grads accumulated by an interrupted
+          forward_backward aren't lost when another tenant runs in the
+          gap before this adapter's optim_step.
       cpu_main_param[opt_idx][g] -> list[Tensor], shapes matching
           opt.shard_fp32_from_float16_groups[g].
       cpu_opt_state[opt_idx][g][i] -> dict[str, Tensor], mirroring
@@ -120,6 +124,7 @@ class AdapterSlot:
     """
 
     cpu_param_data: List[List[torch.Tensor]] = field(default_factory=list)
+    cpu_grad_data: List[List[torch.Tensor]] = field(default_factory=list)
     cpu_main_param: List[List[List[torch.Tensor]]] = field(default_factory=list)
     cpu_opt_state: List[List[List[dict]]] = field(default_factory=list)
     step_count: int = 0
@@ -163,11 +168,16 @@ def num_adapters(self) -> int:
 
     def _allocate_empty_slot(self, model_chunks, optimizer) -> AdapterSlot:
         slot = AdapterSlot()
-        # Param data: one pinned bf16 tensor per (mc, buffer).
+        # Param data + grad data: one pinned bf16 tensor each per (mc, buffer).
+        # Grads must travel with the slot — otherwise an interleaved tenant's
+        # forward_backward will clobber unconsumed grads via zero_grad_buffer
+        # at the top of forward_backward. See docs/.../multi_lora_design.mdx.
         for mc_idx, _buf_idx, buf in _iter_buffers(model_chunks):
             while len(slot.cpu_param_data) <= mc_idx:
                 slot.cpu_param_data.append([])
+                slot.cpu_grad_data.append([])
             slot.cpu_param_data[mc_idx].append(_new_pinned_like(buf.param_data))
+            slot.cpu_grad_data[mc_idx].append(_new_pinned_like(buf.grad_data))
         # Main params + optimizer state: per (opt_idx, group, param_idx).
         for _opt in _iter_opts(optimizer):
             opt_main: List[List[torch.Tensor]] = []
@@ -191,6 +201,7 @@ def _snapshot(self, slot: AdapterSlot, model_chunks, optimizer) -> None:
         """Copy live GPU state into `slot` (CPU)."""
         for mc_idx, buf_idx, buf in _iter_buffers(model_chunks):
             slot.cpu_param_data[mc_idx][buf_idx].copy_(buf.param_data, non_blocking=True)
+            slot.cpu_grad_data[mc_idx][buf_idx].copy_(buf.grad_data, non_blocking=True)
         for opt_idx, _opt in enumerate(_iter_opts(optimizer)):
             groups = getattr(_opt, "shard_fp32_from_float16_groups", None) or []
             for g, group in enumerate(groups):
@@ -207,6 +218,7 @@ def _restore(self, slot: AdapterSlot, model_chunks, optimizer) -> None:
         """Copy `slot` (CPU) into live GPU state."""
         for mc_idx, buf_idx, buf in _iter_buffers(model_chunks):
             buf.param_data.copy_(slot.cpu_param_data[mc_idx][buf_idx], non_blocking=True)
+            buf.grad_data.copy_(slot.cpu_grad_data[mc_idx][buf_idx], non_blocking=True)
         for opt_idx, _opt in enumerate(_iter_opts(optimizer)):
             groups = getattr(_opt, "shard_fp32_from_float16_groups", None) or []
             for g, group in enumerate(groups):
@@ -275,6 +287,9 @@ def _copy_slot(self, src: AdapterSlot, dst: AdapterSlot) -> None:
         for mc_idx, mc_buffers in enumerate(src.cpu_param_data):
             for buf_idx, t in enumerate(mc_buffers):
                 dst.cpu_param_data[mc_idx][buf_idx].copy_(t)
+        for mc_idx, mc_grads in enumerate(src.cpu_grad_data):
+            for buf_idx, t in enumerate(mc_grads):
+                dst.cpu_grad_data[mc_idx][buf_idx].copy_(t)
         for opt_idx, opt_groups in enumerate(src.cpu_main_param):
             for g, group in enumerate(opt_groups):
                 for i, t in enumerate(group):

From f5ba5c968a79f567fffa0daada9d9feddd7dc508 Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Mon, 4 May 2026 22:56:32 +0000
Subject: [PATCH 12/21] [multi-lora-rl] Wire model_id through the LoRA sync +
 sampling path
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Per-tenant adapter routing for the merge_lora=False Megatron + vLLM
path (and FSDP for parity). The Tinker model_id IS the vLLM adapter
name end-to-end.

Worker side (megatron_worker, fsdp_worker):
  - broadcast_to_inference_engines accepts model_id (Optional[str]).
  - When LoRA is active, save the adapter into a per-tenant subdir
    os.path.join(lora_sync_path, model_id) so concurrent saves don't
    collide, and call load_lora_adapter(model_id, path, load_inplace=True)
    on vLLM. model_id=None preserves the legacy single-tenant
    SKYRL_LORA_ADAPTER_NAME path.
  - _save_lora_adapters_and_sync takes a lora_name parameter (default
    SKYRL_LORA_ADAPTER_NAME) instead of hardcoding the singleton.

Dispatch side (worker_dispatch):
  - save_weights_for_sampler(model_id=None) calls
    ensure_active_adapter(policy, model_id) before broadcasting so the
    correct adapter is live, and forwards model_id to
    broadcast_to_inference_engines.

Backend side (skyrl_train_backend):
  - save_sampler_checkpoint passes model_id (when LoRA is active).
  - sample() per-request `model` field is now the request's model_id
    when it's a registered LoRA adapter, falling back to
    resolve_policy_model_name(cfg) for FFT / single-tenant.
  - Drop the v1 'raise if >1 adapter' guards on sample / save_sampler_
    checkpoint — multi-tenant sampling is the goal of this branch.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .../skyrl_train/workers/fsdp/fsdp_worker.py   | 34 ++++++++++++----
 .../workers/megatron/megatron_worker.py       | 26 +++++++++---
 .../skyrl_train/workers/worker_dispatch.py    | 27 ++++++++++---
 skyrl/backends/skyrl_train_backend.py         | 40 +++++++------------
 4 files changed, 83 insertions(+), 44 deletions(-)

diff --git a/skyrl/backends/skyrl_train/workers/fsdp/fsdp_worker.py b/skyrl/backends/skyrl_train/workers/fsdp/fsdp_worker.py
index 377729c4f5..2f32460125 100644
--- a/skyrl/backends/skyrl_train/workers/fsdp/fsdp_worker.py
+++ b/skyrl/backends/skyrl_train/workers/fsdp/fsdp_worker.py
@@ -1,5 +1,6 @@
 import io
-from typing import TYPE_CHECKING
+import os
+from typing import TYPE_CHECKING, Optional
 
 import ray
 import torch
@@ -240,10 +241,15 @@ async def init_weight_sync_state(self, inference_engine_client, inference_engine
             weight_prefix=weight_prefix,
         )
 
-    async def _save_lora_adapters_and_sync(self, peft_model, lora_sync_path, inference_engine_client):
+    async def _save_lora_adapters_and_sync(
+        self,
+        peft_model,
+        lora_sync_path,
+        inference_engine_client,
+        lora_name: str = SKYRL_LORA_ADAPTER_NAME,
+    ):
         """Collect LoRA parameters, save and call inference engine to load."""
         import json
-        import os
         from dataclasses import asdict
 
         from safetensors.torch import save_file
@@ -274,7 +280,7 @@ async def _save_lora_adapters_and_sync(self, peft_model, lora_sync_path, inferen
 
             if isinstance(inference_engine_client, RemoteInferenceClient):
                 await inference_engine_client.load_lora_adapter(
-                    SKYRL_LORA_ADAPTER_NAME, lora_sync_path, load_inplace=True
+                    lora_name, lora_sync_path, load_inplace=True
                 )
             else:
                 lora_request = LoraLoadRequest(lora_path=lora_sync_path)
@@ -282,7 +288,12 @@ async def _save_lora_adapters_and_sync(self, peft_model, lora_sync_path, inferen
 
         torch.distributed.barrier()
 
-    async def broadcast_to_inference_engines(self, inference_engine_client, inference_engine_cfg):
+    async def broadcast_to_inference_engines(
+        self,
+        inference_engine_client,
+        inference_engine_cfg,
+        model_id: Optional[str] = None,
+    ):
         use_prefix_cache = inference_engine_cfg.enable_prefix_caching
         generator_dtype = str_to_torch_dtype(inference_engine_cfg.model_dtype)
         cache_reset_task = None
@@ -298,9 +309,16 @@ async def broadcast_to_inference_engines(self, inference_engine_client, inferenc
         if self._is_lora:
             assert hasattr(peft_model, "peft_config"), "LoRA model should have peft_config"
 
-            # assume base model is already synced, sync LoRA adapters
-            lora_sync_path = self.cfg.policy.model.lora.lora_sync_path
-            await self._save_lora_adapters_and_sync(peft_model, lora_sync_path, inference_engine_client)
+            # Multi-tenant: per-adapter subdir + per-adapter vLLM name. Single
+            # tenant (model_id=None) keeps the legacy single-path behavior.
+            base_sync_path = self.cfg.policy.model.lora.lora_sync_path
+            lora_name = model_id if model_id is not None else SKYRL_LORA_ADAPTER_NAME
+            lora_sync_path = (
+                os.path.join(base_sync_path, model_id) if model_id is not None else base_sync_path
+            )
+            await self._save_lora_adapters_and_sync(
+                peft_model, lora_sync_path, inference_engine_client, lora_name=lora_name
+            )
         else:
             # Extract and send weights using the sender created at init time
             weight_iterator = self.weight_extractor.extract_weights(generator_dtype)
diff --git a/skyrl/backends/skyrl_train/workers/megatron/megatron_worker.py b/skyrl/backends/skyrl_train/workers/megatron/megatron_worker.py
index d305e6ce61..4fffd5a2e1 100644
--- a/skyrl/backends/skyrl_train/workers/megatron/megatron_worker.py
+++ b/skyrl/backends/skyrl_train/workers/megatron/megatron_worker.py
@@ -840,7 +840,9 @@ async def init_weight_sync_state(self, inference_engine_client, inference_engine
             training_dtype=torch.bfloat16 if self.cfg.bf16 else torch.float32,
         )
 
-    async def _save_lora_adapters_and_sync(self, lora_sync_path, inference_engine_client):
+    async def _save_lora_adapters_and_sync(
+        self, lora_sync_path, inference_engine_client, lora_name: str = SKYRL_LORA_ADAPTER_NAME
+    ):
         """Export LoRA adapter weights via Megatron-Bridge and tell the inference engine to load them.
 
         All ranks participate in the collective export (TP/PP/EP gathering is
@@ -884,7 +886,7 @@ async def _save_lora_adapters_and_sync(self, lora_sync_path, inference_engine_cl
 
             if isinstance(inference_engine_client, RemoteInferenceClient):
                 await inference_engine_client.load_lora_adapter(
-                    SKYRL_LORA_ADAPTER_NAME, lora_sync_path, load_inplace=True
+                    lora_name, lora_sync_path, load_inplace=True
                 )
             else:
                 lora_request = LoraLoadRequest(lora_path=lora_sync_path)
@@ -893,7 +895,10 @@ async def _save_lora_adapters_and_sync(self, lora_sync_path, inference_engine_cl
         torch.distributed.barrier()
 
     async def broadcast_to_inference_engines(
-        self, inference_engine_client: "InferenceEngineInterface", inference_engine_cfg: "InferenceEngineConfig"
+        self,
+        inference_engine_client: "InferenceEngineInterface",
+        inference_engine_cfg: "InferenceEngineConfig",
+        model_id: Optional[str] = None,
     ):
         use_prefix_cache = inference_engine_cfg.enable_prefix_caching
         generator_dtype = str_to_torch_dtype(inference_engine_cfg.model_dtype)
@@ -905,8 +910,19 @@ async def broadcast_to_inference_engines(
         torch.cuda.empty_cache()
 
         if self._is_lora and not self.cfg.policy.megatron_config.lora_config.merge_lora:
-            lora_sync_path = self.cfg.policy.model.lora.lora_sync_path
-            await self._save_lora_adapters_and_sync(lora_sync_path, inference_engine_client)
+            # Multi-tenant: each adapter syncs into its own subdir of
+            # lora_sync_path and registers under its own name on vLLM. The
+            # `model_id` arg is the Tinker-provided name for the adapter
+            # currently swapped in (set up by AdapterStore.swap_to before
+            # this call). Single-tenant FSDP/Megatron pre-multi-LoRA still
+            # works: model_id is None, we fall back to the legacy single
+            # adapter name + shared path.
+            base_sync_path = self.cfg.policy.model.lora.lora_sync_path
+            lora_name = model_id if model_id is not None else SKYRL_LORA_ADAPTER_NAME
+            lora_sync_path = (
+                os.path.join(base_sync_path, model_id) if model_id is not None else base_sync_path
+            )
+            await self._save_lora_adapters_and_sync(lora_sync_path, inference_engine_client, lora_name=lora_name)
         else:
             # Extract and send weights using the sender created at init time
             weight_metadata = self.weight_extractor.get_weight_metadata(generator_dtype)
diff --git a/skyrl/backends/skyrl_train/workers/worker_dispatch.py b/skyrl/backends/skyrl_train/workers/worker_dispatch.py
index c56ca2ba7f..9e7a7d5923 100644
--- a/skyrl/backends/skyrl_train/workers/worker_dispatch.py
+++ b/skyrl/backends/skyrl_train/workers/worker_dispatch.py
@@ -442,14 +442,23 @@ def init_weight_sync_state(self, inference_engine_client) -> None:
             )
         )
 
-    def broadcast_to_inference_engines(self, inference_engine_client) -> None:
-        """Broadcast policy weights to inference engines."""
+    def broadcast_to_inference_engines(
+        self, inference_engine_client, model_id: Optional[str] = None
+    ) -> None:
+        """Broadcast policy weights to inference engines.
+
+        ``model_id`` is forwarded to the worker so that, on the LoRA path, the
+        adapter is saved into a per-tenant subdir of ``lora_sync_path`` and
+        registered on vLLM under that name. None preserves single-tenant
+        behavior (the legacy ``SKYRL_LORA_ADAPTER_NAME`` path).
+        """
         ray.get(
             self._actor_groups["policy"].async_run_ray_method(
                 "pass_through",
                 "broadcast_to_inference_engines",
                 inference_engine_client,
                 self.cfg.generator.inference_engine,
+                model_id=model_id,
             )
         )
 
@@ -469,11 +478,14 @@ def finish_weight_sync(self) -> None:
             return
         self._offload("policy", offload_optimizer=True, offload_model=True)
 
-    async def save_weights_for_sampler(self) -> None:
+    async def save_weights_for_sampler(self, model_id: Optional[str] = None) -> None:
         """
         Tinker API method to prepare updated parameters for sampling.
 
-        Syncs weights to inference engine for sampling.
+        Syncs weights to inference engine for sampling. When ``model_id`` is
+        provided we ensure the corresponding LoRA adapter is the live one
+        before broadcasting, and tell the worker to register the adapter on
+        vLLM under ``model_id``.
         """
         if self._inference_engine_client is None:
             raise RuntimeError(
@@ -483,9 +495,12 @@ async def save_weights_for_sampler(self) -> None:
 
         # Sync weights to inference engine
         self.prepare_for_weight_sync()
+        # Make the requested adapter live on every worker before broadcasting
+        # — otherwise we'd export some other tenant's LoRA weights to vLLM.
+        self.ensure_active_adapter("policy", model_id)
         if self.colocate_all:
             await self._inference_engine_client.wake_up(tags=["weights"])
-            self.broadcast_to_inference_engines(self._inference_engine_client)
+            self.broadcast_to_inference_engines(self._inference_engine_client, model_id=model_id)
             self.finish_weight_sync()
             await self._inference_engine_client.wake_up(tags=["kv_cache"])
         else:
@@ -493,7 +508,7 @@ async def save_weights_for_sampler(self) -> None:
             # reading partially-updated weights during the NCCL broadcast.
             await self._inference_engine_client.pause_generation()
             try:
-                self.broadcast_to_inference_engines(self._inference_engine_client)
+                self.broadcast_to_inference_engines(self._inference_engine_client, model_id=model_id)
                 self.finish_weight_sync()
             finally:
                 await self._inference_engine_client.resume_generation()
diff --git a/skyrl/backends/skyrl_train_backend.py b/skyrl/backends/skyrl_train_backend.py
index dc0478557f..626eb71c8c 100644
--- a/skyrl/backends/skyrl_train_backend.py
+++ b/skyrl/backends/skyrl_train_backend.py
@@ -938,18 +938,6 @@ def sample(
         # 1. Ensure inference engines are initialized
         self._ensure_inference_engines()
 
-        # v1 multi-LoRA: sample() is single-tenant. The inference engine path
-        # is not yet adapter-aware, so refuse if more than one adapter exists.
-        if self._base_lora_signature is not None and len(self._model_ids_to_role) > 1:
-            error = types.ErrorResponse(
-                error=(
-                    "sample() is not supported with multiple LoRA adapters in v1. "
-                    "Delete other adapters before sampling."
-                ),
-                status="error",
-            )
-            return {req_id: error for req_id, _, _, _, _ in prepared_batch.request_batch_slices}
-
         # 2. Validate single model
         unique_models = set(prepared_batch.all_model_ids)
         if len(unique_models) != 1:
@@ -1018,12 +1006,15 @@ def _sample_with_remote_client(
     ) -> dict[str, types.SampleOutput | types.ErrorResponse]:
         """Sample using RemoteInferenceClient, forwarding model input chunks directly."""
 
-        # Every sample() body must explicitly identify the target model/LoRA
-        # adapter — the client does not fall back to any default. Resolve
-        # what name the inference engine knows the policy by from config
-        # (LoRA adapter when LoRA weights are sync'd as an adapter, base
-        # model otherwise).
-        model_name = resolve_policy_model_name(self._cfg)
+        # Resolve the inference-engine model name per request. With multi-LoRA
+        # the adapter name on vLLM IS the Tinker model_id (registered by
+        # save_sampler_checkpoint via load_lora_adapter). Single-tenant /
+        # FFT path falls back to resolve_policy_model_name(cfg).
+        fallback_model_name = resolve_policy_model_name(self._cfg)
+        per_request_models = [
+            mid if (self._base_lora_signature is not None and mid in self._model_ids_to_role) else fallback_model_name
+            for mid in prepared_batch.all_model_ids
+        ]
 
         async def sample_all():
             tasks = []
@@ -1033,7 +1024,7 @@ async def sample_all():
 
                 request_payload = {
                     "json": {
-                        "model": model_name,
+                        "model": per_request_models[i],
                         "prompt": model_input.model_dump(),
                         "num_samples": 1,
                         "sampling_params": sampling_params.model_dump(),
@@ -1189,16 +1180,15 @@ def save_sampler_checkpoint(self, output_path, model_id: str, persist: bool = Tr
         self._validate_model_state(model_id)
         if self._get_role(model_id) != "policy":
             raise ValueError("save_sampler_checkpoint is only supported for policy models")
-        if self._base_lora_signature is not None and len(self._model_ids_to_role) > 1:
-            raise ValueError(
-                "save_sampler_checkpoint is not supported with multiple LoRA adapters in v1. "
-                "Delete other adapters before pushing weights to the inference engine."
-            )
 
         # Lazily create inference engines on first sampling-related call
         self._ensure_inference_engines()
 
-        asyncio.run(self._dispatch.save_weights_for_sampler())
+        # Multi-LoRA: pass model_id so the dispatch swaps the right adapter in
+        # before broadcasting and the worker registers it on vLLM under that
+        # name. None for the FFT / single-tenant path uses legacy behavior.
+        sync_id = model_id if self._base_lora_signature is not None else None
+        asyncio.run(self._dispatch.save_weights_for_sampler(model_id=sync_id))
         logger.info(f"Synced weights for {model_id} to inference engines via NCCL")
 
         if persist:

From 40b1ae49ba6c689156ffaae28bd99c8be2404320 Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Mon, 4 May 2026 22:57:21 +0000
Subject: [PATCH 13/21] [multi-lora-rl] Tolerate non-JSON error bodies in
 load/unload_lora_adapter

Both _load_on_server and _unload_on_server called await resp.json()
without try/except, so a non-JSON error body (e.g. a plain-text 5xx
from a proxy in front of vLLM) would raise a generic JSON-parse error
and lose the original status. Mirror the robust pattern from _post:
try resp.json(content_type=None), fall back to resp.text() on parse
failure, then raise_for_status with whichever body we got.

Addresses the gemini-code-assist review note on PR #1579 (see
https://github.com/NovaSky-AI/SkyRL/pull/1579).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .../inference_servers/remote_inference_client.py  | 15 ++++++++++++---
 1 file changed, 12 insertions(+), 3 deletions(-)

diff --git a/skyrl/backends/skyrl_train/inference_servers/remote_inference_client.py b/skyrl/backends/skyrl_train/inference_servers/remote_inference_client.py
index 61f756976d..5914a42ed1 100644
--- a/skyrl/backends/skyrl_train/inference_servers/remote_inference_client.py
+++ b/skyrl/backends/skyrl_train/inference_servers/remote_inference_client.py
@@ -1157,9 +1157,14 @@ async def load_lora_adapter(
         async def _load_on_server(server_url: str):
             url = f"{server_url}/v1/load_lora_adapter"
             async with session.post(url, json=payload) as resp:
-                # vLLM returns 200 with text body on success, or JSON ErrorResponse on failure
+                # vLLM returns 200 with text body on success, or JSON ErrorResponse on failure.
+                # Tolerate non-JSON error bodies (e.g. plain-text 5xx from a proxy):
+                # fall back to the text body so raise_for_status still surfaces it.
                 if resp.status >= 400:
-                    body = await resp.json()
+                    try:
+                        body = await resp.json(content_type=None)
+                    except Exception:
+                        body = await resp.text()
                     raise_for_status(resp, body)
                 return server_url, {"status": resp.status, "body": await resp.text()}
 
@@ -1192,8 +1197,12 @@ async def unload_lora_adapter(self, lora_name: str) -> Dict[str, Any]:
         async def _unload_on_server(server_url: str):
             url = f"{server_url}/v1/unload_lora_adapter"
             async with session.post(url, json=payload) as resp:
+                # See _load_on_server for the JSON/text fallback rationale.
                 if resp.status >= 400:
-                    body = await resp.json()
+                    try:
+                        body = await resp.json(content_type=None)
+                    except Exception:
+                        body = await resp.text()
                     raise_for_status(resp, body)
                 return server_url, {"status": resp.status, "body": await resp.text()}
 

From 8cff746089b756ebc783e65693cf9d25825f1e57 Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Mon, 4 May 2026 22:59:06 +0000
Subject: [PATCH 14/21] [multi-lora-rl] Update design doc + RL two-client smoke
 runbook

Updates docs/content/docs/tinker/multi_lora_design.mdx scope from
"training only" to "training + per-adapter sampling", documents the
Tinker model_id == vLLM adapter name contract, the per-tenant
lora_sync_path layout, the merge_lora=False requirement on Megatron,
and the operator's max_cpu_loras sizing contract. Adds a "PR #1579
foundation" section pointing at the upstream PR.

Adds tests/tinker/test_multi_lora_rl_two_clients.md as the manual gate:
two rl_loop clients training and sampling on independent adapters
against one server, plus contamination check, negative checks, and
troubleshooting.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .../content/docs/tinker/multi_lora_design.mdx |  48 +++++-
 .../tinker/test_multi_lora_rl_two_clients.md  | 152 ++++++++++++++++++
 2 files changed, 193 insertions(+), 7 deletions(-)
 create mode 100644 tests/tinker/test_multi_lora_rl_two_clients.md

diff --git a/docs/content/docs/tinker/multi_lora_design.mdx b/docs/content/docs/tinker/multi_lora_design.mdx
index d850040879..0a9269b312 100644
--- a/docs/content/docs/tinker/multi_lora_design.mdx
+++ b/docs/content/docs/tinker/multi_lora_design.mdx
@@ -12,13 +12,12 @@ Today the SkyRL-Train backend exposed via the Tinker API is single-tenant: a sec
 
 The driver for changing this is Trajectory AI, who want to run Tinker workloads on their own hardware and need many tenants on a shared training pool. There is no Megatron multi-tenant SFT/RL framework today; only Prime-RL has first-class multi-tenancy and only on FSDP/CP/EP. A Megatron-backed solution is therefore both a user-requested feature and a meaningful differentiator.
 
-## v1 scope
+## Scope
 
-- **Training only**, exercised via the `tinker-cookbook` `sl_loop` SFT recipe.
-- One base model, multiple LoRA adapters.
-- Fixed `(rank, alpha, target_modules)` across all adapters. Mismatched configs on a second `create_model` are hard-rejected with a clear `ValueError`.
-- `sample()` and `save_sampler_checkpoint()` raise if more than one adapter is registered. Per-adapter sampling and per-adapter vLLM weight sync are explicitly deferred.
+- **Training and per-adapter sampling** on a single Tinker server, exercised via the `tinker-cookbook` `sl_loop` (SFT) and `rl_loop` (RL) recipes.
+- One base model, multiple LoRA adapters with fixed `(rank, alpha, target_modules)` across all adapters. Mismatched configs on a second `create_model` are hard-rejected with a clear `ValueError`.
 - The FFT (no-LoRA) path stays single-tenant — the relaxation is gated behind `lora_config.rank > 0`.
+- The RL path requires `merge_lora=False` on Megatron so vLLM serves the adapter (not pre-merged weights) and supports multiple LoRA adapters concurrently. PR #1579 contributes the inference-side scaffolding (`load_lora_adapter` / `unload_lora_adapter`, `max_loras`, `max_cpu_loras`, mandatory `model` per data-plane call); this design plumbs the Tinker `model_id` through end-to-end so each tenant's adapter is registered and addressed by its own name on vLLM.
 
 ## Strategy
 
@@ -121,10 +120,45 @@ Snapshotting `grad_data` into the slot on every swap fixes this. After A's fwd_b
 - Existing `tests/tinker/test_api.py` continues to pass (single-tenant path unchanged).
 - End-to-end smoke (manual): launch the Tinker server with `trainer.strategy=megatron`, run two `tinker_cookbook.recipes.sl_loop` clients with distinct `model_id`s in parallel against `base_url=http://localhost:8000`, verify both converge on their respective tasks and GPU memory stays bounded.
 
-## Out of scope (explicit non-goals for v1)
+## Per-adapter sampling and weight sync (RL path)
+
+For RL, sampling has to be per-adapter: each tenant runs `forward_backward → optim_step → save_weights_for_sampler → sample` on its own model_id, and vLLM has to know which weights to use for which call.
+
+The plumbing is straightforward once the SFT scaffolding is in place:
+
+- **Adapter name = Tinker model_id.** The Tinker `model_id` is forwarded as the vLLM adapter name end-to-end. There is no separate naming layer.
+- **Per-tenant `lora_sync_path`.** When `model_id` is set, the worker writes its adapter into `os.path.join(cfg.trainer.policy.model.lora.lora_sync_path, model_id)/` so concurrent saves from different tenants don't collide. Single-tenant calls (`model_id=None`, the FFT path or pre-Tinker callers) keep the legacy shared path.
+- **`save_sampler_checkpoint(model_id)`** swaps the requested adapter live (via `WorkerDispatch.ensure_active_adapter`), then broadcasts to vLLM. The worker calls `RemoteInferenceClient.load_lora_adapter(model_id, lora_sync_path/model_id, load_inplace=True)` so re-syncs of the same adapter overwrite vLLM's slot in place (no fresh int id, no eviction churn).
+- **`sample(prepared_batch)`** now resolves the data-plane `model` per request from `prepared_batch.all_model_ids[i]` when LoRA is active, falling back to `resolve_policy_model_name(cfg)` only on the FFT / single-tenant path. The previous "raise if >1 adapter" guards are gone.
+
+### vLLM capacity contract
+
+vLLM exposes two LoRA capacity knobs:
+
+- `max_loras` — concurrent adapters in a single GPU batch.
+- `max_cpu_loras` — total LoRA capacity in vLLM's CPU LRU cache (defaults to `max_loras` when unset).
+
+Both are config fields on `SkyRLLoraConfig` (`config.py`, surfaced in `ppo_base_config.yaml`). For multi-tenant RL, `max_cpu_loras` MUST be set to at least the expected number of concurrent registered adapters; otherwise vLLM will silently evict an adapter from its CPU cache and the next `sample()` against that adapter will 404. The Tinker server doesn't auto-size this — operators set it explicitly when they expect N tenants. Detecting and reloading-from-disk on 404 is a future improvement.
+
+### Concurrency between training and sampling
+
+With `colocate_all=True`, vLLM is asleep during training and woken in `save_weights_for_sampler` for the broadcast + KV-cache wake. With multi-tenant RL the engine queue's destructive-barrier scheduling (`engine.py`) keeps each model_id's `forward_backward / optim_step / save_weights_for_sampler / sample` chain coherent. For non-colocated mode the same applies — `pause_generation / resume_generation` brackets the broadcast, and per-adapter sample requests fan out independently.
+
+## PR #1579 (foundation)
+
+The RL multi-tenancy work sits on top of [NovaSky-AI/SkyRL#1579](https://github.com/NovaSky-AI/SkyRL/pull/1579) (`hao-aaron:multi-lora`), which provides:
+
+- `RemoteInferenceClient.load_lora_adapter(name, path, load_inplace)` and `unload_lora_adapter(name)`.
+- `SKYRL_LORA_ADAPTER_NAME` constant + `resolve_policy_model_name(cfg)` for the single-tenant fallback.
+- Mandatory `model` field on every data-plane call (`generate / sample / chat_completion / completion / render_chat_completion`).
+- `max_loras` + `max_cpu_loras` config knobs plumbed through to vLLM CLI args.
+
+Our `multi_lora_rl` branch is based on PR #1579's HEAD with our SFT AdapterStore commits cherry-picked on top.
+
+## Out of scope (explicit non-goals)
 
-- Per-adapter `sample()` / vLLM weight sync — `sample` raises if more than one adapter is registered.
 - Adapter-only checkpoint export — `save_checkpoint` still saves the whole base+LoRA state per swap.
 - Variable rank / alpha / target_modules across adapters.
 - Critic role multi-tenancy — Megatron critic is `NotImplementedError` today; no change.
 - `HybridDeviceOptimizer` path — TODO comment only.
+- Auto-reload of LoRA adapters evicted from vLLM's CPU LRU cache; for now `max_cpu_loras` must be set ≥ expected concurrent adapters.
diff --git a/tests/tinker/test_multi_lora_rl_two_clients.md b/tests/tinker/test_multi_lora_rl_two_clients.md
new file mode 100644
index 0000000000..a34c3cfb4b
--- /dev/null
+++ b/tests/tinker/test_multi_lora_rl_two_clients.md
@@ -0,0 +1,152 @@
+# Multi-LoRA Megatron RL — two-client smoke test
+
+The gate before merging the `multi_lora_rl` branch. Extends the SFT smoke
+([`test_multi_lora_smoke_two_clients.md`](./test_multi_lora_smoke_two_clients.md))
+to cover the per-adapter sampling + weight-sync path that powers RL.
+
+Two `tinker-cookbook` `rl_loop` clients, each with their own LoRA adapter,
+should train and sample independently against a single SkyRL Tinker API
+server with the Megatron backend and `merge_lora=False` so vLLM serves
+adapters by name.
+
+## Prerequisites
+
+- At least one CUDA GPU. With `colocate_all=true` (the simplest RL config),
+  training and inference share the same GPUs and vLLM sleeps during
+  training.
+- `tinker-cookbook` and `datasets` available (or invoked via `uv run`).
+- `multi_lora_rl` branch checked out.
+
+## Step 1 — Start the Tinker API server
+
+```bash
+cd /path/to/SkyRL
+
+uv run --extra tinker --extra megatron -m skyrl.tinker.api \
+    --host 0.0.0.0 \
+    --port 8000 \
+    --base-model Qwen/Qwen3-0.6B \
+    --backend megatron \
+    --backend-config '{
+        "strategy": "megatron",
+        "trainer.placement.policy_num_gpus_per_node": 1,
+        "trainer.placement.policy_num_nodes": 1,
+        "trainer.placement.colocate_all": true,
+        "trainer.policy.megatron_config.tensor_model_parallel_size": 1,
+        "trainer.policy.megatron_config.pipeline_model_parallel_size": 1,
+        "trainer.policy.megatron_config.lora_config.merge_lora": false,
+        "trainer.policy.model.lora.max_loras": 4,
+        "trainer.policy.model.lora.max_cpu_loras": 4
+    }'
+```
+
+Critical knobs vs the SFT runbook:
+
+- `merge_lora: false` — Megatron must not pre-merge LoRA into the base
+  weights, otherwise vLLM serves the merged model and there's nothing
+  per-adapter to switch on.
+- `max_loras` ≥ expected concurrent adapters in a single batch (typically 2
+  for two `rl_loop` clients).
+- `max_cpu_loras` ≥ expected total adapters. **Must** be set explicitly;
+  the server doesn't auto-size. If too low, vLLM's CPU LRU evicts an
+  adapter and the next `sample()` against it 404s.
+
+Wait for `init policy model done`. The first client's `create_model` will
+trigger the policy build + AdapterStore bootstrap; subsequent clients
+register additional adapters (`Registered additional LoRA adapter '<id>'`).
+
+## Step 2 — Run two `rl_loop` clients in parallel
+
+```bash
+# Terminal 2 — client A
+TINKER_API_KEY=tml-dummy uv run --with tinker --with tinker-cookbook --with datasets --with torch \
+    python -m tinker_cookbook.recipes.rl_loop \
+    base_url=http://localhost:8000 \
+    model_name="Qwen/Qwen3-0.6B" \
+    lora_rank=32 \
+    log_path=/tmp/rl_loop_a.log
+```
+
+```bash
+# Terminal 3 — client B
+TINKER_API_KEY=tml-dummy uv run --with tinker --with tinker-cookbook --with datasets --with torch \
+    python -m tinker_cookbook.recipes.rl_loop \
+    base_url=http://localhost:8000 \
+    model_name="Qwen/Qwen3-0.6B" \
+    lora_rank=32 \
+    log_path=/tmp/rl_loop_b.log
+```
+
+Stagger the launches by ~10 s. Both clients **must** use the same
+`lora_rank` and `model_name` (mismatches are hard-rejected at
+`create_model` with `LoRA signature mismatch …`).
+
+## Step 3 — What success looks like
+
+- Both clients converge on their respective rewards. Reward trends upward
+  in both `rl_loop_a.log` and `rl_loop_b.log`.
+- Server log shows, for each client, the per-step sequence:
+  - `forward_backward` + `optim_step` (training)
+  - `save_sampler_checkpoint` (writes adapter into
+    `lora_sync_path/<model_id>/`, calls `load_lora_adapter(<model_id>, …,
+    load_inplace=True)` on every vLLM server)
+  - `sample(model=<model_id>)` against the right adapter
+- vLLM server logs show two distinct adapter names registered (the two
+  Tinker `model_id`s) and `sample` requests routed to each.
+- GPU memory stays bounded as the second client connects (single base
+  model, two LoRA adapters; CPU LRU holds the same two).
+
+## Step 4 — Per-adapter contamination check
+
+Stop client B mid-training, let client A continue for ~50 RL steps,
+then restart client B. Client B's reward trajectory should resume from
+where it left off, NOT regress to a fresh-start baseline. If B's reward
+craters back to zero, A's training was bleeding into B's adapter.
+
+## Step 5 — Cleanup
+
+Either let the recipes finish (they call `unload_model` on exit) or kill
+both processes. The session-cleanup loop will eventually unload stale
+adapters. When the *last* registered adapter is unloaded the server tears
+down the Ray runtime via `ray.shutdown()`; subsequent `create_model`
+rebuilds it.
+
+## Negative checks
+
+1. **Mismatched rank.** Client A `lora_rank=32`, client B `lora_rank=16` —
+   second `create_model` must fail with `LoRA signature mismatch …`.
+2. **`max_cpu_loras=1` with two adapters.** Start the server with
+   `max_cpu_loras=1`, run both clients. The second client's first
+   `sample()` should fail (adapter evicted before sampling). This is
+   expected and documents why the operator must size `max_cpu_loras`
+   correctly.
+3. **`merge_lora: true`.** Start the server with `merge_lora=true` and
+   try the same two clients. The second `create_model` will succeed
+   (training is fine), but `sample()` against the *non-active* adapter
+   returns the wrong adapter's output because vLLM only ever sees the
+   merged base. Document; v1 requires `merge_lora=false` for RL.
+
+## Troubleshooting
+
+- **`sample()` 404 on `lora_name=…`.** Either `save_sampler_checkpoint`
+  wasn't called for that model_id before the sample (Tinker recipe bug),
+  or `max_cpu_loras` is too low and vLLM evicted the adapter. Check the
+  vLLM server log for `Adapter X not found / evicted`.
+- **`KeyError: lora_sync_path/<model_id>`** on the worker side — the
+  per-tenant subdir wasn't created. Confirm `model_id` is being passed
+  through to `broadcast_to_inference_engines` (controller log line should
+  read `Synced weights for <model_id> to inference engines via NCCL`).
+- **Server hangs on the second `save_sampler_checkpoint`.** Likely the
+  first client's vLLM wake-up is mid-flight. The dispatch's
+  `prepare_for_weight_sync` + `ensure_active_adapter` should serialize
+  this; if you see deadlocks, capture a thread dump on the controller and
+  the policy worker.
+- **`AttributeError: load_lora_adapter`** on `RemoteInferenceClient` —
+  the server is running an older binary that pre-dates PR #1579. Make
+  sure the server started from this branch.
+
+## Reference
+
+- Design: [`docs/content/docs/tinker/multi_lora_design.mdx`](../../docs/content/docs/tinker/multi_lora_design.mdx).
+- Foundation PR: [NovaSky-AI/SkyRL#1579](https://github.com/NovaSky-AI/SkyRL/pull/1579).
+- SFT runbook: [`test_multi_lora_smoke_two_clients.md`](./test_multi_lora_smoke_two_clients.md).

From a46b5878bcb68d0ec1b4c081703729a7b90355ab Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Tue, 5 May 2026 00:31:23 +0000
Subject: [PATCH 15/21] [multi-lora-rl] Allow mixed model_ids in a single
 sample() batch
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The Tinker engine batches sample requests across model_ids in
find_batchable_sample, then dispatches one prepared_batch to
backend.sample(). Our previous "exactly one model_id per batch" guard
short-circuited multi-tenant RL — when both rl_loop clients had
sample() requests pending in the same engine tick, the batched call
hit the guard and returned 400 to both.

Replaces the unique-model check with a per-request validation: every
model_id must be a known policy, but multiple distinct policy
model_ids in one batch are fine. Routing per request is already
handled by _sample_with_remote_client via the per-request `model`
field on the data plane.

Co-Authored-By: Eric Tang <erictang000@gmail.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 skyrl/backends/skyrl_train_backend.py | 18 +++++++++++-------
 1 file changed, 11 insertions(+), 7 deletions(-)

diff --git a/skyrl/backends/skyrl_train_backend.py b/skyrl/backends/skyrl_train_backend.py
index b0fbb6421b..3bc1893957 100644
--- a/skyrl/backends/skyrl_train_backend.py
+++ b/skyrl/backends/skyrl_train_backend.py
@@ -877,18 +877,22 @@ def sample(
         # 1. Ensure inference engines are initialized
         self._ensure_inference_engines()
 
-        # 2. Validate single model
+        # 2. Validate every model_id in the batch is a known policy. Multi-LoRA
+        # RL legitimately mixes adapters in one batched sample call (the engine
+        # batches across model_ids in find_batchable_sample); we route per
+        # request via the `model` field in _sample_with_remote_client below.
         unique_models = set(prepared_batch.all_model_ids)
-        if len(unique_models) != 1:
+        unknown = [mid for mid in unique_models if mid not in self._model_ids_to_role]
+        if unknown:
             error = types.ErrorResponse(
-                error=f"Expected exactly one model_id for sampling, got {unique_models}", status="error"
+                error=f"Sampling requested for unknown model_id(s): {sorted(unknown)}", status="error"
             )
             return {req_id: error for req_id, _, _, _, _ in prepared_batch.request_batch_slices}
-        model_id = next(iter(unique_models))
-        role = self._model_ids_to_role.get(model_id)
-        if role != "policy":
+        non_policy = [mid for mid in unique_models if self._model_ids_to_role.get(mid) != "policy"]
+        if non_policy:
             error = types.ErrorResponse(
-                error=f"Sampling is only supported for policy models, got '{model_id}'", status="error"
+                error=f"Sampling is only supported for policy models, got non-policy: {sorted(non_policy)}",
+                status="error",
             )
             return {req_id: error for req_id, _, _, _, _ in prepared_batch.request_batch_slices}
 

From edefc84644a41ec40153cbd645f2fb38077a9482 Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Tue, 5 May 2026 00:43:12 +0000
Subject: [PATCH 16/21] [multi-lora-rl] Pass load_inplace=True to vLLM
 load_lora_adapter
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The merged unload-then-load workaround from main wasn't sufficient for
re-syncing an adapter — vLLM still returned 400 "adapter already
loaded" on the second sync of the same name (e.g. when a tenant calls
save_sampler_checkpoint a second time, or when the unload step
returns 200 but the cached LoRARequest hasn't been evicted yet).

vLLM's own error message instructs the caller to set load_inplace=True
in that case, which is what PR #1579 originally did. Restore that
behavior: thread the load_inplace parameter (default True, exposed on
the public API) into the /v1/load_lora_adapter payload, drop the
separate _unload_on_server pre-step. The standalone unload_lora_adapter
method still exists for callers that explicitly want eviction.

Fixes the rl_loop runtime error:
  ClientResponseError: 400, message="The lora adapter '<id>' has
  already been loaded. If you want to load the adapter in place, set
  'load_inplace' to True."

Co-Authored-By: Eric Tang <erictang000@gmail.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .../remote_inference_client.py                | 29 +++++--------------
 1 file changed, 8 insertions(+), 21 deletions(-)

diff --git a/skyrl/backends/skyrl_train/inference_servers/remote_inference_client.py b/skyrl/backends/skyrl_train/inference_servers/remote_inference_client.py
index e5867bd260..4bcd33f1ba 100644
--- a/skyrl/backends/skyrl_train/inference_servers/remote_inference_client.py
+++ b/skyrl/backends/skyrl_train/inference_servers/remote_inference_client.py
@@ -1154,38 +1154,26 @@ async def load_lora_adapter(
         same name is already registered on the server, vLLM replaces it
         inplace, preserving its internal int id.
 
-        TODO(aaron): remove _unload_on_server and add back "load_inplace": True to payload after
-        vllm lora disk load bug is fixed.
-
         Args:
             lora_name: Name to register the adapter under on each server.
             lora_path: Path to the LoRA adapter on disk (must be accessible from servers).
-            load_inplace: When True (default), reloading a previously-loaded and cached
-                adapter with the same name replaces it with the on-disk lora.
+            load_inplace: When True (default), vLLM replaces an existing adapter
+                with the same name in place, preserving its int id. Without
+                this, the second call with the same name returns a 400
+                "adapter already loaded".
 
         Returns:
             Dict mapping server_url to response.
         """
-        # Both endpoints return plain text on success or JSON ErrorResponse on failure,
-        # so we use a session.post directly rather than _call_all_servers (which expects JSON).
-        # ``load_inplace`` is currently unused: until the upstream vLLM in-place reload bug
-        # is fixed (see TODO above) we always do explicit unload-then-load.
-        _ = load_inplace
+        # vLLM /v1/load_lora_adapter returns plain text on success and JSON
+        # ErrorResponse on failure, so we use a session.post directly rather
+        # than _call_all_servers (which expects JSON).
         session = await self._get_session()
 
-        async def _unload_on_server(server_url: str) -> None:
-            """Remove the cached LoRARequest server-side. 404 on the first sync is expected."""
-            url = f"{server_url}/v1/unload_lora_adapter"
-            async with session.post(url, json={"lora_name": lora_name}) as resp:
-                if resp.status == 404:
-                    return
-                if resp.status >= 400:
-                    body = await resp.json()
-                    raise_for_status(resp, body)
+        payload = {"lora_name": lora_name, "lora_path": lora_path, "load_inplace": load_inplace}
 
         async def _load_on_server(server_url: str):
             url = f"{server_url}/v1/load_lora_adapter"
-            payload = {"lora_name": lora_name, "lora_path": lora_path}
             async with session.post(url, json=payload) as resp:
                 # vLLM returns 200 with text body on success, or JSON ErrorResponse on failure.
                 # Tolerate non-JSON error bodies (e.g. plain-text 5xx from a proxy):
@@ -1198,7 +1186,6 @@ async def _load_on_server(server_url: str):
                     raise_for_status(resp, body)
                 return server_url, {"status": resp.status, "body": await resp.text()}
 
-        await asyncio.gather(*[_unload_on_server(url) for url in self.server_urls])
         results = await asyncio.gather(*[_load_on_server(url) for url in self.server_urls])
 
         logger.info(f"Loaded LoRA adapter '{lora_name}' from {lora_path}")

From fa3bfbcc5d715369ba1b835056070d29b5f42238 Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Tue, 5 May 2026 00:44:57 +0000
Subject: [PATCH 17/21] Revert "[multi-lora-rl] Pass load_inplace=True to vLLM
 load_lora_adapter"

This reverts commit edefc84644a41ec40153cbd645f2fb38077a9482.
---
 .../remote_inference_client.py                | 29 ++++++++++++++-----
 1 file changed, 21 insertions(+), 8 deletions(-)

diff --git a/skyrl/backends/skyrl_train/inference_servers/remote_inference_client.py b/skyrl/backends/skyrl_train/inference_servers/remote_inference_client.py
index 4bcd33f1ba..e5867bd260 100644
--- a/skyrl/backends/skyrl_train/inference_servers/remote_inference_client.py
+++ b/skyrl/backends/skyrl_train/inference_servers/remote_inference_client.py
@@ -1154,26 +1154,38 @@ async def load_lora_adapter(
         same name is already registered on the server, vLLM replaces it
         inplace, preserving its internal int id.
 
+        TODO(aaron): remove _unload_on_server and add back "load_inplace": True to payload after
+        vllm lora disk load bug is fixed.
+
         Args:
             lora_name: Name to register the adapter under on each server.
             lora_path: Path to the LoRA adapter on disk (must be accessible from servers).
-            load_inplace: When True (default), vLLM replaces an existing adapter
-                with the same name in place, preserving its int id. Without
-                this, the second call with the same name returns a 400
-                "adapter already loaded".
+            load_inplace: When True (default), reloading a previously-loaded and cached
+                adapter with the same name replaces it with the on-disk lora.
 
         Returns:
             Dict mapping server_url to response.
         """
-        # vLLM /v1/load_lora_adapter returns plain text on success and JSON
-        # ErrorResponse on failure, so we use a session.post directly rather
-        # than _call_all_servers (which expects JSON).
+        # Both endpoints return plain text on success or JSON ErrorResponse on failure,
+        # so we use a session.post directly rather than _call_all_servers (which expects JSON).
+        # ``load_inplace`` is currently unused: until the upstream vLLM in-place reload bug
+        # is fixed (see TODO above) we always do explicit unload-then-load.
+        _ = load_inplace
         session = await self._get_session()
 
-        payload = {"lora_name": lora_name, "lora_path": lora_path, "load_inplace": load_inplace}
+        async def _unload_on_server(server_url: str) -> None:
+            """Remove the cached LoRARequest server-side. 404 on the first sync is expected."""
+            url = f"{server_url}/v1/unload_lora_adapter"
+            async with session.post(url, json={"lora_name": lora_name}) as resp:
+                if resp.status == 404:
+                    return
+                if resp.status >= 400:
+                    body = await resp.json()
+                    raise_for_status(resp, body)
 
         async def _load_on_server(server_url: str):
             url = f"{server_url}/v1/load_lora_adapter"
+            payload = {"lora_name": lora_name, "lora_path": lora_path}
             async with session.post(url, json=payload) as resp:
                 # vLLM returns 200 with text body on success, or JSON ErrorResponse on failure.
                 # Tolerate non-JSON error bodies (e.g. plain-text 5xx from a proxy):
@@ -1186,6 +1198,7 @@ async def _load_on_server(server_url: str):
                     raise_for_status(resp, body)
                 return server_url, {"status": resp.status, "body": await resp.text()}
 
+        await asyncio.gather(*[_unload_on_server(url) for url in self.server_urls])
         results = await asyncio.gather(*[_load_on_server(url) for url in self.server_urls])
 
         logger.info(f"Loaded LoRA adapter '{lora_name}' from {lora_path}")

From cb3861421ae3a09dbd36a9587ffc0f2dfb4b3afb Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Tue, 5 May 2026 01:01:16 +0000
Subject: [PATCH 18/21] x

---
 tests/tinker/test_multi_lora_rl_two_clients.md | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/tests/tinker/test_multi_lora_rl_two_clients.md b/tests/tinker/test_multi_lora_rl_two_clients.md
index 9e098ab6ff..9a036697b1 100644
--- a/tests/tinker/test_multi_lora_rl_two_clients.md
+++ b/tests/tinker/test_multi_lora_rl_two_clients.md
@@ -29,12 +29,16 @@ uv run --extra tinker --extra megatron -m skyrl.tinker.api \
     --backend megatron \
     --backend-config '{
         "strategy": "megatron",
-        "trainer.placement.policy_num_gpus_per_node": 1,
+        "trainer.placement.policy_num_gpus_per_node": 4,
         "trainer.placement.policy_num_nodes": 1,
         "trainer.placement.colocate_all": false,
         "trainer.policy.megatron_config.tensor_model_parallel_size": 1,
         "trainer.policy.megatron_config.pipeline_model_parallel_size": 1,
+        "trainer.micro_train_batch_size_per_gpu": 64,
+        "trainer.micro_forward_batch_size_per_gpu": 64,
         "trainer.policy.megatron_config.lora_config.merge_lora": false,
+        "generator.inference_engine.num_engines": 1,
+        "generator.inference_engine.tensor_parallel_size": 1,
         "trainer.policy.model.lora.max_loras": 4,
         "trainer.policy.model.lora.max_cpu_loras": 4
     }'

From e18e82ba92fd0deacb842b67cf58cb541f4a58b4 Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Wed, 6 May 2026 00:09:50 +0000
Subject: [PATCH 19/21] [docs] Design doc for non-colocated sample routing via
 EXTERNAL path

Adds docs/content/docs/tinker/async_sample_routing.mdx describing the
plan to route SkyRL-Train sample requests through the existing
EXTERNAL fan-out path (api.py:1039-1064) instead of the engine's
synchronous loop. The engine already excludes EXTERNAL futures from
its scheduler; we just need to point a new
BackendForwardingInferenceClient at the engine-managed vLLM.

Covers: synchronization invariants (I1-I4) that already hold via the
SDK + checkpoint validation + vLLM pause/resume, files to add/modify
(EngineStateDB row, BackendForwardingInferenceClient,
SkyRLTrainBackend._publish_engine_state, api.py lifespan wiring),
trade-offs vs. dual-loop-in-engine and full async refactors, failure
modes, testing plan, and explicit non-goals (training-side parallelism,
auto-recovery from vLLM eviction).

Co-Authored-By: Eric Tang <erictang000@gmail.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .../docs/tinker/async_sample_routing.mdx      | 401 ++++++++++++++++++
 1 file changed, 401 insertions(+)
 create mode 100644 docs/content/docs/tinker/async_sample_routing.mdx

diff --git a/docs/content/docs/tinker/async_sample_routing.mdx b/docs/content/docs/tinker/async_sample_routing.mdx
new file mode 100644
index 0000000000..2959743563
--- /dev/null
+++ b/docs/content/docs/tinker/async_sample_routing.mdx
@@ -0,0 +1,401 @@
+---
+title: "Non-colocated Sample Routing — Design"
+---
+
+# Non-colocated Sample Routing — Design
+
+This document describes the design for processing `sample()` requests in
+parallel with training when running the SkyRL Tinker server with the
+SkyRL-Train (FSDP / Megatron) backend in non-colocated mode (`colocate_all=False`).
+
+## Problem
+
+The Tinker engine's main loop (`skyrl/tinker/engine.py:721`) processes pending
+requests synchronously per 100 ms tick, in fixed order:
+
+```python
+self.process_batch_requests(forward_backward_requests, ...)
+self.process_batch_requests(forward_requests, ...)
+self.process_batch_requests(sample_requests, ...)
+self.process_single_requests(other_requests)
+```
+
+This is correct for the JAX backend (single model serves both training and
+sampling — they cannot run concurrently) and for SkyRL-Train colocated mode
+(vLLM is asleep during training, woken only inside `save_weights_for_sampler`
+→ broadcast → `sample`).
+
+For SkyRL-Train **non-colocated** mode the assumption breaks down:
+
+- Training (`forward_backward`, `optim_step`) runs on dedicated policy worker
+  GPUs.
+- vLLM runs on its own GPUs and is **always on** — generation can run at any
+  time.
+- The two interact only at `save_weights_for_sampler`, which broadcasts new
+  weights into vLLM bracketed by `pause_generation` / `resume_generation`.
+
+In this regime, serializing `sample` behind `forward_backward` wastes vLLM
+capacity. A `sample` request submitted at the start of a 30 s training step
+sits idle for that 30 s even though vLLM is free.
+
+For multi-tenant RL (multiple `rl_loop` clients on one Tinker server), the
+problem compounds: every `sample` from any tenant queues behind every other
+tenant's training step.
+
+## Existing infrastructure we can lean on
+
+The Tinker server already has a parallel-sample code path for genuinely
+external vLLM deployments (`engine_config.external_inference_url`):
+
+1. **API handler dispatches sample directly, off the engine queue**
+   (`api.py:1039-1064`). When `app.state.external_inference_client` is
+   non-null, `POST /api/v1/asample` creates a `FutureDB` row with
+   `RequestType.EXTERNAL`, then `asyncio.create_task(...)` to forward to vLLM
+   and return the future_id immediately. The engine never sees the request.
+
+2. **Engine excludes EXTERNAL from its scheduler** (`engine.py:431`):
+   `find_single_requests` filters
+   `where(FutureDB.request_type != types.RequestType.EXTERNAL)`. EXTERNAL
+   futures live entirely in the API process's asyncio loop.
+
+3. **`ExternalInferenceClient.call_and_store_result`** (in
+   `skyrl/tinker/extra/external_inference.py`) writes the result back into
+   `FutureDB` directly when the call returns. The
+   `/api/v1/retrieve_future` polling loop is oblivious to who the writer is.
+
+So the cross-process plumbing for parallel sample is already in place. The
+gap is that it's wired only for an external vLLM URL configured at server
+start. We need to point it at the vLLM that the SkyRL-Train backend stands
+up internally.
+
+## Proposed design
+
+Route SkyRL-Train sample requests through the existing EXTERNAL path,
+adapted to talk to the engine-managed vLLM instead of a user-supplied
+external URL.
+
+### Flow
+
+```
+                    ┌─ POST /api/v1/asample ─────────────────────────┐
+        client ─────┤                                                │
+                    │  api.py:                                       │
+                    │    create FutureDB(type=EXTERNAL)              │
+                    │    asyncio.create_task(client.call_and_store)  │
+                    │    return future_id (immediately)              │
+                    └────────────────────────────────────────────────┘
+                                    │
+                                    ▼
+                    ┌─ BackendForwardingInferenceClient ─────────────┐
+                    │  POST {vllm_proxy_url}/v1/completions          │
+                    │    model=model_id   (LoRA registered by engine)│
+                    │  on completion: write FutureDB.result_data     │
+                    └────────────────────────────────────────────────┘
+
+   (parallel)
+                    ┌─ engine subprocess ────────────────────────────┐
+                    │  process_pending_requests loop:                │
+                    │    forward_backward / optim_step /             │
+                    │    save_weights_for_sampler / ...              │
+                    │  (NO sample work here)                         │
+                    │                                                │
+                    │  vLLM lifecycle owned here. Adapter loaded     │
+                    │  during save_weights_for_sampler via           │
+                    │  RemoteInferenceClient.load_lora_adapter.      │
+                    └────────────────────────────────────────────────┘
+```
+
+The engine still processes `forward_backward`, `optim_step`,
+`save_weights_for_sampler`, `save_weights`, `load_weights`, and the rest of
+the request types in its serial loop. Only `sample` (and only when
+non-colocated) is hoisted into the API process's async loop.
+
+### Synchronization invariants
+
+These are already enforced by the existing API and SDK; calling them out so
+the EXTERNAL refactor preserves them.
+
+**(I1) Sample for checkpoint X requires that save-X has completed.** The
+Tinker SDK serializes `save_weights_for_sampler(name=X)` and
+`sample(checkpoint_id=X)` per training client — the second future depends
+on the first. On the server, the API handler at `api.py:1037` calls
+`validate_checkpoint(...)` before creating the sample future; checkpoints
+are only marked ready inside the engine after the broadcast completes
+(`engine._checkpoint_status_context` in `process_save_weights_for_sampler`).
+Sample-for-X cannot reach the EXTERNAL fan-out before save-X finalises in
+the engine. ✓
+
+**(I2) In-flight sample during a re-broadcast for the same model.** Worker
+side, `WorkerDispatch.save_weights_for_sampler` brackets the broadcast with
+`inference_engine_client.pause_generation()` /
+`resume_generation()`. vLLM's KEEP-mode pause freezes in-flight requests in
+its scheduler; they resume after `resume_generation`. No client-side
+coordination needed. ✓
+
+**(I3) Result writes don't conflict with engine writes.** EXTERNAL sample
+results are written by the API process; all other request results are
+written by the engine. Different `FutureDB` rows, no overlap. SQLite WAL
+(already enabled by `enable_sqlite_wal`) handles concurrent readers.
+
+**(I4) The colocated path is unchanged.** When `colocate_all=True`, the API
+must NOT install the forwarding client — vLLM is asleep during training and
+the engine's synchronous `sample` path is still what wakes it, broadcasts,
+and samples. The lifespan code gates strictly on
+`colocate_all == False && backend == "skyrl_train"`.
+
+### Files to add / modify
+
+#### New: `skyrl/tinker/extra/backend_forwarding_inference.py`
+
+Adapter from the EXTERNAL contract to the SkyRL-Train-managed vLLM. ~80 LOC.
+
+```python
+class BackendForwardingInferenceClient:
+    """Forwards EXTERNAL sample requests to the SkyRL-Train-managed vLLM.
+
+    Differs from ExternalInferenceClient:
+      - Reads the vLLM proxy URL from EngineStateDB (written by the engine
+        after _create_new_inference_client) rather than EngineConfig.
+      - Uses model=<model_id> for LoRA sampling (the engine already loaded
+        the adapter under that name in save_weights_for_sampler via
+        RemoteInferenceClient.load_lora_adapter).
+      - No checkpoint extraction step — the worker's
+        _save_lora_adapters_and_sync wrote the adapter to lora_sync_path
+        and registered it with vLLM directly.
+    """
+
+    def __init__(self, engine_config: EngineConfig, db_engine):
+        self.engine_config = engine_config
+        self.db_engine = db_engine
+        self._cached_proxy_url: str | None = None
+        self._concurrency = asyncio.Semaphore(
+            getattr(engine_config, "max_concurrent_samples", 64)
+        )
+
+    async def _resolve_proxy_url(self) -> str:
+        # Read from EngineStateDB; cache; refresh on 5xx.
+        ...
+
+    async def call_and_store_result(self, request_id, sample_req,
+                                    model_id, checkpoint_id, *, base_model=None):
+        async with self._concurrency:
+            try:
+                proxy_url = await self._resolve_proxy_url()
+                # POST {proxy_url}/v1/completions with model=model_id (or
+                # base_model when present). Reuse the rendering logic from
+                # ExternalInferenceClient._forward_to_engine.
+                result = await self._forward(...)
+                result_data, status = result.model_dump(), RequestStatus.COMPLETED
+            except Exception as e:
+                result_data, status = {"error": str(e), "status": "failed"}, RequestStatus.FAILED
+            await self._write_result(request_id, result_data, status)
+```
+
+#### New DB row: `EngineStateDB`
+
+A single-row table holding inference-side handoff state from engine →
+API. New file `skyrl/tinker/db_models.py` addition:
+
+```python
+class EngineStateDB(SQLModel, table=True):
+    __tablename__ = "engine_state"
+
+    # Always 1; SQLModel needs a primary key.
+    singleton_id: int = Field(default=1, primary_key=True)
+
+    inference_proxy_url: str | None = None
+    """Proxy URL of the engine-managed vLLM, populated after
+    SkyRLTrainBackend._create_new_inference_client. None when no vLLM has
+    been stood up yet (FFT path, colocated, or pre-first-create_model)."""
+
+    inference_server_urls: list[str] = Field(default_factory=list, sa_type=JSON)
+    """Backend vLLM URLs (data-parallel servers). Optional;
+    BackendForwardingInferenceClient only needs proxy_url for /v1/completions."""
+
+    updated_at: datetime = Field(
+        default_factory=lambda: datetime.now(timezone.utc),
+        sa_type=DateTime(timezone=True),
+    )
+```
+
+The engine writes this row inside `_create_new_inference_client` after
+`build_new_inference_client(...)` returns (so the proxy_url is known). The
+API reads it lazily on first sample.
+
+#### Modify: `skyrl/backends/skyrl_train_backend.py`
+
+In `_create_new_inference_client`:
+
+```python
+client, server_setup = build_new_inference_client(
+    self._cfg,
+    self._tokenizer,
+    placement_group=self._colocate_pg if is_colocated else None,
+)
+self._inference_router = server_setup.router
+self._server_groups = server_setup.server_groups
+self._inference_engine_client = client
+
+# NEW: publish vLLM URLs for the API process's BackendForwardingInferenceClient.
+self._publish_engine_state(server_setup)
+```
+
+`_publish_engine_state` writes `EngineStateDB.singleton_id=1` with the
+current `proxy_url` and `server_urls`. Idempotent upsert. On
+`delete_model`-driven `ray.shutdown()`, clear the row (proxy_url=None) so
+the API knows there's no vLLM to talk to.
+
+#### Modify: `skyrl/tinker/api.py` — lifespan
+
+```python
+backend_name = app.state.engine_config.backend
+backend_cfg = app.state.engine_config.backend_config or {}
+is_colocated = bool(backend_cfg.get("trainer.placement.colocate_all", True))
+
+if app.state.engine_config.external_inference_url:
+    app.state.external_inference_client = ExternalInferenceClient(
+        app.state.engine_config, app.state.db_engine,
+    )
+    logger.info(f"External engine configured: {app.state.engine_config.external_inference_url}")
+elif backend_name == "skyrl_train" and not is_colocated:
+    app.state.external_inference_client = BackendForwardingInferenceClient(
+        app.state.engine_config, app.state.db_engine,
+    )
+    logger.info("Engine-managed vLLM forwarding enabled (non-colocated SkyRL-Train)")
+else:
+    app.state.external_inference_client = None
+    logger.info("Using internal engine for inference")
+```
+
+Note: `backend_name` for the SkyRL-Train Megatron path is the string the
+operator passes to `--backend` (currently `"megatron"` or `"fsdp"` per
+`skyrl/tinker/engine.py:177-183`). The check should match those names —
+either by allowlist (`backend_name in ("megatron", "fsdp")`) or by reading
+the resolved backend class. Easiest: allowlist.
+
+Also: the colocate detection from `backend_config` is fragile (assumes the
+key exists). Better: have the engine write `is_colocated` into
+`EngineStateDB` alongside the URL, and have the API read it after first
+write. Until then, default to colocated (current behavior preserved).
+
+#### No changes required to engine.py
+
+The engine's `find_single_requests` already excludes EXTERNAL. The engine
+processes `save_weights_for_sampler` synchronously as today, which is what
+populates the LoRA adapter on vLLM under `model=<model_id>` so subsequent
+sample requests find it.
+
+### Backpressure
+
+Two places to bound concurrency:
+
+1. **Per-client semaphore** inside `BackendForwardingInferenceClient` (the
+   `self._concurrency` above). Default: 64. Configurable via a new
+   `EngineConfig.max_concurrent_samples`.
+2. **vLLM-side `max_num_seqs`** is the ultimate cap; vLLM queues beyond
+   that internally.
+
+The semaphore matters mostly for shielding the FastAPI event loop from
+fan-out from many tenants — once vLLM is saturated the additional
+backpressure is mostly ergonomic (avoid arbitrarily many in-flight HTTP
+requests from one process).
+
+### Failure modes
+
+- **`EngineStateDB.inference_proxy_url is None` when sample arrives.** The
+  engine hasn't built vLLM yet (no `create_model` happened, or last delete
+  triggered teardown). Mark the future failed with a clear message
+  ("inference engine not ready"); the SDK will surface this back to the
+  client.
+- **vLLM-side 404 for `model=<model_id>`.** Adapter wasn't loaded (most
+  likely the client never called `save_weights_for_sampler`, or
+  `max_cpu_loras` is too low and vLLM evicted it). Fail the future with the
+  4xx body verbatim. Documented in
+  [`multi_lora_design.mdx`](./multi_lora_design.mdx#vllm-capacity-contract).
+- **Engine restarts vLLM under us.** The engine's
+  `delete_model`-on-last-adapter path does `ray.shutdown()`. Next
+  `create_model` rebuilds. The
+  `EngineStateDB` row is updated by `_publish_engine_state` on rebuild.
+  `BackendForwardingInferenceClient` should refresh its cached proxy_url
+  on connection errors and retry once.
+- **API process crash / restart.** In-flight `asyncio.create_task`s are
+  lost; their `FutureDB` rows remain `pending`. The
+  existing 5-minute `retrieve_future` timeout reaps them as failures. Same
+  failure mode as the existing external-vLLM path.
+- **EngineStateDB row out of sync with running vLLM.** Mostly mitigated by
+  the engine writing the row inside `_create_new_inference_client` (single
+  source of truth). If the engine crashes after writing but before vLLM
+  comes up, the API will get connection refused → fail the future. The
+  user will retry.
+
+### Testing
+
+- **Unit:** `BackendForwardingInferenceClient.call_and_store_result` against
+  a `FastAPI` stub vLLM (mirroring
+  `tests/backends/skyrl_train/inference_servers/test_remote_inference_client.py`).
+  Verifies model-name plumbing, error handling, semaphore.
+- **Integration (GPU-gated):** extend
+  `tests/tinker/test_multi_lora_rl_two_clients.md` — add a step
+  comparing wall-clock for two `rl_loop` clients sampling vs. a baseline
+  with this routing disabled. The expected delta is "training step time
+  doesn't appear in sample latency."
+- **Negative:** when `external_inference_client is None` (colocated case),
+  sample requests still flow through the engine. Existing test_api.py
+  coverage should pass unchanged.
+
+### Trade-offs vs. dual-loop-in-engine
+
+We considered three alternatives. Why this design wins:
+
+| Alternative | Why we're not picking it |
+|---|---|
+| **Futures + single engine loop.** `process_sample` becomes non-blocking; engine main loop polls in-flight tasks at the top of each tick. | The 100 ms tick is still a hard latency floor for completion. A 30 s `forward_backward` blocks all sample-completion handling for 30 s. Fixing that means making the engine's main loop async — at which point we're 80 % of the way to a full async refactor for less benefit. |
+| **Separate sample loop inside the engine subprocess.** Spawn a second asyncio task in the engine that handles SAMPLE requests; the main loop handles everything else. | Strictly weaker isolation than process-level: a long Ray-side `ray.get` in the main loop releases the GIL but still constrains the engine's overall throughput. SQLite write contention is fine but introduces a per-loop ownership rule we'd need to maintain. |
+| **Full async refactor of `process_pending_requests`.** Each request type as its own async task with per-model locks. | Big diff; backend methods are sync (Ray `ray.get`) and would need executor wrappers. Premature for what we need (parallel sample-vs-training), without a clear next-step requirement that justifies it. |
+
+Routing through EXTERNAL inherits process-level isolation for free
+(API and engine are separate OS processes; one cannot block the other),
+keeps the engine code untouched, and reuses already-validated machinery.
+
+## Implementation order
+
+1. `EngineStateDB` + migration (auto-create via `SQLModel.metadata.create_all`
+   in lifespan, same as other tables).
+2. `_publish_engine_state` in `SkyRLTrainBackend` (called after
+   `_create_new_inference_client` and on `delete_model` teardown).
+3. `BackendForwardingInferenceClient` + unit tests.
+4. Lifespan wiring in `api.py` + colocate detection.
+5. Update RL smoke runbook to verify parallelism with a simple wall-clock
+   comparison. Document the operator-side requirements
+   (`max_concurrent_samples`, `max_cpu_loras`).
+6. Optional follow-up: have the engine also write `is_colocated` to
+   `EngineStateDB` so the API can resolve it without parsing
+   `backend_config`.
+
+## Out of scope (explicit non-goals)
+
+- **Parallelising training requests across model_ids.** This design
+  parallelises sample only. Multi-tenant `forward_backward` /
+  `optim_step` still serialise through the engine's main loop. Adding
+  per-model parallelism on the training side requires a different design
+  (per-model worker pools or per-model dispatch queues) and isn't
+  needed for the multi-LoRA RL workloads we're targeting.
+- **Auto-recovery from vLLM eviction.** If `max_cpu_loras` is set too
+  low and vLLM evicts an adapter mid-run, the next sample 404s and the
+  future fails. Re-loading from disk on demand is a follow-up — for now
+  the operator sizes `max_cpu_loras` ≥ expected concurrent adapters.
+- **Colocated mode acceleration.** Colocated mode keeps the synchronous
+  engine flow as today. There is no parallelism to exploit there (vLLM is
+  asleep during training).
+- **Sample retries.** `BackendForwardingInferenceClient` does at most one
+  proxy_url refresh + retry on connection errors. Application-level retry
+  is the SDK's job (and is already implemented in `tinker/retry_handler`).
+
+## References
+
+- Multi-LoRA design: [`multi_lora_design.mdx`](./multi_lora_design.mdx).
+- PR #1579 (foundation): [NovaSky-AI/SkyRL#1579](https://github.com/NovaSky-AI/SkyRL/pull/1579).
+- Existing external path: `skyrl/tinker/extra/external_inference.py`,
+  `skyrl/tinker/api.py:1057-1062`,
+  `skyrl/tinker/engine.py:431` (EXTERNAL exclusion).
+- RL smoke runbook: [`test_multi_lora_rl_two_clients.md`](../../../tests/tinker/test_multi_lora_rl_two_clients.md).

From 57a474ab20f6a8b680d80fdc3b0ae31c3c0c8352 Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Wed, 6 May 2026 01:38:53 +0000
Subject: [PATCH 20/21] [smoke logs] Snapshot rl_loop / sl_loop runs from
 manual smoke tests

Captures /tmp/rl_loop_{a..i}.log and /tmp/sl_loop_{a..d}.log from the
multi-LoRA RL + SFT smoke runs into tests/tinker/smoke_logs/. Each run
contributes code.diff (the working-tree diff at launch), config.json,
logs.log (full stdout/stderr), and metrics.jsonl.

Force-added because *.log is in the project .gitignore. ~1.2 MB total;
useful as reference output for the runbooks at
tests/tinker/test_multi_lora_{rl_two_clients,smoke_two_clients}.md.

Co-Authored-By: Eric Tang <erictang000@gmail.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .../tinker/smoke_logs/rl_loop_a.log/code.diff |    3 +
 .../smoke_logs/rl_loop_a.log/config.json      |   12 +
 .../tinker/smoke_logs/rl_loop_a.log/logs.log  |  301 +++++
 .../smoke_logs/rl_loop_a.log/metrics.jsonl    |    6 +
 .../tinker/smoke_logs/rl_loop_b.log/code.diff |    3 +
 .../smoke_logs/rl_loop_b.log/config.json      |   12 +
 .../tinker/smoke_logs/rl_loop_b.log/logs.log  |  106 ++
 .../smoke_logs/rl_loop_b.log/metrics.jsonl    |    4 +
 .../tinker/smoke_logs/rl_loop_c.log/code.diff |    3 +
 .../smoke_logs/rl_loop_c.log/config.json      |   12 +
 .../tinker/smoke_logs/rl_loop_c.log/logs.log  |  683 +++++++++++
 .../smoke_logs/rl_loop_c.log/metrics.jsonl    |    7 +
 .../tinker/smoke_logs/rl_loop_d.log/code.diff |    3 +
 .../smoke_logs/rl_loop_d.log/config.json      |   12 +
 .../tinker/smoke_logs/rl_loop_d.log/logs.log  | 1055 +++++++++++++++++
 .../smoke_logs/rl_loop_d.log/metrics.jsonl    |    4 +
 .../rl_loop_e.log/checkpoints.jsonl           |    3 +
 .../tinker/smoke_logs/rl_loop_e.log/code.diff |    3 +
 .../smoke_logs/rl_loop_e.log/config.json      |   12 +
 .../tinker/smoke_logs/rl_loop_e.log/logs.log  |  825 +++++++++++++
 .../smoke_logs/rl_loop_e.log/metrics.jsonl    |   58 +
 .../rl_loop_f.log/checkpoints.jsonl           |    3 +
 .../tinker/smoke_logs/rl_loop_f.log/code.diff |    3 +
 .../smoke_logs/rl_loop_f.log/config.json      |   12 +
 .../tinker/smoke_logs/rl_loop_f.log/logs.log  |  825 +++++++++++++
 .../smoke_logs/rl_loop_f.log/metrics.jsonl    |   58 +
 .../rl_loop_g.log/checkpoints.jsonl           |    3 +
 .../tinker/smoke_logs/rl_loop_g.log/code.diff |    3 +
 .../smoke_logs/rl_loop_g.log/config.json      |   12 +
 .../tinker/smoke_logs/rl_loop_g.log/logs.log  |  825 +++++++++++++
 .../smoke_logs/rl_loop_g.log/metrics.jsonl    |   58 +
 .../rl_loop_h.log/checkpoints.jsonl           |    1 +
 .../tinker/smoke_logs/rl_loop_h.log/code.diff |    3 +
 .../smoke_logs/rl_loop_h.log/config.json      |   12 +
 .../tinker/smoke_logs/rl_loop_h.log/logs.log  |  430 +++++++
 .../smoke_logs/rl_loop_h.log/metrics.jsonl    |   30 +
 .../rl_loop_i.log/checkpoints.jsonl           |    1 +
 .../tinker/smoke_logs/rl_loop_i.log/code.diff |    3 +
 .../smoke_logs/rl_loop_i.log/config.json      |   12 +
 .../tinker/smoke_logs/rl_loop_i.log/logs.log  |  430 +++++++
 .../smoke_logs/rl_loop_i.log/metrics.jsonl    |   30 +
 .../sl_loop_a.log/checkpoints.jsonl           |    2 +
 .../tinker/smoke_logs/sl_loop_a.log/code.diff |    3 +
 .../smoke_logs/sl_loop_a.log/config.json      |   12 +
 .../tinker/smoke_logs/sl_loop_a.log/logs.log  |  729 ++++++++++++
 .../smoke_logs/sl_loop_a.log/metrics.jsonl    |   47 +
 .../sl_loop_b.log/checkpoints.jsonl           |    2 +
 .../tinker/smoke_logs/sl_loop_b.log/code.diff |    3 +
 .../smoke_logs/sl_loop_b.log/config.json      |   12 +
 .../tinker/smoke_logs/sl_loop_b.log/logs.log  |  674 +++++++++++
 .../smoke_logs/sl_loop_b.log/metrics.jsonl    |   44 +
 .../sl_loop_c.log/checkpoints.jsonl           |    2 +
 .../tinker/smoke_logs/sl_loop_c.log/code.diff |    3 +
 .../smoke_logs/sl_loop_c.log/config.json      |   12 +
 .../tinker/smoke_logs/sl_loop_c.log/logs.log  |  674 +++++++++++
 .../smoke_logs/sl_loop_c.log/metrics.jsonl    |   44 +
 .../sl_loop_d.log/checkpoints.jsonl           |    2 +
 .../tinker/smoke_logs/sl_loop_d.log/code.diff |    3 +
 .../smoke_logs/sl_loop_d.log/config.json      |   12 +
 .../tinker/smoke_logs/sl_loop_d.log/logs.log  |  644 ++++++++++
 .../smoke_logs/sl_loop_d.log/metrics.jsonl    |   42 +
 61 files changed, 8847 insertions(+)
 create mode 100644 tests/tinker/smoke_logs/rl_loop_a.log/code.diff
 create mode 100644 tests/tinker/smoke_logs/rl_loop_a.log/config.json
 create mode 100644 tests/tinker/smoke_logs/rl_loop_a.log/logs.log
 create mode 100644 tests/tinker/smoke_logs/rl_loop_a.log/metrics.jsonl
 create mode 100644 tests/tinker/smoke_logs/rl_loop_b.log/code.diff
 create mode 100644 tests/tinker/smoke_logs/rl_loop_b.log/config.json
 create mode 100644 tests/tinker/smoke_logs/rl_loop_b.log/logs.log
 create mode 100644 tests/tinker/smoke_logs/rl_loop_b.log/metrics.jsonl
 create mode 100644 tests/tinker/smoke_logs/rl_loop_c.log/code.diff
 create mode 100644 tests/tinker/smoke_logs/rl_loop_c.log/config.json
 create mode 100644 tests/tinker/smoke_logs/rl_loop_c.log/logs.log
 create mode 100644 tests/tinker/smoke_logs/rl_loop_c.log/metrics.jsonl
 create mode 100644 tests/tinker/smoke_logs/rl_loop_d.log/code.diff
 create mode 100644 tests/tinker/smoke_logs/rl_loop_d.log/config.json
 create mode 100644 tests/tinker/smoke_logs/rl_loop_d.log/logs.log
 create mode 100644 tests/tinker/smoke_logs/rl_loop_d.log/metrics.jsonl
 create mode 100644 tests/tinker/smoke_logs/rl_loop_e.log/checkpoints.jsonl
 create mode 100644 tests/tinker/smoke_logs/rl_loop_e.log/code.diff
 create mode 100644 tests/tinker/smoke_logs/rl_loop_e.log/config.json
 create mode 100644 tests/tinker/smoke_logs/rl_loop_e.log/logs.log
 create mode 100644 tests/tinker/smoke_logs/rl_loop_e.log/metrics.jsonl
 create mode 100644 tests/tinker/smoke_logs/rl_loop_f.log/checkpoints.jsonl
 create mode 100644 tests/tinker/smoke_logs/rl_loop_f.log/code.diff
 create mode 100644 tests/tinker/smoke_logs/rl_loop_f.log/config.json
 create mode 100644 tests/tinker/smoke_logs/rl_loop_f.log/logs.log
 create mode 100644 tests/tinker/smoke_logs/rl_loop_f.log/metrics.jsonl
 create mode 100644 tests/tinker/smoke_logs/rl_loop_g.log/checkpoints.jsonl
 create mode 100644 tests/tinker/smoke_logs/rl_loop_g.log/code.diff
 create mode 100644 tests/tinker/smoke_logs/rl_loop_g.log/config.json
 create mode 100644 tests/tinker/smoke_logs/rl_loop_g.log/logs.log
 create mode 100644 tests/tinker/smoke_logs/rl_loop_g.log/metrics.jsonl
 create mode 100644 tests/tinker/smoke_logs/rl_loop_h.log/checkpoints.jsonl
 create mode 100644 tests/tinker/smoke_logs/rl_loop_h.log/code.diff
 create mode 100644 tests/tinker/smoke_logs/rl_loop_h.log/config.json
 create mode 100644 tests/tinker/smoke_logs/rl_loop_h.log/logs.log
 create mode 100644 tests/tinker/smoke_logs/rl_loop_h.log/metrics.jsonl
 create mode 100644 tests/tinker/smoke_logs/rl_loop_i.log/checkpoints.jsonl
 create mode 100644 tests/tinker/smoke_logs/rl_loop_i.log/code.diff
 create mode 100644 tests/tinker/smoke_logs/rl_loop_i.log/config.json
 create mode 100644 tests/tinker/smoke_logs/rl_loop_i.log/logs.log
 create mode 100644 tests/tinker/smoke_logs/rl_loop_i.log/metrics.jsonl
 create mode 100644 tests/tinker/smoke_logs/sl_loop_a.log/checkpoints.jsonl
 create mode 100644 tests/tinker/smoke_logs/sl_loop_a.log/code.diff
 create mode 100644 tests/tinker/smoke_logs/sl_loop_a.log/config.json
 create mode 100644 tests/tinker/smoke_logs/sl_loop_a.log/logs.log
 create mode 100644 tests/tinker/smoke_logs/sl_loop_a.log/metrics.jsonl
 create mode 100644 tests/tinker/smoke_logs/sl_loop_b.log/checkpoints.jsonl
 create mode 100644 tests/tinker/smoke_logs/sl_loop_b.log/code.diff
 create mode 100644 tests/tinker/smoke_logs/sl_loop_b.log/config.json
 create mode 100644 tests/tinker/smoke_logs/sl_loop_b.log/logs.log
 create mode 100644 tests/tinker/smoke_logs/sl_loop_b.log/metrics.jsonl
 create mode 100644 tests/tinker/smoke_logs/sl_loop_c.log/checkpoints.jsonl
 create mode 100644 tests/tinker/smoke_logs/sl_loop_c.log/code.diff
 create mode 100644 tests/tinker/smoke_logs/sl_loop_c.log/config.json
 create mode 100644 tests/tinker/smoke_logs/sl_loop_c.log/logs.log
 create mode 100644 tests/tinker/smoke_logs/sl_loop_c.log/metrics.jsonl
 create mode 100644 tests/tinker/smoke_logs/sl_loop_d.log/checkpoints.jsonl
 create mode 100644 tests/tinker/smoke_logs/sl_loop_d.log/code.diff
 create mode 100644 tests/tinker/smoke_logs/sl_loop_d.log/config.json
 create mode 100644 tests/tinker/smoke_logs/sl_loop_d.log/logs.log
 create mode 100644 tests/tinker/smoke_logs/sl_loop_d.log/metrics.jsonl

diff --git a/tests/tinker/smoke_logs/rl_loop_a.log/code.diff b/tests/tinker/smoke_logs/rl_loop_a.log/code.diff
new file mode 100644
index 0000000000..d8e559755f
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_a.log/code.diff
@@ -0,0 +1,3 @@
+### repo: /home/etang/tinker-cookbook @ dacb835b31fb3f7a012851a1bc64c950d0495de3
+modules: tinker_cookbook
+(no local changes)
diff --git a/tests/tinker/smoke_logs/rl_loop_a.log/config.json b/tests/tinker/smoke_logs/rl_loop_a.log/config.json
new file mode 100644
index 0000000000..f1d0c32b6a
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_a.log/config.json
@@ -0,0 +1,12 @@
+{
+  "base_url": "http://localhost:8000",
+  "log_path": "/tmp/rl_loop_a.log",
+  "model_name": "Qwen/Qwen3-0.6B",
+  "batch_size": 128,
+  "group_size": 16,
+  "learning_rate": 4e-05,
+  "lora_rank": 32,
+  "save_every": 20,
+  "max_tokens": 256,
+  "ttl_seconds": 604800
+}
\ No newline at end of file
diff --git a/tests/tinker/smoke_logs/rl_loop_a.log/logs.log b/tests/tinker/smoke_logs/rl_loop_a.log/logs.log
new file mode 100644
index 0000000000..b654726695
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_a.log/logs.log
@@ -0,0 +1,301 @@
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/rl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B lora_rank=32 log_path=/tmp/rl_loop_a.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/rl_loop_a.log
+__main__:73 [INFO] Using renderer: qwen3
+__main__:76 [INFO] Loading dataset...
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/rl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B lora_rank=32 log_path=/tmp/rl_loop_a.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/rl_loop_a.log
+__main__:73 [INFO] Using renderer: qwen3
+__main__:76 [INFO] Loading dataset...
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/rl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B lora_rank=32 log_path=/tmp/rl_loop_a.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/rl_loop_a.log
+__main__:73 [INFO] Using renderer: qwen3
+__main__:76 [INFO] Loading dataset...
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_1677eba7
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/rl_loop_a.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/rl_loop_a.log
+tinker.lib.public_interfaces.service_client:159 [INFO] TrainingClient initialized for model model_8babd19c
+__main__:121 [INFO] Training for 58 batches
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 0                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 0           │
+│ progress/done_frac     │ 0.017241    │
+│ reward/total           │ 0.021484    │
+│ skyrl.ai/grad_norm     │ 2767.572763 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 220.875473  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 1                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 1           │
+│ progress/done_frac     │ 0.034483    │
+│ reward/total           │ 0.026855    │
+│ skyrl.ai/grad_norm     │ 2880.009288 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 193.037447  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 2                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value      ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040   │
+│ progress/batch         │ 2          │
+│ progress/done_frac     │ 0.051724   │
+│ reward/total           │ 0.032715   │
+│ skyrl.ai/grad_norm     │ 766.103412 │
+│ skyrl.ai/learning_rate │ 0.000040   │
+│ time/total             │ 381.700626 │
+└────────────────────────┴────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 3                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 3           │
+│ progress/done_frac     │ 0.068966    │
+│ reward/total           │ 0.069824    │
+│ skyrl.ai/grad_norm     │ 4700.790572 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 451.961692  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 4                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 4           │
+│ progress/done_frac     │ 0.086207    │
+│ reward/total           │ 0.188965    │
+│ skyrl.ai/grad_norm     │ 7912.660235 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 518.589346  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 5                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 5           │
+│ progress/done_frac     │ 0.103448    │
+│ reward/total           │ 0.306152    │
+│ skyrl.ai/grad_norm     │ 9853.311322 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 651.922414  │
+└────────────────────────┴─────────────┘
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1845' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1865' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1855' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1882' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1856' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1859' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1836' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1837' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1883' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1841' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1873' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1864' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1900' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1942' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1885' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1955' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1918' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1886' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1902' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1899' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1947' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1910' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1889' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1949' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1906' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1871' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1880' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1911' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1921' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1879' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1909' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1891' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1887' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1934' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1877' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1903' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1919' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1922' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1958' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1833' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1954' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1924' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1961' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1948' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1869' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1932' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1897' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1914' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1927' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1894' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1888' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1933' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1884' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1920' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1905' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1895' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1938' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1890' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1957' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1931' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1893' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1834' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1960' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1936' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1875' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1835' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1848' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1876' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1850' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1839' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1926' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1858' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1935' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1862' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1959' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1940' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1915' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1939' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1878' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1950' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1892' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1872' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1945' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1930' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1908' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1946' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1832' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1928' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1953' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1944' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1896' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1854' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1901' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1849' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1868' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1881' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1904' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1851' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1857' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1866' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1838' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1916' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1925' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1937' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1917' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1941' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_1677eba7
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1913' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1943' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1912' and expected type self.model_cls=<class 'tinker.SampleResponse'>
diff --git a/tests/tinker/smoke_logs/rl_loop_a.log/metrics.jsonl b/tests/tinker/smoke_logs/rl_loop_a.log/metrics.jsonl
new file mode 100644
index 0000000000..7720167023
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_a.log/metrics.jsonl
@@ -0,0 +1,6 @@
+{"step": 0, "progress/batch": 0, "optim/lr": 4e-05, "progress/done_frac": 0.017241379310344827, "skyrl.ai/grad_norm": 2767.572763271094, "skyrl.ai/learning_rate": 4e-05, "time/total": 220.87547278404236, "reward/total": 0.021484375}
+{"step": 1, "progress/batch": 1, "optim/lr": 4e-05, "progress/done_frac": 0.034482758620689655, "skyrl.ai/grad_norm": 2880.009288179467, "skyrl.ai/learning_rate": 4e-05, "time/total": 193.037446975708, "reward/total": 0.02685546875}
+{"step": 2, "progress/batch": 2, "optim/lr": 4e-05, "progress/done_frac": 0.05172413793103448, "skyrl.ai/grad_norm": 766.1034117532698, "skyrl.ai/learning_rate": 4e-05, "time/total": 381.7006256580353, "reward/total": 0.03271484375}
+{"step": 3, "progress/batch": 3, "optim/lr": 4e-05, "progress/done_frac": 0.06896551724137931, "skyrl.ai/grad_norm": 4700.790571808108, "skyrl.ai/learning_rate": 4e-05, "time/total": 451.96169209480286, "reward/total": 0.06982421875}
+{"step": 4, "progress/batch": 4, "optim/lr": 4e-05, "progress/done_frac": 0.08620689655172414, "skyrl.ai/grad_norm": 7912.660235344369, "skyrl.ai/learning_rate": 4e-05, "time/total": 518.5893461704254, "reward/total": 0.18896484375}
+{"step": 5, "progress/batch": 5, "optim/lr": 4e-05, "progress/done_frac": 0.10344827586206896, "skyrl.ai/grad_norm": 9853.311321581186, "skyrl.ai/learning_rate": 4e-05, "time/total": 651.9224138259888, "reward/total": 0.30615234375}
diff --git a/tests/tinker/smoke_logs/rl_loop_b.log/code.diff b/tests/tinker/smoke_logs/rl_loop_b.log/code.diff
new file mode 100644
index 0000000000..d8e559755f
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_b.log/code.diff
@@ -0,0 +1,3 @@
+### repo: /home/etang/tinker-cookbook @ dacb835b31fb3f7a012851a1bc64c950d0495de3
+modules: tinker_cookbook
+(no local changes)
diff --git a/tests/tinker/smoke_logs/rl_loop_b.log/config.json b/tests/tinker/smoke_logs/rl_loop_b.log/config.json
new file mode 100644
index 0000000000..d02ea077f0
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_b.log/config.json
@@ -0,0 +1,12 @@
+{
+  "base_url": "http://localhost:8000",
+  "log_path": "/tmp/rl_loop_b.log",
+  "model_name": "Qwen/Qwen3-0.6B",
+  "batch_size": 128,
+  "group_size": 16,
+  "learning_rate": 4e-05,
+  "lora_rank": 32,
+  "save_every": 20,
+  "max_tokens": 256,
+  "ttl_seconds": 604800
+}
\ No newline at end of file
diff --git a/tests/tinker/smoke_logs/rl_loop_b.log/logs.log b/tests/tinker/smoke_logs/rl_loop_b.log/logs.log
new file mode 100644
index 0000000000..dc6db425c1
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_b.log/logs.log
@@ -0,0 +1,106 @@
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/rl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B lora_rank=32 log_path=/tmp/rl_loop_b.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/rl_loop_b.log
+__main__:73 [INFO] Using renderer: qwen3
+__main__:76 [INFO] Loading dataset...
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_ff83e78a
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/rl_loop_b.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/rl_loop_b.log
+tinker.lib.public_interfaces.service_client:159 [INFO] TrainingClient initialized for model model_19dca9e4
+__main__:121 [INFO] Training for 58 batches
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 0                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 0           │
+│ progress/done_frac     │ 0.017241    │
+│ reward/total           │ 0.016602    │
+│ skyrl.ai/grad_norm     │ 2784.973608 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 329.021244  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 1                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 1           │
+│ progress/done_frac     │ 0.034483    │
+│ reward/total           │ 0.030273    │
+│ skyrl.ai/grad_norm     │ 3172.651888 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 294.953841  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 2                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 2           │
+│ progress/done_frac     │ 0.051724    │
+│ reward/total           │ 0.029297    │
+│ skyrl.ai/grad_norm     │ 2699.619881 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 407.538405  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 3                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 3           │
+│ progress/done_frac     │ 0.068966    │
+│ reward/total           │ 0.054199    │
+│ skyrl.ai/grad_norm     │ 4655.835049 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 686.296854  │
+└────────────────────────┴─────────────┘
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1860' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1772' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1815' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1820' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1758' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1756' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1773' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1790' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1771' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1778' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1768' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1805' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1767' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1812' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1766' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1802' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1808' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1816' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1813' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_ff83e78a
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1829' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1780' and expected type self.model_cls=<class 'tinker.SampleResponse'>
+tinker.lib.retry_handler:241 [ERROR] Request failed with non-retryable error: ValueError: Error retrieving result: Error code: 400 - {'detail': "Expected exactly one model_id for sampling, got {'model_8babd19c', 'model_19dca9e4'}"} with status code e.status_code=400 for self.request_id='1763' and expected type self.model_cls=<class 'tinker.SampleResponse'>
diff --git a/tests/tinker/smoke_logs/rl_loop_b.log/metrics.jsonl b/tests/tinker/smoke_logs/rl_loop_b.log/metrics.jsonl
new file mode 100644
index 0000000000..ffc1adafab
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_b.log/metrics.jsonl
@@ -0,0 +1,4 @@
+{"step": 0, "progress/batch": 0, "optim/lr": 4e-05, "progress/done_frac": 0.017241379310344827, "skyrl.ai/grad_norm": 2784.973608492547, "skyrl.ai/learning_rate": 4e-05, "time/total": 329.02124428749084, "reward/total": 0.0166015625}
+{"step": 1, "progress/batch": 1, "optim/lr": 4e-05, "progress/done_frac": 0.034482758620689655, "skyrl.ai/grad_norm": 3172.6518876170453, "skyrl.ai/learning_rate": 4e-05, "time/total": 294.9538412094116, "reward/total": 0.0302734375}
+{"step": 2, "progress/batch": 2, "optim/lr": 4e-05, "progress/done_frac": 0.05172413793103448, "skyrl.ai/grad_norm": 2699.6198806498664, "skyrl.ai/learning_rate": 4e-05, "time/total": 407.53840494155884, "reward/total": 0.029296875}
+{"step": 3, "progress/batch": 3, "optim/lr": 4e-05, "progress/done_frac": 0.06896551724137931, "skyrl.ai/grad_norm": 4655.835048624468, "skyrl.ai/learning_rate": 4e-05, "time/total": 686.2968535423279, "reward/total": 0.05419921875}
diff --git a/tests/tinker/smoke_logs/rl_loop_c.log/code.diff b/tests/tinker/smoke_logs/rl_loop_c.log/code.diff
new file mode 100644
index 0000000000..d8e559755f
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_c.log/code.diff
@@ -0,0 +1,3 @@
+### repo: /home/etang/tinker-cookbook @ dacb835b31fb3f7a012851a1bc64c950d0495de3
+modules: tinker_cookbook
+(no local changes)
diff --git a/tests/tinker/smoke_logs/rl_loop_c.log/config.json b/tests/tinker/smoke_logs/rl_loop_c.log/config.json
new file mode 100644
index 0000000000..681c18071a
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_c.log/config.json
@@ -0,0 +1,12 @@
+{
+  "base_url": "http://localhost:8000",
+  "log_path": "/tmp/rl_loop_c.log",
+  "model_name": "Qwen/Qwen3-0.6B",
+  "batch_size": 128,
+  "group_size": 16,
+  "learning_rate": 4e-05,
+  "lora_rank": 32,
+  "save_every": 20,
+  "max_tokens": 256,
+  "ttl_seconds": 604800
+}
\ No newline at end of file
diff --git a/tests/tinker/smoke_logs/rl_loop_c.log/logs.log b/tests/tinker/smoke_logs/rl_loop_c.log/logs.log
new file mode 100644
index 0000000000..089d66411c
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_c.log/logs.log
@@ -0,0 +1,683 @@
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/rl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B lora_rank=32 log_path=/tmp/rl_loop_c.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/rl_loop_c.log
+__main__:73 [INFO] Using renderer: qwen3
+__main__:76 [INFO] Loading dataset...
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_060f07d1
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/rl_loop_c.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/rl_loop_c.log
+tinker.lib.public_interfaces.service_client:159 [INFO] TrainingClient initialized for model model_01d93e11
+__main__:121 [INFO] Training for 58 batches
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_060f07d1
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_060f07d1
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/rl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B lora_rank=32 log_path=/tmp/rl_loop_c.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/rl_loop_c.log
+__main__:73 [INFO] Using renderer: qwen3
+__main__:76 [INFO] Loading dataset...
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_d224d066
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/rl_loop_c.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/rl_loop_c.log
+tinker.lib.public_interfaces.service_client:159 [INFO] TrainingClient initialized for model model_c0ac2cf5
+__main__:121 [INFO] Training for 58 batches
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_d224d066
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_d224d066
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/rl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B lora_rank=32 log_path=/tmp/rl_loop_c.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/rl_loop_c.log
+__main__:73 [INFO] Using renderer: qwen3
+__main__:76 [INFO] Loading dataset...
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_7a047dd4
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/rl_loop_c.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/rl_loop_c.log
+tinker.lib.public_interfaces.service_client:159 [INFO] TrainingClient initialized for model model_18c1fcad
+__main__:121 [INFO] Training for 58 batches
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/rl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B lora_rank=32 log_path=/tmp/rl_loop_c.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/rl_loop_c.log
+__main__:73 [INFO] Using renderer: qwen3
+__main__:76 [INFO] Loading dataset...
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_cb2206f9
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/rl_loop_c.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/rl_loop_c.log
+tinker.lib.public_interfaces.service_client:159 [INFO] TrainingClient initialized for model model_4022cc73
+__main__:121 [INFO] Training for 58 batches
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 0                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 0           │
+│ progress/done_frac     │ 0.017241    │
+│ reward/total           │ 0.019043    │
+│ skyrl.ai/grad_norm     │ 2515.858104 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 133.940457  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 1                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 1           │
+│ progress/done_frac     │ 0.034483    │
+│ reward/total           │ 0.023438    │
+│ skyrl.ai/grad_norm     │ 2593.005592 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 80.892934   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 2                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 2           │
+│ progress/done_frac     │ 0.051724    │
+│ reward/total           │ 0.033691    │
+│ skyrl.ai/grad_norm     │ 1459.361847 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 83.688621   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 3                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 3           │
+│ progress/done_frac     │ 0.068966    │
+│ reward/total           │ 0.069824    │
+│ skyrl.ai/grad_norm     │ 3230.374127 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 146.584503  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 4                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 4           │
+│ progress/done_frac     │ 0.086207    │
+│ reward/total           │ 0.225098    │
+│ skyrl.ai/grad_norm     │ 8136.650171 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 157.939979  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 5                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 5           │
+│ progress/done_frac     │ 0.103448    │
+│ reward/total           │ 0.364746    │
+│ skyrl.ai/grad_norm     │ 8247.074148 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 155.924101  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 6                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 6           │
+│ progress/done_frac     │ 0.120690    │
+│ reward/total           │ 0.416016    │
+│ skyrl.ai/grad_norm     │ 8501.772521 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 150.344107  │
+└────────────────────────┴─────────────┘
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 120.04752013500183 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 140.05585919200166 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 160.06073126800038 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 103, in handle_async_request
+    return await self._connection.handle_async_request(request)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 136, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 106, in handle_async_request
+    ) = await self._receive_response_headers(**kwargs)
+        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 177, in _receive_response_headers
+    event = await self._receive_event(timeout=timeout)
+            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 217, in _receive_event
+    data = await self._network_stream.read(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 32, in read
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 180.0704045420025 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 200.08335031200113 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 220.0932662330015 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 240.10563371600074 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 260.113305881001 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 103, in handle_async_request
+    return await self._connection.handle_async_request(request)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 136, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 106, in handle_async_request
+    ) = await self._receive_response_headers(**kwargs)
+        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 177, in _receive_response_headers
+    event = await self._receive_event(timeout=timeout)
+            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 217, in _receive_event
+    data = await self._network_stream.read(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 32, in read
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 280.11982140799955 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 300.12539942100193 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 320.1320038060003 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 340.1399088580001 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 360.1474978940023 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 380.15622666499985 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 103, in handle_async_request
+    return await self._connection.handle_async_request(request)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 136, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 106, in handle_async_request
+    ) = await self._receive_response_headers(**kwargs)
+        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 177, in _receive_response_headers
+    event = await self._receive_event(timeout=timeout)
+            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 217, in _receive_event
+    data = await self._network_stream.read(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 32, in read
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 400.1645123180024 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 420.1727330680005 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 440.1820644870022 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 460.1859169199997 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 480.1972260930015 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 500.20116487500127 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 103, in handle_async_request
+    return await self._connection.handle_async_request(request)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 136, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 106, in handle_async_request
+    ) = await self._receive_response_headers(**kwargs)
+        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 177, in _receive_response_headers
+    event = await self._receive_event(timeout=timeout)
+            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 217, in _receive_event
+    data = await self._network_stream.read(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 32, in read
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 520.2125360690006 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 540.2188946360002 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 560.2234918930008 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 580.2347657700011 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 600.2407169200014 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 103, in handle_async_request
+    return await self._connection.handle_async_request(request)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 136, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 106, in handle_async_request
+    ) = await self._receive_response_headers(**kwargs)
+        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 177, in _receive_response_headers
+    event = await self._receive_event(timeout=timeout)
+            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 217, in _receive_event
+    data = await self._network_stream.read(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 32, in read
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 620.2457274400003 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 640.2519691980015 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 660.2582203860002 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 680.2657900810009 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 700.2808281979997 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 720.289861448 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 103, in handle_async_request
+    return await self._connection.handle_async_request(request)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 136, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 106, in handle_async_request
+    ) = await self._receive_response_headers(**kwargs)
+        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 177, in _receive_response_headers
+    event = await self._receive_event(timeout=timeout)
+            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 217, in _receive_event
+    data = await self._network_stream.read(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 32, in read
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 740.2932220139992 seconds for session session_cb2206f9. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_cb2206f9
diff --git a/tests/tinker/smoke_logs/rl_loop_c.log/metrics.jsonl b/tests/tinker/smoke_logs/rl_loop_c.log/metrics.jsonl
new file mode 100644
index 0000000000..9962cdc017
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_c.log/metrics.jsonl
@@ -0,0 +1,7 @@
+{"step": 0, "progress/batch": 0, "optim/lr": 4e-05, "progress/done_frac": 0.017241379310344827, "skyrl.ai/grad_norm": 2515.8581041068273, "skyrl.ai/learning_rate": 4e-05, "time/total": 133.9404571056366, "reward/total": 0.01904296875}
+{"step": 1, "progress/batch": 1, "optim/lr": 4e-05, "progress/done_frac": 0.034482758620689655, "skyrl.ai/grad_norm": 2593.0055919723736, "skyrl.ai/learning_rate": 4e-05, "time/total": 80.8929340839386, "reward/total": 0.0234375}
+{"step": 2, "progress/batch": 2, "optim/lr": 4e-05, "progress/done_frac": 0.05172413793103448, "skyrl.ai/grad_norm": 1459.361846835801, "skyrl.ai/learning_rate": 4e-05, "time/total": 83.68862080574036, "reward/total": 0.03369140625}
+{"step": 3, "progress/batch": 3, "optim/lr": 4e-05, "progress/done_frac": 0.06896551724137931, "skyrl.ai/grad_norm": 3230.3741269394786, "skyrl.ai/learning_rate": 4e-05, "time/total": 146.5845034122467, "reward/total": 0.06982421875}
+{"step": 4, "progress/batch": 4, "optim/lr": 4e-05, "progress/done_frac": 0.08620689655172414, "skyrl.ai/grad_norm": 8136.650170678348, "skyrl.ai/learning_rate": 4e-05, "time/total": 157.93997859954834, "reward/total": 0.22509765625}
+{"step": 5, "progress/batch": 5, "optim/lr": 4e-05, "progress/done_frac": 0.10344827586206896, "skyrl.ai/grad_norm": 8247.074147841766, "skyrl.ai/learning_rate": 4e-05, "time/total": 155.9241008758545, "reward/total": 0.36474609375}
+{"step": 6, "progress/batch": 6, "optim/lr": 4e-05, "progress/done_frac": 0.1206896551724138, "skyrl.ai/grad_norm": 8501.772521068768, "skyrl.ai/learning_rate": 4e-05, "time/total": 150.34410667419434, "reward/total": 0.416015625}
diff --git a/tests/tinker/smoke_logs/rl_loop_d.log/code.diff b/tests/tinker/smoke_logs/rl_loop_d.log/code.diff
new file mode 100644
index 0000000000..d8e559755f
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_d.log/code.diff
@@ -0,0 +1,3 @@
+### repo: /home/etang/tinker-cookbook @ dacb835b31fb3f7a012851a1bc64c950d0495de3
+modules: tinker_cookbook
+(no local changes)
diff --git a/tests/tinker/smoke_logs/rl_loop_d.log/config.json b/tests/tinker/smoke_logs/rl_loop_d.log/config.json
new file mode 100644
index 0000000000..1a2c902c7f
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_d.log/config.json
@@ -0,0 +1,12 @@
+{
+  "base_url": "http://localhost:8000",
+  "log_path": "/tmp/rl_loop_d.log",
+  "model_name": "Qwen/Qwen3-0.6B",
+  "batch_size": 128,
+  "group_size": 16,
+  "learning_rate": 4e-05,
+  "lora_rank": 32,
+  "save_every": 20,
+  "max_tokens": 256,
+  "ttl_seconds": 604800
+}
\ No newline at end of file
diff --git a/tests/tinker/smoke_logs/rl_loop_d.log/logs.log b/tests/tinker/smoke_logs/rl_loop_d.log/logs.log
new file mode 100644
index 0000000000..3bfd003249
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_d.log/logs.log
@@ -0,0 +1,1055 @@
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/rl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B lora_rank=32 log_path=/tmp/rl_loop_d.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/rl_loop_d.log
+__main__:73 [INFO] Using renderer: qwen3
+__main__:76 [INFO] Loading dataset...
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_ce1e79ca
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/rl_loop_d.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/rl_loop_d.log
+tinker.lib.public_interfaces.service_client:159 [INFO] TrainingClient initialized for model model_050364f7
+__main__:121 [INFO] Training for 58 batches
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 0                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 0           │
+│ progress/done_frac     │ 0.017241    │
+│ reward/total           │ 0.017578    │
+│ skyrl.ai/grad_norm     │ 2725.635981 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 90.791685   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 1                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 1           │
+│ progress/done_frac     │ 0.034483    │
+│ reward/total           │ 0.028809    │
+│ skyrl.ai/grad_norm     │ 2989.493268 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 214.999288  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 2                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 2           │
+│ progress/done_frac     │ 0.051724    │
+│ reward/total           │ 0.032227    │
+│ skyrl.ai/grad_norm     │ 3538.728444 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 89.495047   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 3                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 3           │
+│ progress/done_frac     │ 0.068966    │
+│ reward/total           │ 0.058594    │
+│ skyrl.ai/grad_norm     │ 4845.256856 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 211.909505  │
+└────────────────────────┴─────────────┘
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 120.04485723900143 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 140.04900060500222 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 160.05557763099932 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 103, in handle_async_request
+    return await self._connection.handle_async_request(request)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 136, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 106, in handle_async_request
+    ) = await self._receive_response_headers(**kwargs)
+        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 177, in _receive_response_headers
+    event = await self._receive_event(timeout=timeout)
+            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 217, in _receive_event
+    data = await self._network_stream.read(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 32, in read
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 180.0642175590001 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 200.07351759200174 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 220.0846148849996 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 240.09335962399928 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 260.10152930299955 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 280.1058633040011 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 103, in handle_async_request
+    return await self._connection.handle_async_request(request)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 136, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 106, in handle_async_request
+    ) = await self._receive_response_headers(**kwargs)
+        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 177, in _receive_response_headers
+    event = await self._receive_event(timeout=timeout)
+            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 217, in _receive_event
+    data = await self._network_stream.read(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 32, in read
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 300.1137273230015 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 320.12278748899917 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 340.12942929400015 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 360.1370587860001 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 380.14563508799984 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 103, in handle_async_request
+    return await self._connection.handle_async_request(request)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 136, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 106, in handle_async_request
+    ) = await self._receive_response_headers(**kwargs)
+        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 177, in _receive_response_headers
+    event = await self._receive_event(timeout=timeout)
+            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 217, in _receive_event
+    data = await self._network_stream.read(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 32, in read
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 400.15754274700157 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 420.16348117300004 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 440.1740379190014 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 460.18030684100086 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 480.1916362060001 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 500.2016032320025 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 103, in handle_async_request
+    return await self._connection.handle_async_request(request)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 136, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 106, in handle_async_request
+    ) = await self._receive_response_headers(**kwargs)
+        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 177, in _receive_response_headers
+    event = await self._receive_event(timeout=timeout)
+            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 217, in _receive_event
+    data = await self._network_stream.read(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 32, in read
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 520.2061943459994 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 540.2176004899993 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 560.2313318000015 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 580.2389547450002 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 600.251811164002 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 620.255916858001 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 103, in handle_async_request
+    return await self._connection.handle_async_request(request)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 136, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 106, in handle_async_request
+    ) = await self._receive_response_headers(**kwargs)
+        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 177, in _receive_response_headers
+    event = await self._receive_event(timeout=timeout)
+            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 217, in _receive_event
+    data = await self._network_stream.read(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 32, in read
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 640.2620931620004 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 660.2726048479999 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 680.2792337780011 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 700.2885933140024 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 720.2977807000025 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 103, in handle_async_request
+    return await self._connection.handle_async_request(request)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 136, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 106, in handle_async_request
+    ) = await self._receive_response_headers(**kwargs)
+        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 177, in _receive_response_headers
+    event = await self._receive_event(timeout=timeout)
+            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 217, in _receive_event
+    data = await self._network_stream.read(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 32, in read
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 740.306528338002 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 760.3118081380017 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 780.3204979139991 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 800.3248156610025 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 820.3413785789999 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 840.349503499001 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 103, in handle_async_request
+    return await self._connection.handle_async_request(request)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 136, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 106, in handle_async_request
+    ) = await self._receive_response_headers(**kwargs)
+        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 177, in _receive_response_headers
+    event = await self._receive_event(timeout=timeout)
+            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/http11.py", line 217, in _receive_event
+    data = await self._network_stream.read(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 32, in read
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ReadTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 860.3567000069997 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 880.3628695219995 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 900.373514021001 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 920.3820001849999 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 940.3942476160009 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 960.3973198070016 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 101, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 78, in handle_async_request
+    stream = await self._connect(request)
+             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 124, in _connect
+    stream = await self._network_backend.connect_tcp(**kwargs)
+             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/auto.py", line 31, in connect_tcp
+    return await self._backend.connect_tcp(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 113, in connect_tcp
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ConnectTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ConnectTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 980.5901912549998 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1000.5961541690012 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1020.5987187960018 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1040.6084509050015 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1060.6115265260014 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 101, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 78, in handle_async_request
+    stream = await self._connect(request)
+             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 124, in _connect
+    stream = await self._network_backend.connect_tcp(**kwargs)
+             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/auto.py", line 31, in connect_tcp
+    return await self._backend.connect_tcp(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 113, in connect_tcp
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ConnectTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ConnectTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1080.623052070001 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1100.6259051609995 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1120.6313976990023 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1140.6395486559995 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1160.6467008580003 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1180.657208861001 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 101, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 78, in handle_async_request
+    stream = await self._connect(request)
+             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 124, in _connect
+    stream = await self._network_backend.connect_tcp(**kwargs)
+             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/auto.py", line 31, in connect_tcp
+    return await self._backend.connect_tcp(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 113, in connect_tcp
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ConnectTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ConnectTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1200.6636957090013 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1220.6689744399991 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1240.672793824 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1260.6809952330004 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1280.682395210999 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1300.6857222889994 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
+tinker.lib.telemetry:135 [WARNING] Failed to send telemetry batch
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 101, in map_httpcore_exceptions
+    yield
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 394, in handle_async_request
+    resp = await self._pool.handle_async_request(req)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 256, in handle_async_request
+    raise exc from None
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection_pool.py", line 236, in handle_async_request
+    response = await connection.handle_async_request(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 101, in handle_async_request
+    raise exc
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 78, in handle_async_request
+    stream = await self._connect(request)
+             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_async/connection.py", line 124, in _connect
+    stream = await self._network_backend.connect_tcp(**kwargs)
+             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/auto.py", line 31, in connect_tcp
+    return await self._backend.connect_tcp(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_backends/anyio.py", line 113, in connect_tcp
+    with map_exceptions(exc_map):
+         ^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpcore/_exceptions.py", line 14, in map_exceptions
+    raise to_exc(exc) from exc
+httpcore.ConnectTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 969, in request
+    response = await self._client.send(
+               ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1629, in send
+    response = await self._send_handling_auth(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1657, in _send_handling_auth
+    response = await self._send_handling_redirects(
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1694, in _send_handling_redirects
+    response = await self._send_single_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_client.py", line 1730, in _send_single_request
+    response = await transport.handle_async_request(request)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 393, in handle_async_request
+    with map_httpcore_exceptions():
+         ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 158, in __exit__
+    self.gen.throw(value)
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
+    raise mapped_exc(message) from exc
+httpx.ConnectTimeout
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 133, in _send_batch_with_retry
+    return await self._send_batch(batch)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/lib/telemetry.py", line 142, in _send_batch
+    return await client.telemetry.send(request=request, timeout=HTTP_TIMEOUT_SECONDS)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/resources/telemetry.py", line 68, in send
+    return await self._post(
+           ^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 1232, in post
+    return await self.request(cast_to, opts, stream=stream, stream_cls=stream_cls)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/etang/tinker-cookbook/.venv/lib/python3.12/site-packages/tinker/_base_client.py", line 987, in request
+    raise APITimeoutError(request=request) from err
+tinker.APITimeoutError: Request timed out.
+tinker.lib.internal_client_holder:357 [WARNING] Session heartbeat failed for 1320.6910064229996 seconds for session session_ce1e79ca. Last exception: APITimeoutError: Request timed out..
+Your connection may be unreliable or Tinker is down. If this persists, the session will be terminated.
diff --git a/tests/tinker/smoke_logs/rl_loop_d.log/metrics.jsonl b/tests/tinker/smoke_logs/rl_loop_d.log/metrics.jsonl
new file mode 100644
index 0000000000..7e5d600922
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_d.log/metrics.jsonl
@@ -0,0 +1,4 @@
+{"step": 0, "progress/batch": 0, "optim/lr": 4e-05, "progress/done_frac": 0.017241379310344827, "skyrl.ai/grad_norm": 2725.6359808308957, "skyrl.ai/learning_rate": 4e-05, "time/total": 90.79168486595154, "reward/total": 0.017578125}
+{"step": 1, "progress/batch": 1, "optim/lr": 4e-05, "progress/done_frac": 0.034482758620689655, "skyrl.ai/grad_norm": 2989.493268097455, "skyrl.ai/learning_rate": 4e-05, "time/total": 214.99928832054138, "reward/total": 0.02880859375}
+{"step": 2, "progress/batch": 2, "optim/lr": 4e-05, "progress/done_frac": 0.05172413793103448, "skyrl.ai/grad_norm": 3538.7284439470627, "skyrl.ai/learning_rate": 4e-05, "time/total": 89.49504661560059, "reward/total": 0.0322265625}
+{"step": 3, "progress/batch": 3, "optim/lr": 4e-05, "progress/done_frac": 0.06896551724137931, "skyrl.ai/grad_norm": 4845.256855936535, "skyrl.ai/learning_rate": 4e-05, "time/total": 211.90950536727905, "reward/total": 0.05859375}
diff --git a/tests/tinker/smoke_logs/rl_loop_e.log/checkpoints.jsonl b/tests/tinker/smoke_logs/rl_loop_e.log/checkpoints.jsonl
new file mode 100644
index 0000000000..a362ee607a
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_e.log/checkpoints.jsonl
@@ -0,0 +1,3 @@
+{"name": "000020", "batch": 20, "state_path": "tinker://model_e3750434/weights/000020"}
+{"name": "000040", "batch": 40, "state_path": "tinker://model_e3750434/weights/000040"}
+{"name": "final", "batch": 58, "state_path": "tinker://model_e3750434/weights/final", "sampler_path": "tinker://model_e3750434/final"}
diff --git a/tests/tinker/smoke_logs/rl_loop_e.log/code.diff b/tests/tinker/smoke_logs/rl_loop_e.log/code.diff
new file mode 100644
index 0000000000..d8e559755f
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_e.log/code.diff
@@ -0,0 +1,3 @@
+### repo: /home/etang/tinker-cookbook @ dacb835b31fb3f7a012851a1bc64c950d0495de3
+modules: tinker_cookbook
+(no local changes)
diff --git a/tests/tinker/smoke_logs/rl_loop_e.log/config.json b/tests/tinker/smoke_logs/rl_loop_e.log/config.json
new file mode 100644
index 0000000000..f0e86c83b2
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_e.log/config.json
@@ -0,0 +1,12 @@
+{
+  "base_url": "http://localhost:8000",
+  "log_path": "/tmp/rl_loop_e.log",
+  "model_name": "Qwen/Qwen3-0.6B",
+  "batch_size": 128,
+  "group_size": 16,
+  "learning_rate": 4e-05,
+  "lora_rank": 32,
+  "save_every": 20,
+  "max_tokens": 256,
+  "ttl_seconds": 604800
+}
\ No newline at end of file
diff --git a/tests/tinker/smoke_logs/rl_loop_e.log/logs.log b/tests/tinker/smoke_logs/rl_loop_e.log/logs.log
new file mode 100644
index 0000000000..aa73ca12ac
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_e.log/logs.log
@@ -0,0 +1,825 @@
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/rl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B lora_rank=32 log_path=/tmp/rl_loop_e.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/rl_loop_e.log
+__main__:73 [INFO] Using renderer: qwen3
+__main__:76 [INFO] Loading dataset...
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_997689e1
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/rl_loop_e.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/rl_loop_e.log
+tinker.lib.public_interfaces.service_client:159 [INFO] TrainingClient initialized for model model_e3750434
+__main__:121 [INFO] Training for 58 batches
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 0                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 0           │
+│ progress/done_frac     │ 0.017241    │
+│ reward/total           │ 0.025879    │
+│ skyrl.ai/grad_norm     │ 3311.342930 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 134.581530  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 1                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 1           │
+│ progress/done_frac     │ 0.034483    │
+│ reward/total           │ 0.029785    │
+│ skyrl.ai/grad_norm     │ 2744.224025 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 80.350298   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 2                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 2           │
+│ progress/done_frac     │ 0.051724    │
+│ reward/total           │ 0.034180    │
+│ skyrl.ai/grad_norm     │ 2827.500663 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 82.265805   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 3                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 3           │
+│ progress/done_frac     │ 0.068966    │
+│ reward/total           │ 0.078125    │
+│ skyrl.ai/grad_norm     │ 3308.343997 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 86.257145   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 4                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 4           │
+│ progress/done_frac     │ 0.086207    │
+│ reward/total           │ 0.223145    │
+│ skyrl.ai/grad_norm     │ 7099.649851 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 90.150057   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 5                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 5           │
+│ progress/done_frac     │ 0.103448    │
+│ reward/total           │ 0.415527    │
+│ skyrl.ai/grad_norm     │ 7196.960470 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 88.091412   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 6                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 6           │
+│ progress/done_frac     │ 0.120690    │
+│ reward/total           │ 0.461426    │
+│ skyrl.ai/grad_norm     │ 6807.531124 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 84.784650   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 7                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 7           │
+│ progress/done_frac     │ 0.137931    │
+│ reward/total           │ 0.568359    │
+│ skyrl.ai/grad_norm     │ 4628.584665 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 75.168792   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 8                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 8           │
+│ progress/done_frac     │ 0.155172    │
+│ reward/total           │ 0.513672    │
+│ skyrl.ai/grad_norm     │ 5137.686055 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 76.193191   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 9                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 9           │
+│ progress/done_frac     │ 0.172414    │
+│ reward/total           │ 0.592773    │
+│ skyrl.ai/grad_norm     │ 4486.921885 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 69.551138   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 10                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 10          │
+│ progress/done_frac     │ 0.189655    │
+│ reward/total           │ 0.585449    │
+│ skyrl.ai/grad_norm     │ 4914.633048 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 72.504134   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 11                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 11          │
+│ progress/done_frac     │ 0.206897    │
+│ reward/total           │ 0.625488    │
+│ skyrl.ai/grad_norm     │ 5440.769798 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 65.498356   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 12                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 12          │
+│ progress/done_frac     │ 0.224138    │
+│ reward/total           │ 0.579102    │
+│ skyrl.ai/grad_norm     │ 4611.391981 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 67.348524   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 13                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 13          │
+│ progress/done_frac     │ 0.241379    │
+│ reward/total           │ 0.641113    │
+│ skyrl.ai/grad_norm     │ 4010.038030 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 65.521300   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 14                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 14          │
+│ progress/done_frac     │ 0.258621    │
+│ reward/total           │ 0.570312    │
+│ skyrl.ai/grad_norm     │ 4132.876964 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 67.548914   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 15                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 15          │
+│ progress/done_frac     │ 0.275862    │
+│ reward/total           │ 0.646484    │
+│ skyrl.ai/grad_norm     │ 4244.424343 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.138198   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 16                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 16          │
+│ progress/done_frac     │ 0.293103    │
+│ reward/total           │ 0.668945    │
+│ skyrl.ai/grad_norm     │ 4716.073367 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 133.361657  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 17                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 17          │
+│ progress/done_frac     │ 0.310345    │
+│ reward/total           │ 0.670898    │
+│ skyrl.ai/grad_norm     │ 5110.910682 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 128.878171  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 18                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 18          │
+│ progress/done_frac     │ 0.327586    │
+│ reward/total           │ 0.620605    │
+│ skyrl.ai/grad_norm     │ 5311.529347 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 138.159210  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 19                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 19          │
+│ progress/done_frac     │ 0.344828    │
+│ reward/total           │ 0.658203    │
+│ skyrl.ai/grad_norm     │ 5211.462367 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 140.631799  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_e3750434/weights/000020'}
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 20                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 20          │
+│ progress/done_frac     │ 0.362069    │
+│ reward/total           │ 0.682617    │
+│ skyrl.ai/grad_norm     │ 4430.771942 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 143.243473  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 21                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 21          │
+│ progress/done_frac     │ 0.379310    │
+│ reward/total           │ 0.674316    │
+│ skyrl.ai/grad_norm     │ 4389.845783 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 142.351490  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 22                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 22          │
+│ progress/done_frac     │ 0.396552    │
+│ reward/total           │ 0.713867    │
+│ skyrl.ai/grad_norm     │ 4029.614498 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 138.315871  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 23                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 23          │
+│ progress/done_frac     │ 0.413793    │
+│ reward/total           │ 0.703613    │
+│ skyrl.ai/grad_norm     │ 4087.451406 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 129.937644  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 24                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 24          │
+│ progress/done_frac     │ 0.431034    │
+│ reward/total           │ 0.617188    │
+│ skyrl.ai/grad_norm     │ 4925.083959 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 124.777653  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 25                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 25          │
+│ progress/done_frac     │ 0.448276    │
+│ reward/total           │ 0.687012    │
+│ skyrl.ai/grad_norm     │ 4435.276767 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 116.290545  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 26                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 26          │
+│ progress/done_frac     │ 0.465517    │
+│ reward/total           │ 0.677246    │
+│ skyrl.ai/grad_norm     │ 5004.585298 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 123.367723  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 27                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 27          │
+│ progress/done_frac     │ 0.482759    │
+│ reward/total           │ 0.720703    │
+│ skyrl.ai/grad_norm     │ 4235.038134 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 114.187049  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 28                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 28          │
+│ progress/done_frac     │ 0.500000    │
+│ reward/total           │ 0.719727    │
+│ skyrl.ai/grad_norm     │ 5780.208647 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 95.931007   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 29                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 29          │
+│ progress/done_frac     │ 0.517241    │
+│ reward/total           │ 0.746094    │
+│ skyrl.ai/grad_norm     │ 4432.479216 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 121.491157  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 30                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 30          │
+│ progress/done_frac     │ 0.534483    │
+│ reward/total           │ 0.702637    │
+│ skyrl.ai/grad_norm     │ 5579.194566 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 92.008166   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 31                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 31          │
+│ progress/done_frac     │ 0.551724    │
+│ reward/total           │ 0.751953    │
+│ skyrl.ai/grad_norm     │ 4846.482436 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 103.272193  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 32                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 32          │
+│ progress/done_frac     │ 0.568966    │
+│ reward/total           │ 0.641602    │
+│ skyrl.ai/grad_norm     │ 4850.217315 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 130.145584  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 33                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 33          │
+│ progress/done_frac     │ 0.586207    │
+│ reward/total           │ 0.753418    │
+│ skyrl.ai/grad_norm     │ 4359.924999 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 109.510066  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 34                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 34          │
+│ progress/done_frac     │ 0.603448    │
+│ reward/total           │ 0.680176    │
+│ skyrl.ai/grad_norm     │ 5077.019401 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 115.822206  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 35                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 35          │
+│ progress/done_frac     │ 0.620690    │
+│ reward/total           │ 0.706055    │
+│ skyrl.ai/grad_norm     │ 5074.235312 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 111.309374  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 36                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 36          │
+│ progress/done_frac     │ 0.637931    │
+│ reward/total           │ 0.678711    │
+│ skyrl.ai/grad_norm     │ 4478.695122 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 112.462799  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 37                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 37          │
+│ progress/done_frac     │ 0.655172    │
+│ reward/total           │ 0.702148    │
+│ skyrl.ai/grad_norm     │ 4052.576464 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 116.594501  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 38                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 38          │
+│ progress/done_frac     │ 0.672414    │
+│ reward/total           │ 0.701172    │
+│ skyrl.ai/grad_norm     │ 4067.377288 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 116.648988  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 39                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 39          │
+│ progress/done_frac     │ 0.689655    │
+│ reward/total           │ 0.726562    │
+│ skyrl.ai/grad_norm     │ 5586.897887 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 119.899999  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_e3750434/weights/000040'}
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 40                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 40          │
+│ progress/done_frac     │ 0.706897    │
+│ reward/total           │ 0.740234    │
+│ skyrl.ai/grad_norm     │ 5081.892561 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 122.908623  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 41                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 41          │
+│ progress/done_frac     │ 0.724138    │
+│ reward/total           │ 0.717285    │
+│ skyrl.ai/grad_norm     │ 5443.625263 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 114.647553  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 42                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 42          │
+│ progress/done_frac     │ 0.741379    │
+│ reward/total           │ 0.703613    │
+│ skyrl.ai/grad_norm     │ 5651.898088 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 114.402829  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 43                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 43          │
+│ progress/done_frac     │ 0.758621    │
+│ reward/total           │ 0.668457    │
+│ skyrl.ai/grad_norm     │ 5342.144326 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 108.515467  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 44                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 44          │
+│ progress/done_frac     │ 0.775862    │
+│ reward/total           │ 0.812988    │
+│ skyrl.ai/grad_norm     │ 5312.453859 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 106.809338  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 45                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 45          │
+│ progress/done_frac     │ 0.793103    │
+│ reward/total           │ 0.781738    │
+│ skyrl.ai/grad_norm     │ 5327.349247 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 108.259785  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 46                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 46          │
+│ progress/done_frac     │ 0.810345    │
+│ reward/total           │ 0.700684    │
+│ skyrl.ai/grad_norm     │ 3723.012221 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 114.133801  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 47                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 47          │
+│ progress/done_frac     │ 0.827586    │
+│ reward/total           │ 0.788574    │
+│ skyrl.ai/grad_norm     │ 4958.986187 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 110.261018  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 48                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 48          │
+│ progress/done_frac     │ 0.844828    │
+│ reward/total           │ 0.752441    │
+│ skyrl.ai/grad_norm     │ 4091.877320 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 132.814460  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 49                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 49          │
+│ progress/done_frac     │ 0.862069    │
+│ reward/total           │ 0.705078    │
+│ skyrl.ai/grad_norm     │ 4194.601531 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 112.915834  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 50                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 50          │
+│ progress/done_frac     │ 0.879310    │
+│ reward/total           │ 0.699219    │
+│ skyrl.ai/grad_norm     │ 4843.082902 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 127.590147  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 51                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 51          │
+│ progress/done_frac     │ 0.896552    │
+│ reward/total           │ 0.704590    │
+│ skyrl.ai/grad_norm     │ 4349.735854 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 108.636021  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 52                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 52          │
+│ progress/done_frac     │ 0.913793    │
+│ reward/total           │ 0.672852    │
+│ skyrl.ai/grad_norm     │ 5007.795124 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 95.163089   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 53                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 53          │
+│ progress/done_frac     │ 0.931034    │
+│ reward/total           │ 0.820801    │
+│ skyrl.ai/grad_norm     │ 3999.667236 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 96.708356   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 54                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 54          │
+│ progress/done_frac     │ 0.948276    │
+│ reward/total           │ 0.700195    │
+│ skyrl.ai/grad_norm     │ 4468.568451 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 123.548607  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 55                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 55          │
+│ progress/done_frac     │ 0.965517    │
+│ reward/total           │ 0.781738    │
+│ skyrl.ai/grad_norm     │ 5002.112954 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 113.389676  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 56                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 56          │
+│ progress/done_frac     │ 0.982759    │
+│ reward/total           │ 0.769043    │
+│ skyrl.ai/grad_norm     │ 5835.695674 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 104.909214  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_e.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 57                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 57          │
+│ progress/done_frac     │ 1.000000    │
+│ reward/total           │ 0.734863    │
+│ skyrl.ai/grad_norm     │ 4346.515386 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 120.041633  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_e3750434/weights/final', 'sampler_path': 'tinker://model_e3750434/final'}
+__main__:257 [INFO] Training completed
diff --git a/tests/tinker/smoke_logs/rl_loop_e.log/metrics.jsonl b/tests/tinker/smoke_logs/rl_loop_e.log/metrics.jsonl
new file mode 100644
index 0000000000..3d8f4565a3
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_e.log/metrics.jsonl
@@ -0,0 +1,58 @@
+{"step": 0, "progress/batch": 0, "optim/lr": 4e-05, "progress/done_frac": 0.017241379310344827, "skyrl.ai/grad_norm": 3311.3429299907916, "skyrl.ai/learning_rate": 4e-05, "time/total": 134.58153009414673, "reward/total": 0.02587890625}
+{"step": 1, "progress/batch": 1, "optim/lr": 4e-05, "progress/done_frac": 0.034482758620689655, "skyrl.ai/grad_norm": 2744.2240251116527, "skyrl.ai/learning_rate": 4e-05, "time/total": 80.35029816627502, "reward/total": 0.02978515625}
+{"step": 2, "progress/batch": 2, "optim/lr": 4e-05, "progress/done_frac": 0.05172413793103448, "skyrl.ai/grad_norm": 2827.5006631298957, "skyrl.ai/learning_rate": 4e-05, "time/total": 82.26580548286438, "reward/total": 0.0341796875}
+{"step": 3, "progress/batch": 3, "optim/lr": 4e-05, "progress/done_frac": 0.06896551724137931, "skyrl.ai/grad_norm": 3308.343996624293, "skyrl.ai/learning_rate": 4e-05, "time/total": 86.25714540481567, "reward/total": 0.078125}
+{"step": 4, "progress/batch": 4, "optim/lr": 4e-05, "progress/done_frac": 0.08620689655172414, "skyrl.ai/grad_norm": 7099.649850520799, "skyrl.ai/learning_rate": 4e-05, "time/total": 90.15005731582642, "reward/total": 0.22314453125}
+{"step": 5, "progress/batch": 5, "optim/lr": 4e-05, "progress/done_frac": 0.10344827586206896, "skyrl.ai/grad_norm": 7196.960469531565, "skyrl.ai/learning_rate": 4e-05, "time/total": 88.09141182899475, "reward/total": 0.41552734375}
+{"step": 6, "progress/batch": 6, "optim/lr": 4e-05, "progress/done_frac": 0.1206896551724138, "skyrl.ai/grad_norm": 6807.5311236894095, "skyrl.ai/learning_rate": 4e-05, "time/total": 84.78464961051941, "reward/total": 0.46142578125}
+{"step": 7, "progress/batch": 7, "optim/lr": 4e-05, "progress/done_frac": 0.13793103448275862, "skyrl.ai/grad_norm": 4628.5846648840725, "skyrl.ai/learning_rate": 4e-05, "time/total": 75.16879153251648, "reward/total": 0.568359375}
+{"step": 8, "progress/batch": 8, "optim/lr": 4e-05, "progress/done_frac": 0.15517241379310345, "skyrl.ai/grad_norm": 5137.68605502516, "skyrl.ai/learning_rate": 4e-05, "time/total": 76.19319105148315, "reward/total": 0.513671875}
+{"step": 9, "progress/batch": 9, "optim/lr": 4e-05, "progress/done_frac": 0.1724137931034483, "skyrl.ai/grad_norm": 4486.9218847668835, "skyrl.ai/learning_rate": 4e-05, "time/total": 69.5511384010315, "reward/total": 0.5927734375}
+{"step": 10, "progress/batch": 10, "optim/lr": 4e-05, "progress/done_frac": 0.1896551724137931, "skyrl.ai/grad_norm": 4914.6330483567135, "skyrl.ai/learning_rate": 4e-05, "time/total": 72.50413417816162, "reward/total": 0.58544921875}
+{"step": 11, "progress/batch": 11, "optim/lr": 4e-05, "progress/done_frac": 0.20689655172413793, "skyrl.ai/grad_norm": 5440.769798475212, "skyrl.ai/learning_rate": 4e-05, "time/total": 65.49835634231567, "reward/total": 0.62548828125}
+{"step": 12, "progress/batch": 12, "optim/lr": 4e-05, "progress/done_frac": 0.22413793103448276, "skyrl.ai/grad_norm": 4611.391980736403, "skyrl.ai/learning_rate": 4e-05, "time/total": 67.34852361679077, "reward/total": 0.5791015625}
+{"step": 13, "progress/batch": 13, "optim/lr": 4e-05, "progress/done_frac": 0.2413793103448276, "skyrl.ai/grad_norm": 4010.038029744855, "skyrl.ai/learning_rate": 4e-05, "time/total": 65.52129983901978, "reward/total": 0.64111328125}
+{"step": 14, "progress/batch": 14, "optim/lr": 4e-05, "progress/done_frac": 0.25862068965517243, "skyrl.ai/grad_norm": 4132.876964053007, "skyrl.ai/learning_rate": 4e-05, "time/total": 67.54891395568848, "reward/total": 0.5703125}
+{"step": 15, "progress/batch": 15, "optim/lr": 4e-05, "progress/done_frac": 0.27586206896551724, "skyrl.ai/grad_norm": 4244.424342593469, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.1381983757019, "reward/total": 0.646484375}
+{"step": 16, "progress/batch": 16, "optim/lr": 4e-05, "progress/done_frac": 0.29310344827586204, "skyrl.ai/grad_norm": 4716.073366689709, "skyrl.ai/learning_rate": 4e-05, "time/total": 133.36165714263916, "reward/total": 0.6689453125}
+{"step": 17, "progress/batch": 17, "optim/lr": 4e-05, "progress/done_frac": 0.3103448275862069, "skyrl.ai/grad_norm": 5110.9106820604875, "skyrl.ai/learning_rate": 4e-05, "time/total": 128.87817096710205, "reward/total": 0.6708984375}
+{"step": 18, "progress/batch": 18, "optim/lr": 4e-05, "progress/done_frac": 0.3275862068965517, "skyrl.ai/grad_norm": 5311.5293466194835, "skyrl.ai/learning_rate": 4e-05, "time/total": 138.1592104434967, "reward/total": 0.62060546875}
+{"step": 19, "progress/batch": 19, "optim/lr": 4e-05, "progress/done_frac": 0.3448275862068966, "skyrl.ai/grad_norm": 5211.462366745058, "skyrl.ai/learning_rate": 4e-05, "time/total": 140.63179898262024, "reward/total": 0.658203125}
+{"step": 20, "progress/batch": 20, "optim/lr": 4e-05, "progress/done_frac": 0.3620689655172414, "skyrl.ai/grad_norm": 4430.771941772675, "skyrl.ai/learning_rate": 4e-05, "time/total": 143.24347305297852, "reward/total": 0.6826171875}
+{"step": 21, "progress/batch": 21, "optim/lr": 4e-05, "progress/done_frac": 0.3793103448275862, "skyrl.ai/grad_norm": 4389.845783168243, "skyrl.ai/learning_rate": 4e-05, "time/total": 142.35149002075195, "reward/total": 0.67431640625}
+{"step": 22, "progress/batch": 22, "optim/lr": 4e-05, "progress/done_frac": 0.39655172413793105, "skyrl.ai/grad_norm": 4029.614497690815, "skyrl.ai/learning_rate": 4e-05, "time/total": 138.31587147712708, "reward/total": 0.7138671875}
+{"step": 23, "progress/batch": 23, "optim/lr": 4e-05, "progress/done_frac": 0.41379310344827586, "skyrl.ai/grad_norm": 4087.4514064389805, "skyrl.ai/learning_rate": 4e-05, "time/total": 129.9376437664032, "reward/total": 0.70361328125}
+{"step": 24, "progress/batch": 24, "optim/lr": 4e-05, "progress/done_frac": 0.43103448275862066, "skyrl.ai/grad_norm": 4925.083958675223, "skyrl.ai/learning_rate": 4e-05, "time/total": 124.7776529788971, "reward/total": 0.6171875}
+{"step": 25, "progress/batch": 25, "optim/lr": 4e-05, "progress/done_frac": 0.4482758620689655, "skyrl.ai/grad_norm": 4435.276767012404, "skyrl.ai/learning_rate": 4e-05, "time/total": 116.29054522514343, "reward/total": 0.68701171875}
+{"step": 26, "progress/batch": 26, "optim/lr": 4e-05, "progress/done_frac": 0.46551724137931033, "skyrl.ai/grad_norm": 5004.585297504679, "skyrl.ai/learning_rate": 4e-05, "time/total": 123.36772322654724, "reward/total": 0.67724609375}
+{"step": 27, "progress/batch": 27, "optim/lr": 4e-05, "progress/done_frac": 0.4827586206896552, "skyrl.ai/grad_norm": 4235.038134420987, "skyrl.ai/learning_rate": 4e-05, "time/total": 114.1870493888855, "reward/total": 0.720703125}
+{"step": 28, "progress/batch": 28, "optim/lr": 4e-05, "progress/done_frac": 0.5, "skyrl.ai/grad_norm": 5780.20864675316, "skyrl.ai/learning_rate": 4e-05, "time/total": 95.93100690841675, "reward/total": 0.7197265625}
+{"step": 29, "progress/batch": 29, "optim/lr": 4e-05, "progress/done_frac": 0.5172413793103449, "skyrl.ai/grad_norm": 4432.479215969321, "skyrl.ai/learning_rate": 4e-05, "time/total": 121.4911572933197, "reward/total": 0.74609375}
+{"step": 30, "progress/batch": 30, "optim/lr": 4e-05, "progress/done_frac": 0.5344827586206896, "skyrl.ai/grad_norm": 5579.194565526462, "skyrl.ai/learning_rate": 4e-05, "time/total": 92.00816607475281, "reward/total": 0.70263671875}
+{"step": 31, "progress/batch": 31, "optim/lr": 4e-05, "progress/done_frac": 0.5517241379310345, "skyrl.ai/grad_norm": 4846.4824357465695, "skyrl.ai/learning_rate": 4e-05, "time/total": 103.27219343185425, "reward/total": 0.751953125}
+{"step": 32, "progress/batch": 32, "optim/lr": 4e-05, "progress/done_frac": 0.5689655172413793, "skyrl.ai/grad_norm": 4850.2173147190015, "skyrl.ai/learning_rate": 4e-05, "time/total": 130.14558362960815, "reward/total": 0.6416015625}
+{"step": 33, "progress/batch": 33, "optim/lr": 4e-05, "progress/done_frac": 0.5862068965517241, "skyrl.ai/grad_norm": 4359.92499935492, "skyrl.ai/learning_rate": 4e-05, "time/total": 109.51006555557251, "reward/total": 0.75341796875}
+{"step": 34, "progress/batch": 34, "optim/lr": 4e-05, "progress/done_frac": 0.603448275862069, "skyrl.ai/grad_norm": 5077.019401184124, "skyrl.ai/learning_rate": 4e-05, "time/total": 115.82220578193665, "reward/total": 0.68017578125}
+{"step": 35, "progress/batch": 35, "optim/lr": 4e-05, "progress/done_frac": 0.6206896551724138, "skyrl.ai/grad_norm": 5074.23531184749, "skyrl.ai/learning_rate": 4e-05, "time/total": 111.3093740940094, "reward/total": 0.7060546875}
+{"step": 36, "progress/batch": 36, "optim/lr": 4e-05, "progress/done_frac": 0.6379310344827587, "skyrl.ai/grad_norm": 4478.695122465918, "skyrl.ai/learning_rate": 4e-05, "time/total": 112.4627993106842, "reward/total": 0.6787109375}
+{"step": 37, "progress/batch": 37, "optim/lr": 4e-05, "progress/done_frac": 0.6551724137931034, "skyrl.ai/grad_norm": 4052.576464423589, "skyrl.ai/learning_rate": 4e-05, "time/total": 116.59450078010559, "reward/total": 0.7021484375}
+{"step": 38, "progress/batch": 38, "optim/lr": 4e-05, "progress/done_frac": 0.6724137931034483, "skyrl.ai/grad_norm": 4067.377287638805, "skyrl.ai/learning_rate": 4e-05, "time/total": 116.64898777008057, "reward/total": 0.701171875}
+{"step": 39, "progress/batch": 39, "optim/lr": 4e-05, "progress/done_frac": 0.6896551724137931, "skyrl.ai/grad_norm": 5586.897887021026, "skyrl.ai/learning_rate": 4e-05, "time/total": 119.8999993801117, "reward/total": 0.7265625}
+{"step": 40, "progress/batch": 40, "optim/lr": 4e-05, "progress/done_frac": 0.7068965517241379, "skyrl.ai/grad_norm": 5081.892560847779, "skyrl.ai/learning_rate": 4e-05, "time/total": 122.90862274169922, "reward/total": 0.740234375}
+{"step": 41, "progress/batch": 41, "optim/lr": 4e-05, "progress/done_frac": 0.7241379310344828, "skyrl.ai/grad_norm": 5443.625262635186, "skyrl.ai/learning_rate": 4e-05, "time/total": 114.64755296707153, "reward/total": 0.71728515625}
+{"step": 42, "progress/batch": 42, "optim/lr": 4e-05, "progress/done_frac": 0.7413793103448276, "skyrl.ai/grad_norm": 5651.898088253184, "skyrl.ai/learning_rate": 4e-05, "time/total": 114.40282893180847, "reward/total": 0.70361328125}
+{"step": 43, "progress/batch": 43, "optim/lr": 4e-05, "progress/done_frac": 0.7586206896551724, "skyrl.ai/grad_norm": 5342.144326017409, "skyrl.ai/learning_rate": 4e-05, "time/total": 108.51546716690063, "reward/total": 0.66845703125}
+{"step": 44, "progress/batch": 44, "optim/lr": 4e-05, "progress/done_frac": 0.7758620689655172, "skyrl.ai/grad_norm": 5312.45385862315, "skyrl.ai/learning_rate": 4e-05, "time/total": 106.80933833122253, "reward/total": 0.81298828125}
+{"step": 45, "progress/batch": 45, "optim/lr": 4e-05, "progress/done_frac": 0.7931034482758621, "skyrl.ai/grad_norm": 5327.349247045851, "skyrl.ai/learning_rate": 4e-05, "time/total": 108.25978517532349, "reward/total": 0.78173828125}
+{"step": 46, "progress/batch": 46, "optim/lr": 4e-05, "progress/done_frac": 0.8103448275862069, "skyrl.ai/grad_norm": 3723.012221306828, "skyrl.ai/learning_rate": 4e-05, "time/total": 114.1338005065918, "reward/total": 0.70068359375}
+{"step": 47, "progress/batch": 47, "optim/lr": 4e-05, "progress/done_frac": 0.8275862068965517, "skyrl.ai/grad_norm": 4958.986186711957, "skyrl.ai/learning_rate": 4e-05, "time/total": 110.26101756095886, "reward/total": 0.78857421875}
+{"step": 48, "progress/batch": 48, "optim/lr": 4e-05, "progress/done_frac": 0.8448275862068966, "skyrl.ai/grad_norm": 4091.8773197641203, "skyrl.ai/learning_rate": 4e-05, "time/total": 132.81446027755737, "reward/total": 0.75244140625}
+{"step": 49, "progress/batch": 49, "optim/lr": 4e-05, "progress/done_frac": 0.8620689655172413, "skyrl.ai/grad_norm": 4194.601530538986, "skyrl.ai/learning_rate": 4e-05, "time/total": 112.91583442687988, "reward/total": 0.705078125}
+{"step": 50, "progress/batch": 50, "optim/lr": 4e-05, "progress/done_frac": 0.8793103448275862, "skyrl.ai/grad_norm": 4843.082902449637, "skyrl.ai/learning_rate": 4e-05, "time/total": 127.59014749526978, "reward/total": 0.69921875}
+{"step": 51, "progress/batch": 51, "optim/lr": 4e-05, "progress/done_frac": 0.896551724137931, "skyrl.ai/grad_norm": 4349.73585404907, "skyrl.ai/learning_rate": 4e-05, "time/total": 108.63602137565613, "reward/total": 0.70458984375}
+{"step": 52, "progress/batch": 52, "optim/lr": 4e-05, "progress/done_frac": 0.9137931034482759, "skyrl.ai/grad_norm": 5007.795123604798, "skyrl.ai/learning_rate": 4e-05, "time/total": 95.16308879852295, "reward/total": 0.6728515625}
+{"step": 53, "progress/batch": 53, "optim/lr": 4e-05, "progress/done_frac": 0.9310344827586207, "skyrl.ai/grad_norm": 3999.6672361585283, "skyrl.ai/learning_rate": 4e-05, "time/total": 96.70835590362549, "reward/total": 0.82080078125}
+{"step": 54, "progress/batch": 54, "optim/lr": 4e-05, "progress/done_frac": 0.9482758620689655, "skyrl.ai/grad_norm": 4468.568450857612, "skyrl.ai/learning_rate": 4e-05, "time/total": 123.5486068725586, "reward/total": 0.7001953125}
+{"step": 55, "progress/batch": 55, "optim/lr": 4e-05, "progress/done_frac": 0.9655172413793104, "skyrl.ai/grad_norm": 5002.112953542733, "skyrl.ai/learning_rate": 4e-05, "time/total": 113.38967609405518, "reward/total": 0.78173828125}
+{"step": 56, "progress/batch": 56, "optim/lr": 4e-05, "progress/done_frac": 0.9827586206896551, "skyrl.ai/grad_norm": 5835.695674039214, "skyrl.ai/learning_rate": 4e-05, "time/total": 104.90921354293823, "reward/total": 0.76904296875}
+{"step": 57, "progress/batch": 57, "optim/lr": 4e-05, "progress/done_frac": 1.0, "skyrl.ai/grad_norm": 4346.515385915481, "skyrl.ai/learning_rate": 4e-05, "time/total": 120.04163336753845, "reward/total": 0.73486328125}
diff --git a/tests/tinker/smoke_logs/rl_loop_f.log/checkpoints.jsonl b/tests/tinker/smoke_logs/rl_loop_f.log/checkpoints.jsonl
new file mode 100644
index 0000000000..8f232bbf22
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_f.log/checkpoints.jsonl
@@ -0,0 +1,3 @@
+{"name": "000020", "batch": 20, "state_path": "tinker://model_e06832f2/weights/000020"}
+{"name": "000040", "batch": 40, "state_path": "tinker://model_e06832f2/weights/000040"}
+{"name": "final", "batch": 58, "state_path": "tinker://model_e06832f2/weights/final", "sampler_path": "tinker://model_e06832f2/final"}
diff --git a/tests/tinker/smoke_logs/rl_loop_f.log/code.diff b/tests/tinker/smoke_logs/rl_loop_f.log/code.diff
new file mode 100644
index 0000000000..d8e559755f
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_f.log/code.diff
@@ -0,0 +1,3 @@
+### repo: /home/etang/tinker-cookbook @ dacb835b31fb3f7a012851a1bc64c950d0495de3
+modules: tinker_cookbook
+(no local changes)
diff --git a/tests/tinker/smoke_logs/rl_loop_f.log/config.json b/tests/tinker/smoke_logs/rl_loop_f.log/config.json
new file mode 100644
index 0000000000..472eacbbaa
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_f.log/config.json
@@ -0,0 +1,12 @@
+{
+  "base_url": "http://localhost:8000",
+  "log_path": "/tmp/rl_loop_f.log",
+  "model_name": "Qwen/Qwen3-0.6B",
+  "batch_size": 128,
+  "group_size": 16,
+  "learning_rate": 4e-05,
+  "lora_rank": 32,
+  "save_every": 20,
+  "max_tokens": 256,
+  "ttl_seconds": 604800
+}
\ No newline at end of file
diff --git a/tests/tinker/smoke_logs/rl_loop_f.log/logs.log b/tests/tinker/smoke_logs/rl_loop_f.log/logs.log
new file mode 100644
index 0000000000..0ba8aa403f
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_f.log/logs.log
@@ -0,0 +1,825 @@
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/rl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B lora_rank=32 log_path=/tmp/rl_loop_f.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/rl_loop_f.log
+__main__:73 [INFO] Using renderer: qwen3
+__main__:76 [INFO] Loading dataset...
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_8246be4c
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/rl_loop_f.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/rl_loop_f.log
+tinker.lib.public_interfaces.service_client:159 [INFO] TrainingClient initialized for model model_e06832f2
+__main__:121 [INFO] Training for 58 batches
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 0                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 0           │
+│ progress/done_frac     │ 0.017241    │
+│ reward/total           │ 0.016602    │
+│ skyrl.ai/grad_norm     │ 2294.091214 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 95.201286   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 1                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 1           │
+│ progress/done_frac     │ 0.034483    │
+│ reward/total           │ 0.022461    │
+│ skyrl.ai/grad_norm     │ 2692.103174 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 173.874377  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 2                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 2           │
+│ progress/done_frac     │ 0.051724    │
+│ reward/total           │ 0.025879    │
+│ skyrl.ai/grad_norm     │ 2919.905992 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 103.146296  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 3                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 3           │
+│ progress/done_frac     │ 0.068966    │
+│ reward/total           │ 0.050781    │
+│ skyrl.ai/grad_norm     │ 4679.051827 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 138.495987  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 4                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 4           │
+│ progress/done_frac     │ 0.086207    │
+│ reward/total           │ 0.172852    │
+│ skyrl.ai/grad_norm     │ 7131.992989 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 127.667517  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 5                  
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value        ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040     │
+│ progress/batch         │ 5            │
+│ progress/done_frac     │ 0.103448     │
+│ reward/total           │ 0.315430     │
+│ skyrl.ai/grad_norm     │ 10840.603673 │
+│ skyrl.ai/learning_rate │ 0.000040     │
+│ time/total             │ 142.253343   │
+└────────────────────────┴──────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 6                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 6           │
+│ progress/done_frac     │ 0.120690    │
+│ reward/total           │ 0.411621    │
+│ skyrl.ai/grad_norm     │ 8993.776960 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 138.389953  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 7                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 7           │
+│ progress/done_frac     │ 0.137931    │
+│ reward/total           │ 0.560059    │
+│ skyrl.ai/grad_norm     │ 6986.811004 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 129.658682  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 8                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 8           │
+│ progress/done_frac     │ 0.155172    │
+│ reward/total           │ 0.538086    │
+│ skyrl.ai/grad_norm     │ 5574.307491 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 125.289493  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 9                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 9           │
+│ progress/done_frac     │ 0.172414    │
+│ reward/total           │ 0.591309    │
+│ skyrl.ai/grad_norm     │ 5530.089149 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 115.779778  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 10                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 10          │
+│ progress/done_frac     │ 0.189655    │
+│ reward/total           │ 0.589844    │
+│ skyrl.ai/grad_norm     │ 4833.589350 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 123.424306  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 11                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 11          │
+│ progress/done_frac     │ 0.206897    │
+│ reward/total           │ 0.636719    │
+│ skyrl.ai/grad_norm     │ 5552.164083 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 114.481740  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 12                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 12          │
+│ progress/done_frac     │ 0.224138    │
+│ reward/total           │ 0.587402    │
+│ skyrl.ai/grad_norm     │ 5142.660595 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 109.688283  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 13                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 13          │
+│ progress/done_frac     │ 0.241379    │
+│ reward/total           │ 0.628906    │
+│ skyrl.ai/grad_norm     │ 4498.109603 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 107.482208  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 14                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 14          │
+│ progress/done_frac     │ 0.258621    │
+│ reward/total           │ 0.585449    │
+│ skyrl.ai/grad_norm     │ 5197.242923 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 105.424653  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 15                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 15          │
+│ progress/done_frac     │ 0.275862    │
+│ reward/total           │ 0.646973    │
+│ skyrl.ai/grad_norm     │ 5264.652315 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 106.550719  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 16                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 16          │
+│ progress/done_frac     │ 0.293103    │
+│ reward/total           │ 0.694336    │
+│ skyrl.ai/grad_norm     │ 5170.707882 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 113.230792  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 17                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 17          │
+│ progress/done_frac     │ 0.310345    │
+│ reward/total           │ 0.690430    │
+│ skyrl.ai/grad_norm     │ 5101.805563 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 109.853788  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 18                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 18          │
+│ progress/done_frac     │ 0.327586    │
+│ reward/total           │ 0.625000    │
+│ skyrl.ai/grad_norm     │ 5243.874140 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 115.444906  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 19                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 19          │
+│ progress/done_frac     │ 0.344828    │
+│ reward/total           │ 0.684570    │
+│ skyrl.ai/grad_norm     │ 4356.659730 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 110.799166  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_e06832f2/weights/000020'}
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 20                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 20          │
+│ progress/done_frac     │ 0.362069    │
+│ reward/total           │ 0.699219    │
+│ skyrl.ai/grad_norm     │ 5144.779879 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 113.444232  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 21                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 21          │
+│ progress/done_frac     │ 0.379310    │
+│ reward/total           │ 0.697754    │
+│ skyrl.ai/grad_norm     │ 5902.052863 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 116.014966  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 22                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 22          │
+│ progress/done_frac     │ 0.396552    │
+│ reward/total           │ 0.722656    │
+│ skyrl.ai/grad_norm     │ 5800.156894 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 117.478718  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 23                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 23          │
+│ progress/done_frac     │ 0.413793    │
+│ reward/total           │ 0.712891    │
+│ skyrl.ai/grad_norm     │ 3894.663528 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 119.337539  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 24                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 24          │
+│ progress/done_frac     │ 0.431034    │
+│ reward/total           │ 0.658691    │
+│ skyrl.ai/grad_norm     │ 5355.445453 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 110.459812  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 25                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 25          │
+│ progress/done_frac     │ 0.448276    │
+│ reward/total           │ 0.707520    │
+│ skyrl.ai/grad_norm     │ 4574.888195 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 127.024712  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 26                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 26          │
+│ progress/done_frac     │ 0.465517    │
+│ reward/total           │ 0.700195    │
+│ skyrl.ai/grad_norm     │ 5258.031571 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 101.268171  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 27                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 27          │
+│ progress/done_frac     │ 0.482759    │
+│ reward/total           │ 0.729492    │
+│ skyrl.ai/grad_norm     │ 4962.903586 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 108.927979  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 28                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 28          │
+│ progress/done_frac     │ 0.500000    │
+│ reward/total           │ 0.774902    │
+│ skyrl.ai/grad_norm     │ 5472.180187 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 109.060334  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 29                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 29          │
+│ progress/done_frac     │ 0.517241    │
+│ reward/total           │ 0.754395    │
+│ skyrl.ai/grad_norm     │ 4148.777410 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 118.506368  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 30                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 30          │
+│ progress/done_frac     │ 0.534483    │
+│ reward/total           │ 0.720703    │
+│ skyrl.ai/grad_norm     │ 5368.396409 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 114.272527  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 31                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 31          │
+│ progress/done_frac     │ 0.551724    │
+│ reward/total           │ 0.756836    │
+│ skyrl.ai/grad_norm     │ 3687.510949 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 110.167546  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 32                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 32          │
+│ progress/done_frac     │ 0.568966    │
+│ reward/total           │ 0.661621    │
+│ skyrl.ai/grad_norm     │ 5193.492659 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 133.503203  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 33                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 33          │
+│ progress/done_frac     │ 0.586207    │
+│ reward/total           │ 0.779297    │
+│ skyrl.ai/grad_norm     │ 4541.073662 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 112.264678  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 34                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 34          │
+│ progress/done_frac     │ 0.603448    │
+│ reward/total           │ 0.708984    │
+│ skyrl.ai/grad_norm     │ 5435.408172 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 128.267636  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 35                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 35          │
+│ progress/done_frac     │ 0.620690    │
+│ reward/total           │ 0.708496    │
+│ skyrl.ai/grad_norm     │ 4911.940961 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 108.495283  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 36                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 36          │
+│ progress/done_frac     │ 0.637931    │
+│ reward/total           │ 0.694824    │
+│ skyrl.ai/grad_norm     │ 4770.140878 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 104.190873  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 37                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 37          │
+│ progress/done_frac     │ 0.655172    │
+│ reward/total           │ 0.729980    │
+│ skyrl.ai/grad_norm     │ 5225.622642 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 97.960649   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 38                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 38          │
+│ progress/done_frac     │ 0.672414    │
+│ reward/total           │ 0.714844    │
+│ skyrl.ai/grad_norm     │ 4925.707665 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 112.226573  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 39                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 39          │
+│ progress/done_frac     │ 0.689655    │
+│ reward/total           │ 0.749512    │
+│ skyrl.ai/grad_norm     │ 4603.290997 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 114.274745  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_e06832f2/weights/000040'}
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 40                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 40          │
+│ progress/done_frac     │ 0.706897    │
+│ reward/total           │ 0.742188    │
+│ skyrl.ai/grad_norm     │ 4576.326256 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 160.028438  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 41                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 41          │
+│ progress/done_frac     │ 0.724138    │
+│ reward/total           │ 0.715820    │
+│ skyrl.ai/grad_norm     │ 5315.205734 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 87.291408   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 42                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 42          │
+│ progress/done_frac     │ 0.741379    │
+│ reward/total           │ 0.731934    │
+│ skyrl.ai/grad_norm     │ 4193.670469 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 68.908308   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 43                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 43          │
+│ progress/done_frac     │ 0.758621    │
+│ reward/total           │ 0.703613    │
+│ skyrl.ai/grad_norm     │ 4106.626109 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 68.086455   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 44                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 44          │
+│ progress/done_frac     │ 0.775862    │
+│ reward/total           │ 0.830078    │
+│ skyrl.ai/grad_norm     │ 4323.928307 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.415823   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 45                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 45          │
+│ progress/done_frac     │ 0.793103    │
+│ reward/total           │ 0.776367    │
+│ skyrl.ai/grad_norm     │ 4000.091249 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 63.905457   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 46                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 46          │
+│ progress/done_frac     │ 0.810345    │
+│ reward/total           │ 0.727051    │
+│ skyrl.ai/grad_norm     │ 3997.601656 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 67.781055   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 47                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 47          │
+│ progress/done_frac     │ 0.827586    │
+│ reward/total           │ 0.796875    │
+│ skyrl.ai/grad_norm     │ 3721.394497 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 59.962005   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 48                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 48          │
+│ progress/done_frac     │ 0.844828    │
+│ reward/total           │ 0.791504    │
+│ skyrl.ai/grad_norm     │ 4527.206203 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.772083   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 49                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 49          │
+│ progress/done_frac     │ 0.862069    │
+│ reward/total           │ 0.709961    │
+│ skyrl.ai/grad_norm     │ 4433.696877 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 67.995457   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 50                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 50          │
+│ progress/done_frac     │ 0.879310    │
+│ reward/total           │ 0.738770    │
+│ skyrl.ai/grad_norm     │ 4101.841538 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 62.500658   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 51                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 51          │
+│ progress/done_frac     │ 0.896552    │
+│ reward/total           │ 0.734375    │
+│ skyrl.ai/grad_norm     │ 4450.920130 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 65.566620   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 52                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value        ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040     │
+│ progress/batch         │ 52           │
+│ progress/done_frac     │ 0.913793     │
+│ reward/total           │ 0.690430     │
+│ skyrl.ai/grad_norm     │ 17508.627359 │
+│ skyrl.ai/learning_rate │ 0.000040     │
+│ time/total             │ 64.398955    │
+└────────────────────────┴──────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 53                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 53          │
+│ progress/done_frac     │ 0.931034    │
+│ reward/total           │ 0.812500    │
+│ skyrl.ai/grad_norm     │ 4251.743642 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 63.225998   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 54                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 54          │
+│ progress/done_frac     │ 0.948276    │
+│ reward/total           │ 0.721680    │
+│ skyrl.ai/grad_norm     │ 4362.466275 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.625412   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 55                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 55          │
+│ progress/done_frac     │ 0.965517    │
+│ reward/total           │ 0.772461    │
+│ skyrl.ai/grad_norm     │ 4851.632097 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.751420   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 56                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 56          │
+│ progress/done_frac     │ 0.982759    │
+│ reward/total           │ 0.781738    │
+│ skyrl.ai/grad_norm     │ 4099.095510 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 65.074497   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_f.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 57                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 57          │
+│ progress/done_frac     │ 1.000000    │
+│ reward/total           │ 0.749023    │
+│ skyrl.ai/grad_norm     │ 7324.760474 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.397691   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_e06832f2/weights/final', 'sampler_path': 'tinker://model_e06832f2/final'}
+__main__:257 [INFO] Training completed
diff --git a/tests/tinker/smoke_logs/rl_loop_f.log/metrics.jsonl b/tests/tinker/smoke_logs/rl_loop_f.log/metrics.jsonl
new file mode 100644
index 0000000000..4a1587b4fd
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_f.log/metrics.jsonl
@@ -0,0 +1,58 @@
+{"step": 0, "progress/batch": 0, "optim/lr": 4e-05, "progress/done_frac": 0.017241379310344827, "skyrl.ai/grad_norm": 2294.091214402775, "skyrl.ai/learning_rate": 4e-05, "time/total": 95.20128583908081, "reward/total": 0.0166015625}
+{"step": 1, "progress/batch": 1, "optim/lr": 4e-05, "progress/done_frac": 0.034482758620689655, "skyrl.ai/grad_norm": 2692.103174100131, "skyrl.ai/learning_rate": 4e-05, "time/total": 173.87437748908997, "reward/total": 0.0224609375}
+{"step": 2, "progress/batch": 2, "optim/lr": 4e-05, "progress/done_frac": 0.05172413793103448, "skyrl.ai/grad_norm": 2919.905991637402, "skyrl.ai/learning_rate": 4e-05, "time/total": 103.14629554748535, "reward/total": 0.02587890625}
+{"step": 3, "progress/batch": 3, "optim/lr": 4e-05, "progress/done_frac": 0.06896551724137931, "skyrl.ai/grad_norm": 4679.051827026497, "skyrl.ai/learning_rate": 4e-05, "time/total": 138.49598717689514, "reward/total": 0.05078125}
+{"step": 4, "progress/batch": 4, "optim/lr": 4e-05, "progress/done_frac": 0.08620689655172414, "skyrl.ai/grad_norm": 7131.992989340357, "skyrl.ai/learning_rate": 4e-05, "time/total": 127.66751670837402, "reward/total": 0.1728515625}
+{"step": 5, "progress/batch": 5, "optim/lr": 4e-05, "progress/done_frac": 0.10344827586206896, "skyrl.ai/grad_norm": 10840.603673227797, "skyrl.ai/learning_rate": 4e-05, "time/total": 142.25334334373474, "reward/total": 0.3154296875}
+{"step": 6, "progress/batch": 6, "optim/lr": 4e-05, "progress/done_frac": 0.1206896551724138, "skyrl.ai/grad_norm": 8993.776959653826, "skyrl.ai/learning_rate": 4e-05, "time/total": 138.3899528980255, "reward/total": 0.41162109375}
+{"step": 7, "progress/batch": 7, "optim/lr": 4e-05, "progress/done_frac": 0.13793103448275862, "skyrl.ai/grad_norm": 6986.811003598136, "skyrl.ai/learning_rate": 4e-05, "time/total": 129.658682346344, "reward/total": 0.56005859375}
+{"step": 8, "progress/batch": 8, "optim/lr": 4e-05, "progress/done_frac": 0.15517241379310345, "skyrl.ai/grad_norm": 5574.3074906215925, "skyrl.ai/learning_rate": 4e-05, "time/total": 125.28949284553528, "reward/total": 0.5380859375}
+{"step": 9, "progress/batch": 9, "optim/lr": 4e-05, "progress/done_frac": 0.1724137931034483, "skyrl.ai/grad_norm": 5530.089149371825, "skyrl.ai/learning_rate": 4e-05, "time/total": 115.77977848052979, "reward/total": 0.59130859375}
+{"step": 10, "progress/batch": 10, "optim/lr": 4e-05, "progress/done_frac": 0.1896551724137931, "skyrl.ai/grad_norm": 4833.589349541395, "skyrl.ai/learning_rate": 4e-05, "time/total": 123.42430639266968, "reward/total": 0.58984375}
+{"step": 11, "progress/batch": 11, "optim/lr": 4e-05, "progress/done_frac": 0.20689655172413793, "skyrl.ai/grad_norm": 5552.164082589778, "skyrl.ai/learning_rate": 4e-05, "time/total": 114.48173975944519, "reward/total": 0.63671875}
+{"step": 12, "progress/batch": 12, "optim/lr": 4e-05, "progress/done_frac": 0.22413793103448276, "skyrl.ai/grad_norm": 5142.6605954505685, "skyrl.ai/learning_rate": 4e-05, "time/total": 109.68828344345093, "reward/total": 0.58740234375}
+{"step": 13, "progress/batch": 13, "optim/lr": 4e-05, "progress/done_frac": 0.2413793103448276, "skyrl.ai/grad_norm": 4498.109602933215, "skyrl.ai/learning_rate": 4e-05, "time/total": 107.48220753669739, "reward/total": 0.62890625}
+{"step": 14, "progress/batch": 14, "optim/lr": 4e-05, "progress/done_frac": 0.25862068965517243, "skyrl.ai/grad_norm": 5197.242922935198, "skyrl.ai/learning_rate": 4e-05, "time/total": 105.42465329170227, "reward/total": 0.58544921875}
+{"step": 15, "progress/batch": 15, "optim/lr": 4e-05, "progress/done_frac": 0.27586206896551724, "skyrl.ai/grad_norm": 5264.65231520563, "skyrl.ai/learning_rate": 4e-05, "time/total": 106.55071878433228, "reward/total": 0.64697265625}
+{"step": 16, "progress/batch": 16, "optim/lr": 4e-05, "progress/done_frac": 0.29310344827586204, "skyrl.ai/grad_norm": 5170.707881905532, "skyrl.ai/learning_rate": 4e-05, "time/total": 113.2307915687561, "reward/total": 0.6943359375}
+{"step": 17, "progress/batch": 17, "optim/lr": 4e-05, "progress/done_frac": 0.3103448275862069, "skyrl.ai/grad_norm": 5101.805562739529, "skyrl.ai/learning_rate": 4e-05, "time/total": 109.85378813743591, "reward/total": 0.6904296875}
+{"step": 18, "progress/batch": 18, "optim/lr": 4e-05, "progress/done_frac": 0.3275862068965517, "skyrl.ai/grad_norm": 5243.874140366071, "skyrl.ai/learning_rate": 4e-05, "time/total": 115.44490599632263, "reward/total": 0.625}
+{"step": 19, "progress/batch": 19, "optim/lr": 4e-05, "progress/done_frac": 0.3448275862068966, "skyrl.ai/grad_norm": 4356.659729655278, "skyrl.ai/learning_rate": 4e-05, "time/total": 110.79916596412659, "reward/total": 0.6845703125}
+{"step": 20, "progress/batch": 20, "optim/lr": 4e-05, "progress/done_frac": 0.3620689655172414, "skyrl.ai/grad_norm": 5144.779878673139, "skyrl.ai/learning_rate": 4e-05, "time/total": 113.44423246383667, "reward/total": 0.69921875}
+{"step": 21, "progress/batch": 21, "optim/lr": 4e-05, "progress/done_frac": 0.3793103448275862, "skyrl.ai/grad_norm": 5902.05286319938, "skyrl.ai/learning_rate": 4e-05, "time/total": 116.01496648788452, "reward/total": 0.69775390625}
+{"step": 22, "progress/batch": 22, "optim/lr": 4e-05, "progress/done_frac": 0.39655172413793105, "skyrl.ai/grad_norm": 5800.156894429667, "skyrl.ai/learning_rate": 4e-05, "time/total": 117.47871780395508, "reward/total": 0.72265625}
+{"step": 23, "progress/batch": 23, "optim/lr": 4e-05, "progress/done_frac": 0.41379310344827586, "skyrl.ai/grad_norm": 3894.66352847072, "skyrl.ai/learning_rate": 4e-05, "time/total": 119.33753895759583, "reward/total": 0.712890625}
+{"step": 24, "progress/batch": 24, "optim/lr": 4e-05, "progress/done_frac": 0.43103448275862066, "skyrl.ai/grad_norm": 5355.44545299455, "skyrl.ai/learning_rate": 4e-05, "time/total": 110.45981216430664, "reward/total": 0.65869140625}
+{"step": 25, "progress/batch": 25, "optim/lr": 4e-05, "progress/done_frac": 0.4482758620689655, "skyrl.ai/grad_norm": 4574.888195355161, "skyrl.ai/learning_rate": 4e-05, "time/total": 127.02471208572388, "reward/total": 0.70751953125}
+{"step": 26, "progress/batch": 26, "optim/lr": 4e-05, "progress/done_frac": 0.46551724137931033, "skyrl.ai/grad_norm": 5258.03157084474, "skyrl.ai/learning_rate": 4e-05, "time/total": 101.26817059516907, "reward/total": 0.7001953125}
+{"step": 27, "progress/batch": 27, "optim/lr": 4e-05, "progress/done_frac": 0.4827586206896552, "skyrl.ai/grad_norm": 4962.903585603895, "skyrl.ai/learning_rate": 4e-05, "time/total": 108.92797946929932, "reward/total": 0.7294921875}
+{"step": 28, "progress/batch": 28, "optim/lr": 4e-05, "progress/done_frac": 0.5, "skyrl.ai/grad_norm": 5472.180187091795, "skyrl.ai/learning_rate": 4e-05, "time/total": 109.06033444404602, "reward/total": 0.77490234375}
+{"step": 29, "progress/batch": 29, "optim/lr": 4e-05, "progress/done_frac": 0.5172413793103449, "skyrl.ai/grad_norm": 4148.7774102740195, "skyrl.ai/learning_rate": 4e-05, "time/total": 118.50636839866638, "reward/total": 0.75439453125}
+{"step": 30, "progress/batch": 30, "optim/lr": 4e-05, "progress/done_frac": 0.5344827586206896, "skyrl.ai/grad_norm": 5368.396408612166, "skyrl.ai/learning_rate": 4e-05, "time/total": 114.27252721786499, "reward/total": 0.720703125}
+{"step": 31, "progress/batch": 31, "optim/lr": 4e-05, "progress/done_frac": 0.5517241379310345, "skyrl.ai/grad_norm": 3687.510949136287, "skyrl.ai/learning_rate": 4e-05, "time/total": 110.1675455570221, "reward/total": 0.7568359375}
+{"step": 32, "progress/batch": 32, "optim/lr": 4e-05, "progress/done_frac": 0.5689655172413793, "skyrl.ai/grad_norm": 5193.492659087909, "skyrl.ai/learning_rate": 4e-05, "time/total": 133.5032033920288, "reward/total": 0.66162109375}
+{"step": 33, "progress/batch": 33, "optim/lr": 4e-05, "progress/done_frac": 0.5862068965517241, "skyrl.ai/grad_norm": 4541.073661591497, "skyrl.ai/learning_rate": 4e-05, "time/total": 112.26467752456665, "reward/total": 0.779296875}
+{"step": 34, "progress/batch": 34, "optim/lr": 4e-05, "progress/done_frac": 0.603448275862069, "skyrl.ai/grad_norm": 5435.408172345477, "skyrl.ai/learning_rate": 4e-05, "time/total": 128.2676362991333, "reward/total": 0.708984375}
+{"step": 35, "progress/batch": 35, "optim/lr": 4e-05, "progress/done_frac": 0.6206896551724138, "skyrl.ai/grad_norm": 4911.940960557242, "skyrl.ai/learning_rate": 4e-05, "time/total": 108.49528288841248, "reward/total": 0.70849609375}
+{"step": 36, "progress/batch": 36, "optim/lr": 4e-05, "progress/done_frac": 0.6379310344827587, "skyrl.ai/grad_norm": 4770.140878422774, "skyrl.ai/learning_rate": 4e-05, "time/total": 104.19087266921997, "reward/total": 0.69482421875}
+{"step": 37, "progress/batch": 37, "optim/lr": 4e-05, "progress/done_frac": 0.6551724137931034, "skyrl.ai/grad_norm": 5225.622642326941, "skyrl.ai/learning_rate": 4e-05, "time/total": 97.96064853668213, "reward/total": 0.72998046875}
+{"step": 38, "progress/batch": 38, "optim/lr": 4e-05, "progress/done_frac": 0.6724137931034483, "skyrl.ai/grad_norm": 4925.707664894457, "skyrl.ai/learning_rate": 4e-05, "time/total": 112.22657251358032, "reward/total": 0.71484375}
+{"step": 39, "progress/batch": 39, "optim/lr": 4e-05, "progress/done_frac": 0.6896551724137931, "skyrl.ai/grad_norm": 4603.290996667492, "skyrl.ai/learning_rate": 4e-05, "time/total": 114.27474546432495, "reward/total": 0.74951171875}
+{"step": 40, "progress/batch": 40, "optim/lr": 4e-05, "progress/done_frac": 0.7068965517241379, "skyrl.ai/grad_norm": 4576.326255851958, "skyrl.ai/learning_rate": 4e-05, "time/total": 160.0284378528595, "reward/total": 0.7421875}
+{"step": 41, "progress/batch": 41, "optim/lr": 4e-05, "progress/done_frac": 0.7241379310344828, "skyrl.ai/grad_norm": 5315.205734494198, "skyrl.ai/learning_rate": 4e-05, "time/total": 87.2914080619812, "reward/total": 0.7158203125}
+{"step": 42, "progress/batch": 42, "optim/lr": 4e-05, "progress/done_frac": 0.7413793103448276, "skyrl.ai/grad_norm": 4193.670468694459, "skyrl.ai/learning_rate": 4e-05, "time/total": 68.90830826759338, "reward/total": 0.73193359375}
+{"step": 43, "progress/batch": 43, "optim/lr": 4e-05, "progress/done_frac": 0.7586206896551724, "skyrl.ai/grad_norm": 4106.6261091070855, "skyrl.ai/learning_rate": 4e-05, "time/total": 68.08645534515381, "reward/total": 0.70361328125}
+{"step": 44, "progress/batch": 44, "optim/lr": 4e-05, "progress/done_frac": 0.7758620689655172, "skyrl.ai/grad_norm": 4323.928306528683, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.41582298278809, "reward/total": 0.830078125}
+{"step": 45, "progress/batch": 45, "optim/lr": 4e-05, "progress/done_frac": 0.7931034482758621, "skyrl.ai/grad_norm": 4000.0912489592033, "skyrl.ai/learning_rate": 4e-05, "time/total": 63.90545701980591, "reward/total": 0.7763671875}
+{"step": 46, "progress/batch": 46, "optim/lr": 4e-05, "progress/done_frac": 0.8103448275862069, "skyrl.ai/grad_norm": 3997.601655993253, "skyrl.ai/learning_rate": 4e-05, "time/total": 67.78105521202087, "reward/total": 0.72705078125}
+{"step": 47, "progress/batch": 47, "optim/lr": 4e-05, "progress/done_frac": 0.8275862068965517, "skyrl.ai/grad_norm": 3721.394496690723, "skyrl.ai/learning_rate": 4e-05, "time/total": 59.96200513839722, "reward/total": 0.796875}
+{"step": 48, "progress/batch": 48, "optim/lr": 4e-05, "progress/done_frac": 0.8448275862068966, "skyrl.ai/grad_norm": 4527.206202505028, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.77208304405212, "reward/total": 0.79150390625}
+{"step": 49, "progress/batch": 49, "optim/lr": 4e-05, "progress/done_frac": 0.8620689655172413, "skyrl.ai/grad_norm": 4433.696877324835, "skyrl.ai/learning_rate": 4e-05, "time/total": 67.99545741081238, "reward/total": 0.7099609375}
+{"step": 50, "progress/batch": 50, "optim/lr": 4e-05, "progress/done_frac": 0.8793103448275862, "skyrl.ai/grad_norm": 4101.8415376511075, "skyrl.ai/learning_rate": 4e-05, "time/total": 62.50065779685974, "reward/total": 0.73876953125}
+{"step": 51, "progress/batch": 51, "optim/lr": 4e-05, "progress/done_frac": 0.896551724137931, "skyrl.ai/grad_norm": 4450.9201295911835, "skyrl.ai/learning_rate": 4e-05, "time/total": 65.5666196346283, "reward/total": 0.734375}
+{"step": 52, "progress/batch": 52, "optim/lr": 4e-05, "progress/done_frac": 0.9137931034482759, "skyrl.ai/grad_norm": 17508.627359104998, "skyrl.ai/learning_rate": 4e-05, "time/total": 64.39895486831665, "reward/total": 0.6904296875}
+{"step": 53, "progress/batch": 53, "optim/lr": 4e-05, "progress/done_frac": 0.9310344827586207, "skyrl.ai/grad_norm": 4251.743642318996, "skyrl.ai/learning_rate": 4e-05, "time/total": 63.22599792480469, "reward/total": 0.8125}
+{"step": 54, "progress/batch": 54, "optim/lr": 4e-05, "progress/done_frac": 0.9482758620689655, "skyrl.ai/grad_norm": 4362.466274941275, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.62541174888611, "reward/total": 0.7216796875}
+{"step": 55, "progress/batch": 55, "optim/lr": 4e-05, "progress/done_frac": 0.9655172413793104, "skyrl.ai/grad_norm": 4851.6320965217465, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.75142049789429, "reward/total": 0.7724609375}
+{"step": 56, "progress/batch": 56, "optim/lr": 4e-05, "progress/done_frac": 0.9827586206896551, "skyrl.ai/grad_norm": 4099.095509987539, "skyrl.ai/learning_rate": 4e-05, "time/total": 65.07449650764465, "reward/total": 0.78173828125}
+{"step": 57, "progress/batch": 57, "optim/lr": 4e-05, "progress/done_frac": 1.0, "skyrl.ai/grad_norm": 7324.7604738994705, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.39769148826599, "reward/total": 0.7490234375}
diff --git a/tests/tinker/smoke_logs/rl_loop_g.log/checkpoints.jsonl b/tests/tinker/smoke_logs/rl_loop_g.log/checkpoints.jsonl
new file mode 100644
index 0000000000..4c7109497f
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_g.log/checkpoints.jsonl
@@ -0,0 +1,3 @@
+{"name": "000020", "batch": 20, "state_path": "tinker://model_33022fca/weights/000020"}
+{"name": "000040", "batch": 40, "state_path": "tinker://model_33022fca/weights/000040"}
+{"name": "final", "batch": 58, "state_path": "tinker://model_33022fca/weights/final", "sampler_path": "tinker://model_33022fca/final"}
diff --git a/tests/tinker/smoke_logs/rl_loop_g.log/code.diff b/tests/tinker/smoke_logs/rl_loop_g.log/code.diff
new file mode 100644
index 0000000000..d8e559755f
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_g.log/code.diff
@@ -0,0 +1,3 @@
+### repo: /home/etang/tinker-cookbook @ dacb835b31fb3f7a012851a1bc64c950d0495de3
+modules: tinker_cookbook
+(no local changes)
diff --git a/tests/tinker/smoke_logs/rl_loop_g.log/config.json b/tests/tinker/smoke_logs/rl_loop_g.log/config.json
new file mode 100644
index 0000000000..d5c9146cd5
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_g.log/config.json
@@ -0,0 +1,12 @@
+{
+  "base_url": "http://localhost:8000",
+  "log_path": "/tmp/rl_loop_g.log",
+  "model_name": "Qwen/Qwen3-0.6B",
+  "batch_size": 128,
+  "group_size": 16,
+  "learning_rate": 4e-05,
+  "lora_rank": 32,
+  "save_every": 20,
+  "max_tokens": 256,
+  "ttl_seconds": 604800
+}
\ No newline at end of file
diff --git a/tests/tinker/smoke_logs/rl_loop_g.log/logs.log b/tests/tinker/smoke_logs/rl_loop_g.log/logs.log
new file mode 100644
index 0000000000..37bf57a3e7
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_g.log/logs.log
@@ -0,0 +1,825 @@
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/rl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B lora_rank=32 log_path=/tmp/rl_loop_g.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/rl_loop_g.log
+__main__:73 [INFO] Using renderer: qwen3
+__main__:76 [INFO] Loading dataset...
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_43f00dc0
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/rl_loop_g.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/rl_loop_g.log
+tinker.lib.public_interfaces.service_client:159 [INFO] TrainingClient initialized for model model_33022fca
+__main__:121 [INFO] Training for 58 batches
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 0                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 0           │
+│ progress/done_frac     │ 0.017241    │
+│ reward/total           │ 0.019043    │
+│ skyrl.ai/grad_norm     │ 3047.044306 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 81.935303   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 1                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 1           │
+│ progress/done_frac     │ 0.034483    │
+│ reward/total           │ 0.034668    │
+│ skyrl.ai/grad_norm     │ 2863.267714 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 79.756085   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 2                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 2           │
+│ progress/done_frac     │ 0.051724    │
+│ reward/total           │ 0.080078    │
+│ skyrl.ai/grad_norm     │ 3862.287017 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 84.182842   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 3                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 3           │
+│ progress/done_frac     │ 0.068966    │
+│ reward/total           │ 0.257812    │
+│ skyrl.ai/grad_norm     │ 8041.708525 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 91.647291   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 4                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 4           │
+│ progress/done_frac     │ 0.086207    │
+│ reward/total           │ 0.547363    │
+│ skyrl.ai/grad_norm     │ 4885.541730 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 76.656733   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 5                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 5           │
+│ progress/done_frac     │ 0.103448    │
+│ reward/total           │ 0.603027    │
+│ skyrl.ai/grad_norm     │ 4959.070881 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 70.430493   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 6                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 6           │
+│ progress/done_frac     │ 0.120690    │
+│ reward/total           │ 0.568848    │
+│ skyrl.ai/grad_norm     │ 3898.757494 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 65.931934   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 7                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 7           │
+│ progress/done_frac     │ 0.137931    │
+│ reward/total           │ 0.600586    │
+│ skyrl.ai/grad_norm     │ 4328.296894 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 59.117071   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 8                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 8           │
+│ progress/done_frac     │ 0.155172    │
+│ reward/total           │ 0.535156    │
+│ skyrl.ai/grad_norm     │ 4568.654288 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 62.654777   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 9                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 9           │
+│ progress/done_frac     │ 0.172414    │
+│ reward/total           │ 0.617676    │
+│ skyrl.ai/grad_norm     │ 4046.557673 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 63.533596   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 10                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 10          │
+│ progress/done_frac     │ 0.189655    │
+│ reward/total           │ 0.625977    │
+│ skyrl.ai/grad_norm     │ 3957.613675 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 64.592402   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 11                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 11          │
+│ progress/done_frac     │ 0.206897    │
+│ reward/total           │ 0.677734    │
+│ skyrl.ai/grad_norm     │ 4522.047103 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 63.164687   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 12                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 12          │
+│ progress/done_frac     │ 0.224138    │
+│ reward/total           │ 0.622559    │
+│ skyrl.ai/grad_norm     │ 4453.164717 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 67.606588   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 13                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 13          │
+│ progress/done_frac     │ 0.241379    │
+│ reward/total           │ 0.668457    │
+│ skyrl.ai/grad_norm     │ 4064.819307 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.253005   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 14                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 14          │
+│ progress/done_frac     │ 0.258621    │
+│ reward/total           │ 0.635742    │
+│ skyrl.ai/grad_norm     │ 4596.063751 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 69.474511   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 15                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 15          │
+│ progress/done_frac     │ 0.275862    │
+│ reward/total           │ 0.718750    │
+│ skyrl.ai/grad_norm     │ 4935.121072 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 67.929768   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 16                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 16          │
+│ progress/done_frac     │ 0.293103    │
+│ reward/total           │ 0.740723    │
+│ skyrl.ai/grad_norm     │ 4247.624277 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.258889   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 17                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 17          │
+│ progress/done_frac     │ 0.310345    │
+│ reward/total           │ 0.705566    │
+│ skyrl.ai/grad_norm     │ 4071.461163 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.827817   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 18                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 18          │
+│ progress/done_frac     │ 0.327586    │
+│ reward/total           │ 0.689941    │
+│ skyrl.ai/grad_norm     │ 3900.221019 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 63.241114   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 19                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 19          │
+│ progress/done_frac     │ 0.344828    │
+│ reward/total           │ 0.720703    │
+│ skyrl.ai/grad_norm     │ 4153.697630 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.975040   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_33022fca/weights/000020'}
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 20                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 20          │
+│ progress/done_frac     │ 0.362069    │
+│ reward/total           │ 0.729980    │
+│ skyrl.ai/grad_norm     │ 3834.269292 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 65.836580   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 21                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 21          │
+│ progress/done_frac     │ 0.379310    │
+│ reward/total           │ 0.731934    │
+│ skyrl.ai/grad_norm     │ 4056.313967 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 65.094682   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 22                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 22          │
+│ progress/done_frac     │ 0.396552    │
+│ reward/total           │ 0.743652    │
+│ skyrl.ai/grad_norm     │ 4249.544681 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.107702   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 23                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 23          │
+│ progress/done_frac     │ 0.413793    │
+│ reward/total           │ 0.750977    │
+│ skyrl.ai/grad_norm     │ 3919.666567 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 63.815610   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 24                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 24          │
+│ progress/done_frac     │ 0.431034    │
+│ reward/total           │ 0.668457    │
+│ skyrl.ai/grad_norm     │ 4038.148462 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 67.941240   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 25                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 25          │
+│ progress/done_frac     │ 0.448276    │
+│ reward/total           │ 0.745117    │
+│ skyrl.ai/grad_norm     │ 4261.364570 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 65.611496   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 26                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 26          │
+│ progress/done_frac     │ 0.465517    │
+│ reward/total           │ 0.730469    │
+│ skyrl.ai/grad_norm     │ 3908.385472 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.793354   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 27                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 27          │
+│ progress/done_frac     │ 0.482759    │
+│ reward/total           │ 0.728516    │
+│ skyrl.ai/grad_norm     │ 4198.391835 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 64.184751   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 28                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 28          │
+│ progress/done_frac     │ 0.500000    │
+│ reward/total           │ 0.759766    │
+│ skyrl.ai/grad_norm     │ 3855.957858 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 67.562127   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 29                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 29          │
+│ progress/done_frac     │ 0.517241    │
+│ reward/total           │ 0.775391    │
+│ skyrl.ai/grad_norm     │ 3629.153483 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 64.316827   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 30                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 30          │
+│ progress/done_frac     │ 0.534483    │
+│ reward/total           │ 0.742188    │
+│ skyrl.ai/grad_norm     │ 4626.587295 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.431585   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 31                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 31          │
+│ progress/done_frac     │ 0.551724    │
+│ reward/total           │ 0.779785    │
+│ skyrl.ai/grad_norm     │ 3172.986763 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 61.603456   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 32                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 32          │
+│ progress/done_frac     │ 0.568966    │
+│ reward/total           │ 0.664551    │
+│ skyrl.ai/grad_norm     │ 4532.925766 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 70.215890   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 33                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 33          │
+│ progress/done_frac     │ 0.586207    │
+│ reward/total           │ 0.782227    │
+│ skyrl.ai/grad_norm     │ 4165.589514 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 65.486906   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 34                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 34          │
+│ progress/done_frac     │ 0.603448    │
+│ reward/total           │ 0.710449    │
+│ skyrl.ai/grad_norm     │ 4447.996853 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 73.745564   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 35                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 35          │
+│ progress/done_frac     │ 0.620690    │
+│ reward/total           │ 0.719238    │
+│ skyrl.ai/grad_norm     │ 4060.280532 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 67.713121   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 36                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 36          │
+│ progress/done_frac     │ 0.637931    │
+│ reward/total           │ 0.728027    │
+│ skyrl.ai/grad_norm     │ 4412.181547 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 64.371590   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 37                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 37          │
+│ progress/done_frac     │ 0.655172    │
+│ reward/total           │ 0.740723    │
+│ skyrl.ai/grad_norm     │ 3785.710765 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.436951   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 38                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 38          │
+│ progress/done_frac     │ 0.672414    │
+│ reward/total           │ 0.696289    │
+│ skyrl.ai/grad_norm     │ 4860.556141 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 69.290205   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 39                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 39          │
+│ progress/done_frac     │ 0.689655    │
+│ reward/total           │ 0.774902    │
+│ skyrl.ai/grad_norm     │ 4201.451892 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 63.281941   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_33022fca/weights/000040'}
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 40                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 40          │
+│ progress/done_frac     │ 0.706897    │
+│ reward/total           │ 0.736816    │
+│ skyrl.ai/grad_norm     │ 4512.850097 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 67.527777   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 41                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 41          │
+│ progress/done_frac     │ 0.724138    │
+│ reward/total           │ 0.736816    │
+│ skyrl.ai/grad_norm     │ 4640.554924 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.977870   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 42                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 42          │
+│ progress/done_frac     │ 0.741379    │
+│ reward/total           │ 0.731445    │
+│ skyrl.ai/grad_norm     │ 4721.049248 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 68.467704   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 43                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 43          │
+│ progress/done_frac     │ 0.758621    │
+│ reward/total           │ 0.703613    │
+│ skyrl.ai/grad_norm     │ 4301.910738 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 67.849135   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 44                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 44          │
+│ progress/done_frac     │ 0.775862    │
+│ reward/total           │ 0.808105    │
+│ skyrl.ai/grad_norm     │ 4144.882387 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 65.650478   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 45                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 45          │
+│ progress/done_frac     │ 0.793103    │
+│ reward/total           │ 0.785645    │
+│ skyrl.ai/grad_norm     │ 4613.070344 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 63.609451   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 46                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 46          │
+│ progress/done_frac     │ 0.810345    │
+│ reward/total           │ 0.735840    │
+│ skyrl.ai/grad_norm     │ 4161.312293 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 67.403885   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 47                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 47          │
+│ progress/done_frac     │ 0.827586    │
+│ reward/total           │ 0.793945    │
+│ skyrl.ai/grad_norm     │ 3778.356653 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 63.203141   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 48                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 48          │
+│ progress/done_frac     │ 0.844828    │
+│ reward/total           │ 0.770508    │
+│ skyrl.ai/grad_norm     │ 4912.503639 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.865016   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 49                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 49          │
+│ progress/done_frac     │ 0.862069    │
+│ reward/total           │ 0.699707    │
+│ skyrl.ai/grad_norm     │ 4550.238895 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 65.346623   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 50                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 50          │
+│ progress/done_frac     │ 0.879310    │
+│ reward/total           │ 0.737793    │
+│ skyrl.ai/grad_norm     │ 5372.916154 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 64.003319   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 51                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 51          │
+│ progress/done_frac     │ 0.896552    │
+│ reward/total           │ 0.731934    │
+│ skyrl.ai/grad_norm     │ 4387.628517 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 64.909327   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 52                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 52          │
+│ progress/done_frac     │ 0.913793    │
+│ reward/total           │ 0.691406    │
+│ skyrl.ai/grad_norm     │ 4861.483930 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.043161   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 53                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 53          │
+│ progress/done_frac     │ 0.931034    │
+│ reward/total           │ 0.809082    │
+│ skyrl.ai/grad_norm     │ 3259.830210 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 62.288203   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 54                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 54          │
+│ progress/done_frac     │ 0.948276    │
+│ reward/total           │ 0.718750    │
+│ skyrl.ai/grad_norm     │ 4360.203894 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 66.404381   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 55                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 55          │
+│ progress/done_frac     │ 0.965517    │
+│ reward/total           │ 0.782715    │
+│ skyrl.ai/grad_norm     │ 4557.082619 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 65.916927   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 56                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 56          │
+│ progress/done_frac     │ 0.982759    │
+│ reward/total           │ 0.790527    │
+│ skyrl.ai/grad_norm     │ 3941.068891 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 67.037578   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_g.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 57                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 57          │
+│ progress/done_frac     │ 1.000000    │
+│ reward/total           │ 0.739258    │
+│ skyrl.ai/grad_norm     │ 4977.277569 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 67.706937   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_33022fca/weights/final', 'sampler_path': 'tinker://model_33022fca/final'}
+__main__:257 [INFO] Training completed
diff --git a/tests/tinker/smoke_logs/rl_loop_g.log/metrics.jsonl b/tests/tinker/smoke_logs/rl_loop_g.log/metrics.jsonl
new file mode 100644
index 0000000000..069f5b8f83
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_g.log/metrics.jsonl
@@ -0,0 +1,58 @@
+{"step": 0, "progress/batch": 0, "optim/lr": 4e-05, "progress/done_frac": 0.017241379310344827, "skyrl.ai/grad_norm": 3047.044305552514, "skyrl.ai/learning_rate": 4e-05, "time/total": 81.935302734375, "reward/total": 0.01904296875}
+{"step": 1, "progress/batch": 1, "optim/lr": 4e-05, "progress/done_frac": 0.034482758620689655, "skyrl.ai/grad_norm": 2863.267713644674, "skyrl.ai/learning_rate": 4e-05, "time/total": 79.75608515739441, "reward/total": 0.03466796875}
+{"step": 2, "progress/batch": 2, "optim/lr": 4e-05, "progress/done_frac": 0.05172413793103448, "skyrl.ai/grad_norm": 3862.2870167816372, "skyrl.ai/learning_rate": 4e-05, "time/total": 84.18284177780151, "reward/total": 0.080078125}
+{"step": 3, "progress/batch": 3, "optim/lr": 4e-05, "progress/done_frac": 0.06896551724137931, "skyrl.ai/grad_norm": 8041.708524934238, "skyrl.ai/learning_rate": 4e-05, "time/total": 91.6472909450531, "reward/total": 0.2578125}
+{"step": 4, "progress/batch": 4, "optim/lr": 4e-05, "progress/done_frac": 0.08620689655172414, "skyrl.ai/grad_norm": 4885.541730453236, "skyrl.ai/learning_rate": 4e-05, "time/total": 76.65673303604126, "reward/total": 0.54736328125}
+{"step": 5, "progress/batch": 5, "optim/lr": 4e-05, "progress/done_frac": 0.10344827586206896, "skyrl.ai/grad_norm": 4959.070880719492, "skyrl.ai/learning_rate": 4e-05, "time/total": 70.4304928779602, "reward/total": 0.60302734375}
+{"step": 6, "progress/batch": 6, "optim/lr": 4e-05, "progress/done_frac": 0.1206896551724138, "skyrl.ai/grad_norm": 3898.757494382024, "skyrl.ai/learning_rate": 4e-05, "time/total": 65.93193435668945, "reward/total": 0.56884765625}
+{"step": 7, "progress/batch": 7, "optim/lr": 4e-05, "progress/done_frac": 0.13793103448275862, "skyrl.ai/grad_norm": 4328.29689369849, "skyrl.ai/learning_rate": 4e-05, "time/total": 59.11707139015198, "reward/total": 0.6005859375}
+{"step": 8, "progress/batch": 8, "optim/lr": 4e-05, "progress/done_frac": 0.15517241379310345, "skyrl.ai/grad_norm": 4568.654287643135, "skyrl.ai/learning_rate": 4e-05, "time/total": 62.65477657318115, "reward/total": 0.53515625}
+{"step": 9, "progress/batch": 9, "optim/lr": 4e-05, "progress/done_frac": 0.1724137931034483, "skyrl.ai/grad_norm": 4046.557672886919, "skyrl.ai/learning_rate": 4e-05, "time/total": 63.53359603881836, "reward/total": 0.61767578125}
+{"step": 10, "progress/batch": 10, "optim/lr": 4e-05, "progress/done_frac": 0.1896551724137931, "skyrl.ai/grad_norm": 3957.613674930892, "skyrl.ai/learning_rate": 4e-05, "time/total": 64.59240198135376, "reward/total": 0.6259765625}
+{"step": 11, "progress/batch": 11, "optim/lr": 4e-05, "progress/done_frac": 0.20689655172413793, "skyrl.ai/grad_norm": 4522.047102806427, "skyrl.ai/learning_rate": 4e-05, "time/total": 63.16468691825867, "reward/total": 0.677734375}
+{"step": 12, "progress/batch": 12, "optim/lr": 4e-05, "progress/done_frac": 0.22413793103448276, "skyrl.ai/grad_norm": 4453.164717366741, "skyrl.ai/learning_rate": 4e-05, "time/total": 67.6065878868103, "reward/total": 0.62255859375}
+{"step": 13, "progress/batch": 13, "optim/lr": 4e-05, "progress/done_frac": 0.2413793103448276, "skyrl.ai/grad_norm": 4064.8193071771348, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.25300526618958, "reward/total": 0.66845703125}
+{"step": 14, "progress/batch": 14, "optim/lr": 4e-05, "progress/done_frac": 0.25862068965517243, "skyrl.ai/grad_norm": 4596.063750645763, "skyrl.ai/learning_rate": 4e-05, "time/total": 69.47451114654541, "reward/total": 0.6357421875}
+{"step": 15, "progress/batch": 15, "optim/lr": 4e-05, "progress/done_frac": 0.27586206896551724, "skyrl.ai/grad_norm": 4935.121072476338, "skyrl.ai/learning_rate": 4e-05, "time/total": 67.92976808547974, "reward/total": 0.71875}
+{"step": 16, "progress/batch": 16, "optim/lr": 4e-05, "progress/done_frac": 0.29310344827586204, "skyrl.ai/grad_norm": 4247.624277169533, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.2588894367218, "reward/total": 0.74072265625}
+{"step": 17, "progress/batch": 17, "optim/lr": 4e-05, "progress/done_frac": 0.3103448275862069, "skyrl.ai/grad_norm": 4071.461162776823, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.82781744003296, "reward/total": 0.70556640625}
+{"step": 18, "progress/batch": 18, "optim/lr": 4e-05, "progress/done_frac": 0.3275862068965517, "skyrl.ai/grad_norm": 3900.2210193782607, "skyrl.ai/learning_rate": 4e-05, "time/total": 63.241114139556885, "reward/total": 0.68994140625}
+{"step": 19, "progress/batch": 19, "optim/lr": 4e-05, "progress/done_frac": 0.3448275862068966, "skyrl.ai/grad_norm": 4153.697629823336, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.97503972053528, "reward/total": 0.720703125}
+{"step": 20, "progress/batch": 20, "optim/lr": 4e-05, "progress/done_frac": 0.3620689655172414, "skyrl.ai/grad_norm": 3834.269291533916, "skyrl.ai/learning_rate": 4e-05, "time/total": 65.83657956123352, "reward/total": 0.72998046875}
+{"step": 21, "progress/batch": 21, "optim/lr": 4e-05, "progress/done_frac": 0.3793103448275862, "skyrl.ai/grad_norm": 4056.3139671381455, "skyrl.ai/learning_rate": 4e-05, "time/total": 65.09468173980713, "reward/total": 0.73193359375}
+{"step": 22, "progress/batch": 22, "optim/lr": 4e-05, "progress/done_frac": 0.39655172413793105, "skyrl.ai/grad_norm": 4249.54468149236, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.10770153999329, "reward/total": 0.74365234375}
+{"step": 23, "progress/batch": 23, "optim/lr": 4e-05, "progress/done_frac": 0.41379310344827586, "skyrl.ai/grad_norm": 3919.6665674518795, "skyrl.ai/learning_rate": 4e-05, "time/total": 63.81561017036438, "reward/total": 0.7509765625}
+{"step": 24, "progress/batch": 24, "optim/lr": 4e-05, "progress/done_frac": 0.43103448275862066, "skyrl.ai/grad_norm": 4038.148461857241, "skyrl.ai/learning_rate": 4e-05, "time/total": 67.94123959541321, "reward/total": 0.66845703125}
+{"step": 25, "progress/batch": 25, "optim/lr": 4e-05, "progress/done_frac": 0.4482758620689655, "skyrl.ai/grad_norm": 4261.364570181716, "skyrl.ai/learning_rate": 4e-05, "time/total": 65.61149597167969, "reward/total": 0.7451171875}
+{"step": 26, "progress/batch": 26, "optim/lr": 4e-05, "progress/done_frac": 0.46551724137931033, "skyrl.ai/grad_norm": 3908.3854722890374, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.79335403442383, "reward/total": 0.73046875}
+{"step": 27, "progress/batch": 27, "optim/lr": 4e-05, "progress/done_frac": 0.4827586206896552, "skyrl.ai/grad_norm": 4198.391834976816, "skyrl.ai/learning_rate": 4e-05, "time/total": 64.1847505569458, "reward/total": 0.728515625}
+{"step": 28, "progress/batch": 28, "optim/lr": 4e-05, "progress/done_frac": 0.5, "skyrl.ai/grad_norm": 3855.95785765353, "skyrl.ai/learning_rate": 4e-05, "time/total": 67.56212735176086, "reward/total": 0.759765625}
+{"step": 29, "progress/batch": 29, "optim/lr": 4e-05, "progress/done_frac": 0.5172413793103449, "skyrl.ai/grad_norm": 3629.153482563117, "skyrl.ai/learning_rate": 4e-05, "time/total": 64.31682705879211, "reward/total": 0.775390625}
+{"step": 30, "progress/batch": 30, "optim/lr": 4e-05, "progress/done_frac": 0.5344827586206896, "skyrl.ai/grad_norm": 4626.5872951885385, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.43158459663391, "reward/total": 0.7421875}
+{"step": 31, "progress/batch": 31, "optim/lr": 4e-05, "progress/done_frac": 0.5517241379310345, "skyrl.ai/grad_norm": 3172.9867632878645, "skyrl.ai/learning_rate": 4e-05, "time/total": 61.603456258773804, "reward/total": 0.77978515625}
+{"step": 32, "progress/batch": 32, "optim/lr": 4e-05, "progress/done_frac": 0.5689655172413793, "skyrl.ai/grad_norm": 4532.925765992644, "skyrl.ai/learning_rate": 4e-05, "time/total": 70.21588969230652, "reward/total": 0.66455078125}
+{"step": 33, "progress/batch": 33, "optim/lr": 4e-05, "progress/done_frac": 0.5862068965517241, "skyrl.ai/grad_norm": 4165.589514102416, "skyrl.ai/learning_rate": 4e-05, "time/total": 65.48690557479858, "reward/total": 0.7822265625}
+{"step": 34, "progress/batch": 34, "optim/lr": 4e-05, "progress/done_frac": 0.603448275862069, "skyrl.ai/grad_norm": 4447.996852516872, "skyrl.ai/learning_rate": 4e-05, "time/total": 73.74556350708008, "reward/total": 0.71044921875}
+{"step": 35, "progress/batch": 35, "optim/lr": 4e-05, "progress/done_frac": 0.6206896551724138, "skyrl.ai/grad_norm": 4060.280532180012, "skyrl.ai/learning_rate": 4e-05, "time/total": 67.71312117576599, "reward/total": 0.71923828125}
+{"step": 36, "progress/batch": 36, "optim/lr": 4e-05, "progress/done_frac": 0.6379310344827587, "skyrl.ai/grad_norm": 4412.181546582144, "skyrl.ai/learning_rate": 4e-05, "time/total": 64.37158989906311, "reward/total": 0.72802734375}
+{"step": 37, "progress/batch": 37, "optim/lr": 4e-05, "progress/done_frac": 0.6551724137931034, "skyrl.ai/grad_norm": 3785.7107654970155, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.43695116043091, "reward/total": 0.74072265625}
+{"step": 38, "progress/batch": 38, "optim/lr": 4e-05, "progress/done_frac": 0.6724137931034483, "skyrl.ai/grad_norm": 4860.556141019256, "skyrl.ai/learning_rate": 4e-05, "time/total": 69.2902045249939, "reward/total": 0.6962890625}
+{"step": 39, "progress/batch": 39, "optim/lr": 4e-05, "progress/done_frac": 0.6896551724137931, "skyrl.ai/grad_norm": 4201.45189190594, "skyrl.ai/learning_rate": 4e-05, "time/total": 63.281941175460815, "reward/total": 0.77490234375}
+{"step": 40, "progress/batch": 40, "optim/lr": 4e-05, "progress/done_frac": 0.7068965517241379, "skyrl.ai/grad_norm": 4512.850097222375, "skyrl.ai/learning_rate": 4e-05, "time/total": 67.5277771949768, "reward/total": 0.73681640625}
+{"step": 41, "progress/batch": 41, "optim/lr": 4e-05, "progress/done_frac": 0.7241379310344828, "skyrl.ai/grad_norm": 4640.554923713326, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.97787046432495, "reward/total": 0.73681640625}
+{"step": 42, "progress/batch": 42, "optim/lr": 4e-05, "progress/done_frac": 0.7413793103448276, "skyrl.ai/grad_norm": 4721.049247783802, "skyrl.ai/learning_rate": 4e-05, "time/total": 68.46770405769348, "reward/total": 0.7314453125}
+{"step": 43, "progress/batch": 43, "optim/lr": 4e-05, "progress/done_frac": 0.7586206896551724, "skyrl.ai/grad_norm": 4301.910738265033, "skyrl.ai/learning_rate": 4e-05, "time/total": 67.84913516044617, "reward/total": 0.70361328125}
+{"step": 44, "progress/batch": 44, "optim/lr": 4e-05, "progress/done_frac": 0.7758620689655172, "skyrl.ai/grad_norm": 4144.882386751161, "skyrl.ai/learning_rate": 4e-05, "time/total": 65.65047812461853, "reward/total": 0.80810546875}
+{"step": 45, "progress/batch": 45, "optim/lr": 4e-05, "progress/done_frac": 0.7931034482758621, "skyrl.ai/grad_norm": 4613.070344141741, "skyrl.ai/learning_rate": 4e-05, "time/total": 63.60945105552673, "reward/total": 0.78564453125}
+{"step": 46, "progress/batch": 46, "optim/lr": 4e-05, "progress/done_frac": 0.8103448275862069, "skyrl.ai/grad_norm": 4161.3122930152695, "skyrl.ai/learning_rate": 4e-05, "time/total": 67.40388464927673, "reward/total": 0.73583984375}
+{"step": 47, "progress/batch": 47, "optim/lr": 4e-05, "progress/done_frac": 0.8275862068965517, "skyrl.ai/grad_norm": 3778.3566533613525, "skyrl.ai/learning_rate": 4e-05, "time/total": 63.2031409740448, "reward/total": 0.7939453125}
+{"step": 48, "progress/batch": 48, "optim/lr": 4e-05, "progress/done_frac": 0.8448275862068966, "skyrl.ai/grad_norm": 4912.503638675497, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.86501574516296, "reward/total": 0.7705078125}
+{"step": 49, "progress/batch": 49, "optim/lr": 4e-05, "progress/done_frac": 0.8620689655172413, "skyrl.ai/grad_norm": 4550.238894827391, "skyrl.ai/learning_rate": 4e-05, "time/total": 65.3466227054596, "reward/total": 0.69970703125}
+{"step": 50, "progress/batch": 50, "optim/lr": 4e-05, "progress/done_frac": 0.8793103448275862, "skyrl.ai/grad_norm": 5372.916154194108, "skyrl.ai/learning_rate": 4e-05, "time/total": 64.0033187866211, "reward/total": 0.73779296875}
+{"step": 51, "progress/batch": 51, "optim/lr": 4e-05, "progress/done_frac": 0.896551724137931, "skyrl.ai/grad_norm": 4387.628516636293, "skyrl.ai/learning_rate": 4e-05, "time/total": 64.9093267917633, "reward/total": 0.73193359375}
+{"step": 52, "progress/batch": 52, "optim/lr": 4e-05, "progress/done_frac": 0.9137931034482759, "skyrl.ai/grad_norm": 4861.48392983048, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.04316091537476, "reward/total": 0.69140625}
+{"step": 53, "progress/batch": 53, "optim/lr": 4e-05, "progress/done_frac": 0.9310344827586207, "skyrl.ai/grad_norm": 3259.830210302371, "skyrl.ai/learning_rate": 4e-05, "time/total": 62.28820252418518, "reward/total": 0.80908203125}
+{"step": 54, "progress/batch": 54, "optim/lr": 4e-05, "progress/done_frac": 0.9482758620689655, "skyrl.ai/grad_norm": 4360.2038943150355, "skyrl.ai/learning_rate": 4e-05, "time/total": 66.40438055992126, "reward/total": 0.71875}
+{"step": 55, "progress/batch": 55, "optim/lr": 4e-05, "progress/done_frac": 0.9655172413793104, "skyrl.ai/grad_norm": 4557.082619395878, "skyrl.ai/learning_rate": 4e-05, "time/total": 65.91692686080933, "reward/total": 0.78271484375}
+{"step": 56, "progress/batch": 56, "optim/lr": 4e-05, "progress/done_frac": 0.9827586206896551, "skyrl.ai/grad_norm": 3941.06889054226, "skyrl.ai/learning_rate": 4e-05, "time/total": 67.03757810592651, "reward/total": 0.79052734375}
+{"step": 57, "progress/batch": 57, "optim/lr": 4e-05, "progress/done_frac": 1.0, "skyrl.ai/grad_norm": 4977.27756911346, "skyrl.ai/learning_rate": 4e-05, "time/total": 67.70693683624268, "reward/total": 0.7392578125}
diff --git a/tests/tinker/smoke_logs/rl_loop_h.log/checkpoints.jsonl b/tests/tinker/smoke_logs/rl_loop_h.log/checkpoints.jsonl
new file mode 100644
index 0000000000..03afd02453
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_h.log/checkpoints.jsonl
@@ -0,0 +1 @@
+{"name": "000020", "batch": 20, "state_path": "tinker://model_388eefaf/weights/000020"}
diff --git a/tests/tinker/smoke_logs/rl_loop_h.log/code.diff b/tests/tinker/smoke_logs/rl_loop_h.log/code.diff
new file mode 100644
index 0000000000..d8e559755f
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_h.log/code.diff
@@ -0,0 +1,3 @@
+### repo: /home/etang/tinker-cookbook @ dacb835b31fb3f7a012851a1bc64c950d0495de3
+modules: tinker_cookbook
+(no local changes)
diff --git a/tests/tinker/smoke_logs/rl_loop_h.log/config.json b/tests/tinker/smoke_logs/rl_loop_h.log/config.json
new file mode 100644
index 0000000000..fee60c6b68
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_h.log/config.json
@@ -0,0 +1,12 @@
+{
+  "base_url": "http://localhost:8000",
+  "log_path": "/tmp/rl_loop_h.log",
+  "model_name": "Qwen/Qwen3-0.6B",
+  "batch_size": 128,
+  "group_size": 16,
+  "learning_rate": 4e-05,
+  "lora_rank": 32,
+  "save_every": 20,
+  "max_tokens": 256,
+  "ttl_seconds": 604800
+}
\ No newline at end of file
diff --git a/tests/tinker/smoke_logs/rl_loop_h.log/logs.log b/tests/tinker/smoke_logs/rl_loop_h.log/logs.log
new file mode 100644
index 0000000000..bf5aebff3f
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_h.log/logs.log
@@ -0,0 +1,430 @@
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/rl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B lora_rank=32 log_path=/tmp/rl_loop_h.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/rl_loop_h.log
+__main__:73 [INFO] Using renderer: qwen3
+__main__:76 [INFO] Loading dataset...
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_70c19c3c
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/rl_loop_h.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/rl_loop_h.log
+tinker.lib.public_interfaces.service_client:159 [INFO] TrainingClient initialized for model model_388eefaf
+__main__:121 [INFO] Training for 58 batches
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 0                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 0           │
+│ progress/done_frac     │ 0.017241    │
+│ reward/total           │ 0.017578    │
+│ skyrl.ai/grad_norm     │ 2986.160076 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 199.903373  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 1                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 1           │
+│ progress/done_frac     │ 0.034483    │
+│ reward/total           │ 0.027832    │
+│ skyrl.ai/grad_norm     │ 3232.538322 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 141.800122  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 2                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 2           │
+│ progress/done_frac     │ 0.051724    │
+│ reward/total           │ 0.028809    │
+│ skyrl.ai/grad_norm     │ 2605.609526 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 98.730768   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 3                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 3           │
+│ progress/done_frac     │ 0.068966    │
+│ reward/total           │ 0.066895    │
+│ skyrl.ai/grad_norm     │ 2888.340873 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 148.096477  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 4                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 4           │
+│ progress/done_frac     │ 0.086207    │
+│ reward/total           │ 0.193848    │
+│ skyrl.ai/grad_norm     │ 6988.557505 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 218.452928  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 5                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 5           │
+│ progress/done_frac     │ 0.103448    │
+│ reward/total           │ 0.339355    │
+│ skyrl.ai/grad_norm     │ 9294.383680 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 162.523287  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 6                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 6           │
+│ progress/done_frac     │ 0.120690    │
+│ reward/total           │ 0.389648    │
+│ skyrl.ai/grad_norm     │ 7166.919282 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 165.164537  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 7                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 7           │
+│ progress/done_frac     │ 0.137931    │
+│ reward/total           │ 0.520996    │
+│ skyrl.ai/grad_norm     │ 7366.206351 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 152.435220  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 8                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 8           │
+│ progress/done_frac     │ 0.155172    │
+│ reward/total           │ 0.509277    │
+│ skyrl.ai/grad_norm     │ 5313.993414 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 140.352420  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 9                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 9           │
+│ progress/done_frac     │ 0.172414    │
+│ reward/total           │ 0.583984    │
+│ skyrl.ai/grad_norm     │ 4483.872210 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 124.565117  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 10                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 10          │
+│ progress/done_frac     │ 0.189655    │
+│ reward/total           │ 0.573730    │
+│ skyrl.ai/grad_norm     │ 4581.698157 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 131.392579  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 11                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 11          │
+│ progress/done_frac     │ 0.206897    │
+│ reward/total           │ 0.619629    │
+│ skyrl.ai/grad_norm     │ 5430.633481 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 122.801665  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 12                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 12          │
+│ progress/done_frac     │ 0.224138    │
+│ reward/total           │ 0.562012    │
+│ skyrl.ai/grad_norm     │ 4113.346083 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 128.322746  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 13                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 13          │
+│ progress/done_frac     │ 0.241379    │
+│ reward/total           │ 0.614746    │
+│ skyrl.ai/grad_norm     │ 5131.263392 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 121.569723  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 14                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 14          │
+│ progress/done_frac     │ 0.258621    │
+│ reward/total           │ 0.566406    │
+│ skyrl.ai/grad_norm     │ 5241.210738 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 128.441487  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 15                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 15          │
+│ progress/done_frac     │ 0.275862    │
+│ reward/total           │ 0.619141    │
+│ skyrl.ai/grad_norm     │ 4036.870446 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 129.937497  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 16                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 16          │
+│ progress/done_frac     │ 0.293103    │
+│ reward/total           │ 0.637695    │
+│ skyrl.ai/grad_norm     │ 5042.480342 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 121.023461  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 17                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 17          │
+│ progress/done_frac     │ 0.310345    │
+│ reward/total           │ 0.656250    │
+│ skyrl.ai/grad_norm     │ 4292.369276 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 135.823786  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 18                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 18          │
+│ progress/done_frac     │ 0.327586    │
+│ reward/total           │ 0.619629    │
+│ skyrl.ai/grad_norm     │ 4199.793566 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 112.505669  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 19                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 19          │
+│ progress/done_frac     │ 0.344828    │
+│ reward/total           │ 0.663086    │
+│ skyrl.ai/grad_norm     │ 4492.620616 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 133.195554  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_388eefaf/weights/000020'}
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 20                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 20          │
+│ progress/done_frac     │ 0.362069    │
+│ reward/total           │ 0.673340    │
+│ skyrl.ai/grad_norm     │ 5127.105421 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 133.592353  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 21                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 21          │
+│ progress/done_frac     │ 0.379310    │
+│ reward/total           │ 0.697266    │
+│ skyrl.ai/grad_norm     │ 4158.499729 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 110.293463  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 22                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 22          │
+│ progress/done_frac     │ 0.396552    │
+│ reward/total           │ 0.699707    │
+│ skyrl.ai/grad_norm     │ 5291.736199 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 127.729033  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 23                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 23          │
+│ progress/done_frac     │ 0.413793    │
+│ reward/total           │ 0.701172    │
+│ skyrl.ai/grad_norm     │ 5152.130239 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 118.040159  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 24                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 24          │
+│ progress/done_frac     │ 0.431034    │
+│ reward/total           │ 0.630859    │
+│ skyrl.ai/grad_norm     │ 4435.361090 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 136.918935  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 25                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 25          │
+│ progress/done_frac     │ 0.448276    │
+│ reward/total           │ 0.686523    │
+│ skyrl.ai/grad_norm     │ 5201.739901 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 125.109281  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 26                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 26          │
+│ progress/done_frac     │ 0.465517    │
+│ reward/total           │ 0.705566    │
+│ skyrl.ai/grad_norm     │ 4135.996132 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 124.709949  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 27                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 27          │
+│ progress/done_frac     │ 0.482759    │
+│ reward/total           │ 0.729980    │
+│ skyrl.ai/grad_norm     │ 4415.818384 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 131.896928  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 28                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 28          │
+│ progress/done_frac     │ 0.500000    │
+│ reward/total           │ 0.760742    │
+│ skyrl.ai/grad_norm     │ 4766.756339 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 133.826097  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_h.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 29                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 29          │
+│ progress/done_frac     │ 0.517241    │
+│ reward/total           │ 0.760254    │
+│ skyrl.ai/grad_norm     │ 3994.311831 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 126.998264  │
+└────────────────────────┴─────────────┘
diff --git a/tests/tinker/smoke_logs/rl_loop_h.log/metrics.jsonl b/tests/tinker/smoke_logs/rl_loop_h.log/metrics.jsonl
new file mode 100644
index 0000000000..9e72712816
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_h.log/metrics.jsonl
@@ -0,0 +1,30 @@
+{"step": 0, "progress/batch": 0, "optim/lr": 4e-05, "progress/done_frac": 0.017241379310344827, "skyrl.ai/grad_norm": 2986.160076084335, "skyrl.ai/learning_rate": 4e-05, "time/total": 199.90337252616882, "reward/total": 0.017578125}
+{"step": 1, "progress/batch": 1, "optim/lr": 4e-05, "progress/done_frac": 0.034482758620689655, "skyrl.ai/grad_norm": 3232.5383215052534, "skyrl.ai/learning_rate": 4e-05, "time/total": 141.80012226104736, "reward/total": 0.02783203125}
+{"step": 2, "progress/batch": 2, "optim/lr": 4e-05, "progress/done_frac": 0.05172413793103448, "skyrl.ai/grad_norm": 2605.6095256196772, "skyrl.ai/learning_rate": 4e-05, "time/total": 98.73076844215393, "reward/total": 0.02880859375}
+{"step": 3, "progress/batch": 3, "optim/lr": 4e-05, "progress/done_frac": 0.06896551724137931, "skyrl.ai/grad_norm": 2888.3408732350135, "skyrl.ai/learning_rate": 4e-05, "time/total": 148.09647703170776, "reward/total": 0.06689453125}
+{"step": 4, "progress/batch": 4, "optim/lr": 4e-05, "progress/done_frac": 0.08620689655172414, "skyrl.ai/grad_norm": 6988.557504950503, "skyrl.ai/learning_rate": 4e-05, "time/total": 218.45292806625366, "reward/total": 0.19384765625}
+{"step": 5, "progress/batch": 5, "optim/lr": 4e-05, "progress/done_frac": 0.10344827586206896, "skyrl.ai/grad_norm": 9294.383680481455, "skyrl.ai/learning_rate": 4e-05, "time/total": 162.52328658103943, "reward/total": 0.33935546875}
+{"step": 6, "progress/batch": 6, "optim/lr": 4e-05, "progress/done_frac": 0.1206896551724138, "skyrl.ai/grad_norm": 7166.9192823695175, "skyrl.ai/learning_rate": 4e-05, "time/total": 165.1645369529724, "reward/total": 0.3896484375}
+{"step": 7, "progress/batch": 7, "optim/lr": 4e-05, "progress/done_frac": 0.13793103448275862, "skyrl.ai/grad_norm": 7366.206350625809, "skyrl.ai/learning_rate": 4e-05, "time/total": 152.43522000312805, "reward/total": 0.52099609375}
+{"step": 8, "progress/batch": 8, "optim/lr": 4e-05, "progress/done_frac": 0.15517241379310345, "skyrl.ai/grad_norm": 5313.9934136203065, "skyrl.ai/learning_rate": 4e-05, "time/total": 140.35242009162903, "reward/total": 0.50927734375}
+{"step": 9, "progress/batch": 9, "optim/lr": 4e-05, "progress/done_frac": 0.1724137931034483, "skyrl.ai/grad_norm": 4483.8722104895005, "skyrl.ai/learning_rate": 4e-05, "time/total": 124.56511735916138, "reward/total": 0.583984375}
+{"step": 10, "progress/batch": 10, "optim/lr": 4e-05, "progress/done_frac": 0.1896551724137931, "skyrl.ai/grad_norm": 4581.698156797324, "skyrl.ai/learning_rate": 4e-05, "time/total": 131.39257907867432, "reward/total": 0.57373046875}
+{"step": 11, "progress/batch": 11, "optim/lr": 4e-05, "progress/done_frac": 0.20689655172413793, "skyrl.ai/grad_norm": 5430.633480543499, "skyrl.ai/learning_rate": 4e-05, "time/total": 122.80166530609131, "reward/total": 0.61962890625}
+{"step": 12, "progress/batch": 12, "optim/lr": 4e-05, "progress/done_frac": 0.22413793103448276, "skyrl.ai/grad_norm": 4113.346083178511, "skyrl.ai/learning_rate": 4e-05, "time/total": 128.32274556159973, "reward/total": 0.56201171875}
+{"step": 13, "progress/batch": 13, "optim/lr": 4e-05, "progress/done_frac": 0.2413793103448276, "skyrl.ai/grad_norm": 5131.263392187152, "skyrl.ai/learning_rate": 4e-05, "time/total": 121.56972336769104, "reward/total": 0.61474609375}
+{"step": 14, "progress/batch": 14, "optim/lr": 4e-05, "progress/done_frac": 0.25862068965517243, "skyrl.ai/grad_norm": 5241.210737987932, "skyrl.ai/learning_rate": 4e-05, "time/total": 128.4414873123169, "reward/total": 0.56640625}
+{"step": 15, "progress/batch": 15, "optim/lr": 4e-05, "progress/done_frac": 0.27586206896551724, "skyrl.ai/grad_norm": 4036.870446273945, "skyrl.ai/learning_rate": 4e-05, "time/total": 129.9374966621399, "reward/total": 0.619140625}
+{"step": 16, "progress/batch": 16, "optim/lr": 4e-05, "progress/done_frac": 0.29310344827586204, "skyrl.ai/grad_norm": 5042.480342053898, "skyrl.ai/learning_rate": 4e-05, "time/total": 121.02346110343933, "reward/total": 0.6376953125}
+{"step": 17, "progress/batch": 17, "optim/lr": 4e-05, "progress/done_frac": 0.3103448275862069, "skyrl.ai/grad_norm": 4292.369275819591, "skyrl.ai/learning_rate": 4e-05, "time/total": 135.82378602027893, "reward/total": 0.65625}
+{"step": 18, "progress/batch": 18, "optim/lr": 4e-05, "progress/done_frac": 0.3275862068965517, "skyrl.ai/grad_norm": 4199.793566355375, "skyrl.ai/learning_rate": 4e-05, "time/total": 112.50566911697388, "reward/total": 0.61962890625}
+{"step": 19, "progress/batch": 19, "optim/lr": 4e-05, "progress/done_frac": 0.3448275862068966, "skyrl.ai/grad_norm": 4492.620616076991, "skyrl.ai/learning_rate": 4e-05, "time/total": 133.19555377960205, "reward/total": 0.6630859375}
+{"step": 20, "progress/batch": 20, "optim/lr": 4e-05, "progress/done_frac": 0.3620689655172414, "skyrl.ai/grad_norm": 5127.105421190401, "skyrl.ai/learning_rate": 4e-05, "time/total": 133.59235310554504, "reward/total": 0.67333984375}
+{"step": 21, "progress/batch": 21, "optim/lr": 4e-05, "progress/done_frac": 0.3793103448275862, "skyrl.ai/grad_norm": 4158.499729469752, "skyrl.ai/learning_rate": 4e-05, "time/total": 110.2934627532959, "reward/total": 0.697265625}
+{"step": 22, "progress/batch": 22, "optim/lr": 4e-05, "progress/done_frac": 0.39655172413793105, "skyrl.ai/grad_norm": 5291.736199018239, "skyrl.ai/learning_rate": 4e-05, "time/total": 127.72903275489807, "reward/total": 0.69970703125}
+{"step": 23, "progress/batch": 23, "optim/lr": 4e-05, "progress/done_frac": 0.41379310344827586, "skyrl.ai/grad_norm": 5152.130239037053, "skyrl.ai/learning_rate": 4e-05, "time/total": 118.04015898704529, "reward/total": 0.701171875}
+{"step": 24, "progress/batch": 24, "optim/lr": 4e-05, "progress/done_frac": 0.43103448275862066, "skyrl.ai/grad_norm": 4435.361090148129, "skyrl.ai/learning_rate": 4e-05, "time/total": 136.91893458366394, "reward/total": 0.630859375}
+{"step": 25, "progress/batch": 25, "optim/lr": 4e-05, "progress/done_frac": 0.4482758620689655, "skyrl.ai/grad_norm": 5201.739901225358, "skyrl.ai/learning_rate": 4e-05, "time/total": 125.10928058624268, "reward/total": 0.6865234375}
+{"step": 26, "progress/batch": 26, "optim/lr": 4e-05, "progress/done_frac": 0.46551724137931033, "skyrl.ai/grad_norm": 4135.996131526237, "skyrl.ai/learning_rate": 4e-05, "time/total": 124.70994901657104, "reward/total": 0.70556640625}
+{"step": 27, "progress/batch": 27, "optim/lr": 4e-05, "progress/done_frac": 0.4827586206896552, "skyrl.ai/grad_norm": 4415.818383946514, "skyrl.ai/learning_rate": 4e-05, "time/total": 131.89692759513855, "reward/total": 0.72998046875}
+{"step": 28, "progress/batch": 28, "optim/lr": 4e-05, "progress/done_frac": 0.5, "skyrl.ai/grad_norm": 4766.756339482857, "skyrl.ai/learning_rate": 4e-05, "time/total": 133.826096534729, "reward/total": 0.7607421875}
+{"step": 29, "progress/batch": 29, "optim/lr": 4e-05, "progress/done_frac": 0.5172413793103449, "skyrl.ai/grad_norm": 3994.311830591097, "skyrl.ai/learning_rate": 4e-05, "time/total": 126.9982635974884, "reward/total": 0.76025390625}
diff --git a/tests/tinker/smoke_logs/rl_loop_i.log/checkpoints.jsonl b/tests/tinker/smoke_logs/rl_loop_i.log/checkpoints.jsonl
new file mode 100644
index 0000000000..f33bedb477
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_i.log/checkpoints.jsonl
@@ -0,0 +1 @@
+{"name": "000020", "batch": 20, "state_path": "tinker://model_8d63bf6f/weights/000020"}
diff --git a/tests/tinker/smoke_logs/rl_loop_i.log/code.diff b/tests/tinker/smoke_logs/rl_loop_i.log/code.diff
new file mode 100644
index 0000000000..d8e559755f
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_i.log/code.diff
@@ -0,0 +1,3 @@
+### repo: /home/etang/tinker-cookbook @ dacb835b31fb3f7a012851a1bc64c950d0495de3
+modules: tinker_cookbook
+(no local changes)
diff --git a/tests/tinker/smoke_logs/rl_loop_i.log/config.json b/tests/tinker/smoke_logs/rl_loop_i.log/config.json
new file mode 100644
index 0000000000..ce74455801
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_i.log/config.json
@@ -0,0 +1,12 @@
+{
+  "base_url": "http://localhost:8000",
+  "log_path": "/tmp/rl_loop_i.log",
+  "model_name": "Qwen/Qwen3-0.6B",
+  "batch_size": 128,
+  "group_size": 16,
+  "learning_rate": 4e-05,
+  "lora_rank": 32,
+  "save_every": 20,
+  "max_tokens": 256,
+  "ttl_seconds": 604800
+}
\ No newline at end of file
diff --git a/tests/tinker/smoke_logs/rl_loop_i.log/logs.log b/tests/tinker/smoke_logs/rl_loop_i.log/logs.log
new file mode 100644
index 0000000000..08f955b1db
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_i.log/logs.log
@@ -0,0 +1,430 @@
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/rl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B lora_rank=32 log_path=/tmp/rl_loop_i.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/rl_loop_i.log
+__main__:73 [INFO] Using renderer: qwen3
+__main__:76 [INFO] Loading dataset...
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_1f83fcde
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/rl_loop_i.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/rl_loop_i.log
+tinker.lib.public_interfaces.service_client:159 [INFO] TrainingClient initialized for model model_8d63bf6f
+__main__:121 [INFO] Training for 58 batches
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 0                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 0           │
+│ progress/done_frac     │ 0.017241    │
+│ reward/total           │ 0.016602    │
+│ skyrl.ai/grad_norm     │ 2500.282184 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 92.174990   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 1                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 1           │
+│ progress/done_frac     │ 0.034483    │
+│ reward/total           │ 0.028809    │
+│ skyrl.ai/grad_norm     │ 3717.477236 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 212.645627  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 2                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 2           │
+│ progress/done_frac     │ 0.051724    │
+│ reward/total           │ 0.026855    │
+│ skyrl.ai/grad_norm     │ 3123.280167 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 154.771824  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 3                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 3           │
+│ progress/done_frac     │ 0.068966    │
+│ reward/total           │ 0.060059    │
+│ skyrl.ai/grad_norm     │ 5379.627311 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 151.328907  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 4                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 4           │
+│ progress/done_frac     │ 0.086207    │
+│ reward/total           │ 0.203613    │
+│ skyrl.ai/grad_norm     │ 7820.956975 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 165.573668  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 5                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 5           │
+│ progress/done_frac     │ 0.103448    │
+│ reward/total           │ 0.337891    │
+│ skyrl.ai/grad_norm     │ 9302.311541 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 161.446375  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 6                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 6           │
+│ progress/done_frac     │ 0.120690    │
+│ reward/total           │ 0.383789    │
+│ skyrl.ai/grad_norm     │ 9008.423613 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 158.094930  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 7                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 7           │
+│ progress/done_frac     │ 0.137931    │
+│ reward/total           │ 0.522461    │
+│ skyrl.ai/grad_norm     │ 6418.265498 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 150.219928  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 8                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 8           │
+│ progress/done_frac     │ 0.155172    │
+│ reward/total           │ 0.495117    │
+│ skyrl.ai/grad_norm     │ 5129.178102 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 130.450021  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                 Step 9                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 9           │
+│ progress/done_frac     │ 0.172414    │
+│ reward/total           │ 0.578125    │
+│ skyrl.ai/grad_norm     │ 5509.330994 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 126.062569  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 10                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 10          │
+│ progress/done_frac     │ 0.189655    │
+│ reward/total           │ 0.578125    │
+│ skyrl.ai/grad_norm     │ 5093.188589 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 128.825426  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 11                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 11          │
+│ progress/done_frac     │ 0.206897    │
+│ reward/total           │ 0.601074    │
+│ skyrl.ai/grad_norm     │ 5185.033462 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 121.206484  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 12                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 12          │
+│ progress/done_frac     │ 0.224138    │
+│ reward/total           │ 0.561035    │
+│ skyrl.ai/grad_norm     │ 4451.828838 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 123.993158  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 13                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 13          │
+│ progress/done_frac     │ 0.241379    │
+│ reward/total           │ 0.602539    │
+│ skyrl.ai/grad_norm     │ 5261.430984 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 129.587321  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 14                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 14          │
+│ progress/done_frac     │ 0.258621    │
+│ reward/total           │ 0.562500    │
+│ skyrl.ai/grad_norm     │ 4213.133751 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 130.133244  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 15                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 15          │
+│ progress/done_frac     │ 0.275862    │
+│ reward/total           │ 0.609863    │
+│ skyrl.ai/grad_norm     │ 4431.559545 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 119.718005  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 16                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 16          │
+│ progress/done_frac     │ 0.293103    │
+│ reward/total           │ 0.634766    │
+│ skyrl.ai/grad_norm     │ 4251.883818 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 131.653447  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 17                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 17          │
+│ progress/done_frac     │ 0.310345    │
+│ reward/total           │ 0.654785    │
+│ skyrl.ai/grad_norm     │ 4336.820264 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 127.127751  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 18                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 18          │
+│ progress/done_frac     │ 0.327586    │
+│ reward/total           │ 0.601074    │
+│ skyrl.ai/grad_norm     │ 4390.180634 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 119.283090  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 19                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 19          │
+│ progress/done_frac     │ 0.344828    │
+│ reward/total           │ 0.656250    │
+│ skyrl.ai/grad_norm     │ 5138.570229 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 87.947657   │
+└────────────────────────┴─────────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_8d63bf6f/weights/000020'}
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 20                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 20          │
+│ progress/done_frac     │ 0.362069    │
+│ reward/total           │ 0.652832    │
+│ skyrl.ai/grad_norm     │ 4787.003656 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 171.219129  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 21                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 21          │
+│ progress/done_frac     │ 0.379310    │
+│ reward/total           │ 0.664551    │
+│ skyrl.ai/grad_norm     │ 5361.994405 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 116.019150  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 22                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 22          │
+│ progress/done_frac     │ 0.396552    │
+│ reward/total           │ 0.705566    │
+│ skyrl.ai/grad_norm     │ 4536.794683 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 119.559554  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 23                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 23          │
+│ progress/done_frac     │ 0.413793    │
+│ reward/total           │ 0.707520    │
+│ skyrl.ai/grad_norm     │ 3972.942738 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 131.468635  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 24                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 24          │
+│ progress/done_frac     │ 0.431034    │
+│ reward/total           │ 0.628418    │
+│ skyrl.ai/grad_norm     │ 4835.404637 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 126.411474  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 25                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 25          │
+│ progress/done_frac     │ 0.448276    │
+│ reward/total           │ 0.701660    │
+│ skyrl.ai/grad_norm     │ 4455.773782 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 131.987321  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 26                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 26          │
+│ progress/done_frac     │ 0.465517    │
+│ reward/total           │ 0.688477    │
+│ skyrl.ai/grad_norm     │ 6312.734431 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 126.180908  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 27                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 27          │
+│ progress/done_frac     │ 0.482759    │
+│ reward/total           │ 0.716797    │
+│ skyrl.ai/grad_norm     │ 4511.742679 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 133.081262  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 28                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 28          │
+│ progress/done_frac     │ 0.500000    │
+│ reward/total           │ 0.748535    │
+│ skyrl.ai/grad_norm     │ 4176.975700 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 125.895678  │
+└────────────────────────┴─────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/rl_loop_i.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 29                 
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value       ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
+│ optim/lr               │ 0.000040    │
+│ progress/batch         │ 29          │
+│ progress/done_frac     │ 0.517241    │
+│ reward/total           │ 0.733887    │
+│ skyrl.ai/grad_norm     │ 4329.369700 │
+│ skyrl.ai/learning_rate │ 0.000040    │
+│ time/total             │ 136.384873  │
+└────────────────────────┴─────────────┘
diff --git a/tests/tinker/smoke_logs/rl_loop_i.log/metrics.jsonl b/tests/tinker/smoke_logs/rl_loop_i.log/metrics.jsonl
new file mode 100644
index 0000000000..e1ad5a641f
--- /dev/null
+++ b/tests/tinker/smoke_logs/rl_loop_i.log/metrics.jsonl
@@ -0,0 +1,30 @@
+{"step": 0, "progress/batch": 0, "optim/lr": 4e-05, "progress/done_frac": 0.017241379310344827, "skyrl.ai/grad_norm": 2500.2821840744296, "skyrl.ai/learning_rate": 4e-05, "time/total": 92.17498970031738, "reward/total": 0.0166015625}
+{"step": 1, "progress/batch": 1, "optim/lr": 4e-05, "progress/done_frac": 0.034482758620689655, "skyrl.ai/grad_norm": 3717.4772359760323, "skyrl.ai/learning_rate": 4e-05, "time/total": 212.64562678337097, "reward/total": 0.02880859375}
+{"step": 2, "progress/batch": 2, "optim/lr": 4e-05, "progress/done_frac": 0.05172413793103448, "skyrl.ai/grad_norm": 3123.2801667477734, "skyrl.ai/learning_rate": 4e-05, "time/total": 154.7718243598938, "reward/total": 0.02685546875}
+{"step": 3, "progress/batch": 3, "optim/lr": 4e-05, "progress/done_frac": 0.06896551724137931, "skyrl.ai/grad_norm": 5379.627310511389, "skyrl.ai/learning_rate": 4e-05, "time/total": 151.3289074897766, "reward/total": 0.06005859375}
+{"step": 4, "progress/batch": 4, "optim/lr": 4e-05, "progress/done_frac": 0.08620689655172414, "skyrl.ai/grad_norm": 7820.956974693058, "skyrl.ai/learning_rate": 4e-05, "time/total": 165.57366752624512, "reward/total": 0.20361328125}
+{"step": 5, "progress/batch": 5, "optim/lr": 4e-05, "progress/done_frac": 0.10344827586206896, "skyrl.ai/grad_norm": 9302.311540687078, "skyrl.ai/learning_rate": 4e-05, "time/total": 161.44637537002563, "reward/total": 0.337890625}
+{"step": 6, "progress/batch": 6, "optim/lr": 4e-05, "progress/done_frac": 0.1206896551724138, "skyrl.ai/grad_norm": 9008.423613485325, "skyrl.ai/learning_rate": 4e-05, "time/total": 158.09492993354797, "reward/total": 0.3837890625}
+{"step": 7, "progress/batch": 7, "optim/lr": 4e-05, "progress/done_frac": 0.13793103448275862, "skyrl.ai/grad_norm": 6418.265497780533, "skyrl.ai/learning_rate": 4e-05, "time/total": 150.21992802619934, "reward/total": 0.5224609375}
+{"step": 8, "progress/batch": 8, "optim/lr": 4e-05, "progress/done_frac": 0.15517241379310345, "skyrl.ai/grad_norm": 5129.178101801496, "skyrl.ai/learning_rate": 4e-05, "time/total": 130.45002102851868, "reward/total": 0.4951171875}
+{"step": 9, "progress/batch": 9, "optim/lr": 4e-05, "progress/done_frac": 0.1724137931034483, "skyrl.ai/grad_norm": 5509.330993868493, "skyrl.ai/learning_rate": 4e-05, "time/total": 126.06256937980652, "reward/total": 0.578125}
+{"step": 10, "progress/batch": 10, "optim/lr": 4e-05, "progress/done_frac": 0.1896551724137931, "skyrl.ai/grad_norm": 5093.188588693727, "skyrl.ai/learning_rate": 4e-05, "time/total": 128.8254256248474, "reward/total": 0.578125}
+{"step": 11, "progress/batch": 11, "optim/lr": 4e-05, "progress/done_frac": 0.20689655172413793, "skyrl.ai/grad_norm": 5185.033461801379, "skyrl.ai/learning_rate": 4e-05, "time/total": 121.2064836025238, "reward/total": 0.60107421875}
+{"step": 12, "progress/batch": 12, "optim/lr": 4e-05, "progress/done_frac": 0.22413793103448276, "skyrl.ai/grad_norm": 4451.828837680083, "skyrl.ai/learning_rate": 4e-05, "time/total": 123.99315762519836, "reward/total": 0.56103515625}
+{"step": 13, "progress/batch": 13, "optim/lr": 4e-05, "progress/done_frac": 0.2413793103448276, "skyrl.ai/grad_norm": 5261.430984057474, "skyrl.ai/learning_rate": 4e-05, "time/total": 129.5873212814331, "reward/total": 0.6025390625}
+{"step": 14, "progress/batch": 14, "optim/lr": 4e-05, "progress/done_frac": 0.25862068965517243, "skyrl.ai/grad_norm": 4213.13375054721, "skyrl.ai/learning_rate": 4e-05, "time/total": 130.1332437992096, "reward/total": 0.5625}
+{"step": 15, "progress/batch": 15, "optim/lr": 4e-05, "progress/done_frac": 0.27586206896551724, "skyrl.ai/grad_norm": 4431.559544900643, "skyrl.ai/learning_rate": 4e-05, "time/total": 119.71800494194031, "reward/total": 0.60986328125}
+{"step": 16, "progress/batch": 16, "optim/lr": 4e-05, "progress/done_frac": 0.29310344827586204, "skyrl.ai/grad_norm": 4251.883817791827, "skyrl.ai/learning_rate": 4e-05, "time/total": 131.6534469127655, "reward/total": 0.634765625}
+{"step": 17, "progress/batch": 17, "optim/lr": 4e-05, "progress/done_frac": 0.3103448275862069, "skyrl.ai/grad_norm": 4336.820263741628, "skyrl.ai/learning_rate": 4e-05, "time/total": 127.1277506351471, "reward/total": 0.65478515625}
+{"step": 18, "progress/batch": 18, "optim/lr": 4e-05, "progress/done_frac": 0.3275862068965517, "skyrl.ai/grad_norm": 4390.1806340969615, "skyrl.ai/learning_rate": 4e-05, "time/total": 119.28309035301208, "reward/total": 0.60107421875}
+{"step": 19, "progress/batch": 19, "optim/lr": 4e-05, "progress/done_frac": 0.3448275862068966, "skyrl.ai/grad_norm": 5138.57022915908, "skyrl.ai/learning_rate": 4e-05, "time/total": 87.9476568698883, "reward/total": 0.65625}
+{"step": 20, "progress/batch": 20, "optim/lr": 4e-05, "progress/done_frac": 0.3620689655172414, "skyrl.ai/grad_norm": 4787.003655732884, "skyrl.ai/learning_rate": 4e-05, "time/total": 171.21912908554077, "reward/total": 0.65283203125}
+{"step": 21, "progress/batch": 21, "optim/lr": 4e-05, "progress/done_frac": 0.3793103448275862, "skyrl.ai/grad_norm": 5361.994405069815, "skyrl.ai/learning_rate": 4e-05, "time/total": 116.01915049552917, "reward/total": 0.66455078125}
+{"step": 22, "progress/batch": 22, "optim/lr": 4e-05, "progress/done_frac": 0.39655172413793105, "skyrl.ai/grad_norm": 4536.794683474226, "skyrl.ai/learning_rate": 4e-05, "time/total": 119.55955386161804, "reward/total": 0.70556640625}
+{"step": 23, "progress/batch": 23, "optim/lr": 4e-05, "progress/done_frac": 0.41379310344827586, "skyrl.ai/grad_norm": 3972.942738072121, "skyrl.ai/learning_rate": 4e-05, "time/total": 131.4686348438263, "reward/total": 0.70751953125}
+{"step": 24, "progress/batch": 24, "optim/lr": 4e-05, "progress/done_frac": 0.43103448275862066, "skyrl.ai/grad_norm": 4835.404636635904, "skyrl.ai/learning_rate": 4e-05, "time/total": 126.41147351264954, "reward/total": 0.62841796875}
+{"step": 25, "progress/batch": 25, "optim/lr": 4e-05, "progress/done_frac": 0.4482758620689655, "skyrl.ai/grad_norm": 4455.773782408618, "skyrl.ai/learning_rate": 4e-05, "time/total": 131.98732089996338, "reward/total": 0.70166015625}
+{"step": 26, "progress/batch": 26, "optim/lr": 4e-05, "progress/done_frac": 0.46551724137931033, "skyrl.ai/grad_norm": 6312.734431290453, "skyrl.ai/learning_rate": 4e-05, "time/total": 126.180908203125, "reward/total": 0.6884765625}
+{"step": 27, "progress/batch": 27, "optim/lr": 4e-05, "progress/done_frac": 0.4827586206896552, "skyrl.ai/grad_norm": 4511.742678832648, "skyrl.ai/learning_rate": 4e-05, "time/total": 133.08126187324524, "reward/total": 0.716796875}
+{"step": 28, "progress/batch": 28, "optim/lr": 4e-05, "progress/done_frac": 0.5, "skyrl.ai/grad_norm": 4176.975700192665, "skyrl.ai/learning_rate": 4e-05, "time/total": 125.8956778049469, "reward/total": 0.74853515625}
+{"step": 29, "progress/batch": 29, "optim/lr": 4e-05, "progress/done_frac": 0.5172413793103449, "skyrl.ai/grad_norm": 4329.369700083374, "skyrl.ai/learning_rate": 4e-05, "time/total": 136.38487315177917, "reward/total": 0.73388671875}
diff --git a/tests/tinker/smoke_logs/sl_loop_a.log/checkpoints.jsonl b/tests/tinker/smoke_logs/sl_loop_a.log/checkpoints.jsonl
new file mode 100644
index 0000000000..f33af574cf
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_a.log/checkpoints.jsonl
@@ -0,0 +1,2 @@
+{"name": "000020", "batch": 20, "state_path": "tinker://model_a6e27bdf/weights/000020"}
+{"name": "000040", "batch": 40, "state_path": "tinker://model_a6e27bdf/weights/000040"}
diff --git a/tests/tinker/smoke_logs/sl_loop_a.log/code.diff b/tests/tinker/smoke_logs/sl_loop_a.log/code.diff
new file mode 100644
index 0000000000..d8e559755f
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_a.log/code.diff
@@ -0,0 +1,3 @@
+### repo: /home/etang/tinker-cookbook @ dacb835b31fb3f7a012851a1bc64c950d0495de3
+modules: tinker_cookbook
+(no local changes)
diff --git a/tests/tinker/smoke_logs/sl_loop_a.log/config.json b/tests/tinker/smoke_logs/sl_loop_a.log/config.json
new file mode 100644
index 0000000000..a74eec435c
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_a.log/config.json
@@ -0,0 +1,12 @@
+{
+  "base_url": "http://localhost:8000",
+  "log_path": "/tmp/sl_loop_a.log",
+  "model_name": "Qwen/Qwen3-0.6B",
+  "batch_size": 128,
+  "learning_rate": 0.0001,
+  "max_length": 32768,
+  "train_on_what": "last_assistant_message",
+  "lora_rank": 32,
+  "save_every": 20,
+  "ttl_seconds": 604800
+}
\ No newline at end of file
diff --git a/tests/tinker/smoke_logs/sl_loop_a.log/logs.log b/tests/tinker/smoke_logs/sl_loop_a.log/logs.log
new file mode 100644
index 0000000000..83759c41ca
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_a.log/logs.log
@@ -0,0 +1,729 @@
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/sl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B train_on_what=LAST_ASSISTANT_MESSAGE lora_rank=32 log_path=/tmp/sl_loop_a.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/sl_loop_a.log
+__main__:51 [INFO] Using renderer: qwen3
+__main__:54 [INFO] Loading dataset...
+__main__:64 [INFO] Dropping last 28 examples to keep batch size uniform at 128
+__main__:67 [INFO] Train batches: 74
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_b2801142
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/sl_loop_a.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/sl_loop_a.log
+tinker.lib.telemetry:204 [INFO] Exception logged for session ID: session_b2801142
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/sl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B train_on_what=LAST_ASSISTANT_MESSAGE lora_rank=32 log_path=/tmp/sl_loop_a.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/sl_loop_a.log
+__main__:51 [INFO] Using renderer: qwen3
+__main__:54 [INFO] Loading dataset...
+__main__:64 [INFO] Dropping last 28 examples to keep batch size uniform at 128
+__main__:67 [INFO] Train batches: 74
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_29a17128
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/sl_loop_a.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/sl_loop_a.log
+tinker.lib.public_interfaces.service_client:159 [INFO] TrainingClient initialized for model model_a6e27bdf
+__main__:87 [INFO] Training for 74 steps
+tinker_cookbook.supervised.common:188 [INFO] Weight reduction: 'mean' (token-mean loss)
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 0                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000100  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37834     │
+│ progress               │ 0.000000  │
+│ skyrl.ai/grad_norm     │ 1.419505  │
+│ skyrl.ai/learning_rate │ 0.000100  │
+│ time_total             │ 87.547961 │
+│ train_mean_nll         │ 2.937314  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 1                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000099  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35654     │
+│ progress               │ 0.013514  │
+│ skyrl.ai/grad_norm     │ 1.757945  │
+│ skyrl.ai/learning_rate │ 0.000099  │
+│ time_total             │ 40.488416 │
+│ train_mean_nll         │ 3.141781  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 2                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000097  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38474     │
+│ progress               │ 0.027027  │
+│ skyrl.ai/grad_norm     │ 1.375920  │
+│ skyrl.ai/learning_rate │ 0.000097  │
+│ time_total             │ 40.696287 │
+│ train_mean_nll         │ 2.757976  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 3                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000096  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39781     │
+│ progress               │ 0.040541  │
+│ skyrl.ai/grad_norm     │ 0.992602  │
+│ skyrl.ai/learning_rate │ 0.000096  │
+│ time_total             │ 40.512256 │
+│ train_mean_nll         │ 2.664960  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 4                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000095  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37346     │
+│ progress               │ 0.054054  │
+│ skyrl.ai/grad_norm     │ 1.000068  │
+│ skyrl.ai/learning_rate │ 0.000095  │
+│ time_total             │ 76.539217 │
+│ train_mean_nll         │ 2.643194  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 5                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000093  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38710     │
+│ progress               │ 0.067568  │
+│ skyrl.ai/grad_norm     │ 0.864887  │
+│ skyrl.ai/learning_rate │ 0.000093  │
+│ time_total             │ 77.631970 │
+│ train_mean_nll         │ 2.497554  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 6                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000092  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33356     │
+│ progress               │ 0.081081  │
+│ skyrl.ai/grad_norm     │ 0.622200  │
+│ skyrl.ai/learning_rate │ 0.000092  │
+│ time_total             │ 76.481162 │
+│ train_mean_nll         │ 2.663911  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 7                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000091  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39115     │
+│ progress               │ 0.094595  │
+│ skyrl.ai/grad_norm     │ 0.546005  │
+│ skyrl.ai/learning_rate │ 0.000091  │
+│ time_total             │ 76.499547 │
+│ train_mean_nll         │ 2.479411  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 8                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000089  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38640     │
+│ progress               │ 0.108108  │
+│ skyrl.ai/grad_norm     │ 0.589701  │
+│ skyrl.ai/learning_rate │ 0.000089  │
+│ time_total             │ 76.493416 │
+│ train_mean_nll         │ 2.446202  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 9                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000088  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38863     │
+│ progress               │ 0.121622  │
+│ skyrl.ai/grad_norm     │ 0.481519  │
+│ skyrl.ai/learning_rate │ 0.000088  │
+│ time_total             │ 76.621217 │
+│ train_mean_nll         │ 2.381734  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 10                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000086  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33155     │
+│ progress               │ 0.135135  │
+│ skyrl.ai/grad_norm     │ 0.576803  │
+│ skyrl.ai/learning_rate │ 0.000086  │
+│ time_total             │ 76.480710 │
+│ train_mean_nll         │ 2.520661  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 11                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000085  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38406     │
+│ progress               │ 0.148649  │
+│ skyrl.ai/grad_norm     │ 0.641170  │
+│ skyrl.ai/learning_rate │ 0.000085  │
+│ time_total             │ 76.499414 │
+│ train_mean_nll         │ 2.459423  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 12                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000084  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38843     │
+│ progress               │ 0.162162  │
+│ skyrl.ai/grad_norm     │ 0.567629  │
+│ skyrl.ai/learning_rate │ 0.000084  │
+│ time_total             │ 76.519642 │
+│ train_mean_nll         │ 2.331488  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 13                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000082  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35583     │
+│ progress               │ 0.175676  │
+│ skyrl.ai/grad_norm     │ 0.533097  │
+│ skyrl.ai/learning_rate │ 0.000082  │
+│ time_total             │ 76.485970 │
+│ train_mean_nll         │ 2.370727  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 14                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000081  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 36717     │
+│ progress               │ 0.189189  │
+│ skyrl.ai/grad_norm     │ 0.442445  │
+│ skyrl.ai/learning_rate │ 0.000081  │
+│ time_total             │ 76.499796 │
+│ train_mean_nll         │ 2.284867  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 15                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000080  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34002     │
+│ progress               │ 0.202703  │
+│ skyrl.ai/grad_norm     │ 0.459619  │
+│ skyrl.ai/learning_rate │ 0.000080  │
+│ time_total             │ 76.715802 │
+│ train_mean_nll         │ 2.411974  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 16                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000078  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38620     │
+│ progress               │ 0.216216  │
+│ skyrl.ai/grad_norm     │ 0.343792  │
+│ skyrl.ai/learning_rate │ 0.000078  │
+│ time_total             │ 76.496818 │
+│ train_mean_nll         │ 2.385261  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 17                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000077  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37639     │
+│ progress               │ 0.229730  │
+│ skyrl.ai/grad_norm     │ 0.328837  │
+│ skyrl.ai/learning_rate │ 0.000077  │
+│ time_total             │ 75.498580 │
+│ train_mean_nll         │ 2.256649  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 18                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000076  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34118     │
+│ progress               │ 0.243243  │
+│ skyrl.ai/grad_norm     │ 0.384255  │
+│ skyrl.ai/learning_rate │ 0.000076  │
+│ time_total             │ 77.494641 │
+│ train_mean_nll         │ 2.406415  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 19                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000074  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37178     │
+│ progress               │ 0.256757  │
+│ skyrl.ai/grad_norm     │ 0.364690  │
+│ skyrl.ai/learning_rate │ 0.000074  │
+│ time_total             │ 75.483256 │
+│ train_mean_nll         │ 2.389909  │
+└────────────────────────┴───────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_a6e27bdf/weights/000020'}
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 20                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value      ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
+│ learning_rate          │ 0.000073   │
+│ num_sequences          │ 128        │
+│ num_tokens             │ 35309      │
+│ progress               │ 0.270270   │
+│ skyrl.ai/grad_norm     │ 0.341943   │
+│ skyrl.ai/learning_rate │ 0.000073   │
+│ time_total             │ 118.789298 │
+│ train_mean_nll         │ 2.354352   │
+└────────────────────────┴────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 21                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000072  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37229     │
+│ progress               │ 0.283784  │
+│ skyrl.ai/grad_norm     │ 0.345487  │
+│ skyrl.ai/learning_rate │ 0.000072  │
+│ time_total             │ 77.501071 │
+│ train_mean_nll         │ 2.356468  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 22                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000070  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 36087     │
+│ progress               │ 0.297297  │
+│ skyrl.ai/grad_norm     │ 0.434072  │
+│ skyrl.ai/learning_rate │ 0.000070  │
+│ time_total             │ 76.563729 │
+│ train_mean_nll         │ 2.349741  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 23                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000069  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33348     │
+│ progress               │ 0.310811  │
+│ skyrl.ai/grad_norm     │ 0.666360  │
+│ skyrl.ai/learning_rate │ 0.000069  │
+│ time_total             │ 39.467893 │
+│ train_mean_nll         │ 2.315340  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 24                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000068  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35210     │
+│ progress               │ 0.324324  │
+│ skyrl.ai/grad_norm     │ 0.362909  │
+│ skyrl.ai/learning_rate │ 0.000068  │
+│ time_total             │ 76.473245 │
+│ train_mean_nll         │ 2.381596  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 25                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000066  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34833     │
+│ progress               │ 0.337838  │
+│ skyrl.ai/grad_norm     │ 0.389343  │
+│ skyrl.ai/learning_rate │ 0.000066  │
+│ time_total             │ 77.588004 │
+│ train_mean_nll         │ 2.444775  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 26                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000065  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37256     │
+│ progress               │ 0.351351  │
+│ skyrl.ai/grad_norm     │ 0.379360  │
+│ skyrl.ai/learning_rate │ 0.000065  │
+│ time_total             │ 76.494458 │
+│ train_mean_nll         │ 2.408405  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 27                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000064  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33947     │
+│ progress               │ 0.364865  │
+│ skyrl.ai/grad_norm     │ 0.327722  │
+│ skyrl.ai/learning_rate │ 0.000064  │
+│ time_total             │ 76.535758 │
+│ train_mean_nll         │ 2.349726  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 28                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000062  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 32694     │
+│ progress               │ 0.378378  │
+│ skyrl.ai/grad_norm     │ 0.331114  │
+│ skyrl.ai/learning_rate │ 0.000062  │
+│ time_total             │ 75.470683 │
+│ train_mean_nll         │ 2.433801  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 29                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000061  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37440     │
+│ progress               │ 0.391892  │
+│ skyrl.ai/grad_norm     │ 0.354970  │
+│ skyrl.ai/learning_rate │ 0.000061  │
+│ time_total             │ 76.488313 │
+│ train_mean_nll         │ 2.169731  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 30                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000059  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39453     │
+│ progress               │ 0.405405  │
+│ skyrl.ai/grad_norm     │ 0.323725  │
+│ skyrl.ai/learning_rate │ 0.000059  │
+│ time_total             │ 77.668248 │
+│ train_mean_nll         │ 2.282022  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 31                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000058  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 36143     │
+│ progress               │ 0.418919  │
+│ skyrl.ai/grad_norm     │ 0.326314  │
+│ skyrl.ai/learning_rate │ 0.000058  │
+│ time_total             │ 77.478847 │
+│ train_mean_nll         │ 2.280007  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 32                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000057  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34808     │
+│ progress               │ 0.432432  │
+│ skyrl.ai/grad_norm     │ 0.385605  │
+│ skyrl.ai/learning_rate │ 0.000057  │
+│ time_total             │ 76.566763 │
+│ train_mean_nll         │ 2.355149  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 33                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000055  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 40403     │
+│ progress               │ 0.445946  │
+│ skyrl.ai/grad_norm     │ 0.292981  │
+│ skyrl.ai/learning_rate │ 0.000055  │
+│ time_total             │ 77.524634 │
+│ train_mean_nll         │ 2.233092  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 34                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000054  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33581     │
+│ progress               │ 0.459459  │
+│ skyrl.ai/grad_norm     │ 0.330763  │
+│ skyrl.ai/learning_rate │ 0.000054  │
+│ time_total             │ 76.473376 │
+│ train_mean_nll         │ 2.357910  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 35                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000053  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35478     │
+│ progress               │ 0.472973  │
+│ skyrl.ai/grad_norm     │ 0.270800  │
+│ skyrl.ai/learning_rate │ 0.000053  │
+│ time_total             │ 75.483867 │
+│ train_mean_nll         │ 2.342329  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 36                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000051  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37613     │
+│ progress               │ 0.486486  │
+│ skyrl.ai/grad_norm     │ 0.442881  │
+│ skyrl.ai/learning_rate │ 0.000051  │
+│ time_total             │ 77.512969 │
+│ train_mean_nll         │ 2.252668  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 37                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000050  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37243     │
+│ progress               │ 0.500000  │
+│ skyrl.ai/grad_norm     │ 0.338969  │
+│ skyrl.ai/learning_rate │ 0.000050  │
+│ time_total             │ 77.658975 │
+│ train_mean_nll         │ 2.455121  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 38                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000049  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34551     │
+│ progress               │ 0.513514  │
+│ skyrl.ai/grad_norm     │ 0.351700  │
+│ skyrl.ai/learning_rate │ 0.000049  │
+│ time_total             │ 76.483724 │
+│ train_mean_nll         │ 2.280617  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 39                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000047  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34720     │
+│ progress               │ 0.527027  │
+│ skyrl.ai/grad_norm     │ 0.324459  │
+│ skyrl.ai/learning_rate │ 0.000047  │
+│ time_total             │ 76.485910 │
+│ train_mean_nll         │ 2.223614  │
+└────────────────────────┴───────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_a6e27bdf/weights/000040'}
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 40                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value      ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
+│ learning_rate          │ 0.000046   │
+│ num_sequences          │ 128        │
+│ num_tokens             │ 37150      │
+│ progress               │ 0.540541   │
+│ skyrl.ai/grad_norm     │ 0.291667   │
+│ skyrl.ai/learning_rate │ 0.000046   │
+│ time_total             │ 116.702335 │
+│ train_mean_nll         │ 2.320501   │
+└────────────────────────┴────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 41                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000045  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39832     │
+│ progress               │ 0.554054  │
+│ skyrl.ai/grad_norm     │ 0.330698  │
+│ skyrl.ai/learning_rate │ 0.000045  │
+│ time_total             │ 76.535000 │
+│ train_mean_nll         │ 2.251225  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 42                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000043  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 41346     │
+│ progress               │ 0.567568  │
+│ skyrl.ai/grad_norm     │ 0.344484  │
+│ skyrl.ai/learning_rate │ 0.000043  │
+│ time_total             │ 76.549524 │
+│ train_mean_nll         │ 2.185086  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 43                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000042  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35728     │
+│ progress               │ 0.581081  │
+│ skyrl.ai/grad_norm     │ 0.307271  │
+│ skyrl.ai/learning_rate │ 0.000042  │
+│ time_total             │ 40.502563 │
+│ train_mean_nll         │ 2.388796  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 44                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000041  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37469     │
+│ progress               │ 0.594595  │
+│ skyrl.ai/grad_norm     │ 0.304323  │
+│ skyrl.ai/learning_rate │ 0.000041  │
+│ time_total             │ 75.640097 │
+│ train_mean_nll         │ 2.405290  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 45                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000039  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 36253     │
+│ progress               │ 0.608108  │
+│ skyrl.ai/grad_norm     │ 0.296620  │
+│ skyrl.ai/learning_rate │ 0.000039  │
+│ time_total             │ 76.666610 │
+│ train_mean_nll         │ 2.326724  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_a.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 46                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000038  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39163     │
+│ progress               │ 0.621622  │
+│ skyrl.ai/grad_norm     │ 0.335458  │
+│ skyrl.ai/learning_rate │ 0.000038  │
+│ time_total             │ 77.500199 │
+│ train_mean_nll         │ 2.402814  │
+└────────────────────────┴───────────┘
diff --git a/tests/tinker/smoke_logs/sl_loop_a.log/metrics.jsonl b/tests/tinker/smoke_logs/sl_loop_a.log/metrics.jsonl
new file mode 100644
index 0000000000..24345b00bb
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_a.log/metrics.jsonl
@@ -0,0 +1,47 @@
+{"step": 0, "skyrl.ai/grad_norm": 1.4195045582052144, "skyrl.ai/learning_rate": 0.0001, "num_sequences": 128, "num_tokens": 37834, "learning_rate": 0.0001, "train_mean_nll": 2.9373137950897217, "progress": 0.0, "time_total": 87.54796123504639}
+{"step": 1, "skyrl.ai/grad_norm": 1.7579454159817616, "skyrl.ai/learning_rate": 9.864864864864865e-05, "num_sequences": 128, "num_tokens": 35654, "learning_rate": 9.864864864864865e-05, "train_mean_nll": 3.1417813301086426, "progress": 0.013513513513513514, "time_total": 40.48841595649719}
+{"step": 2, "skyrl.ai/grad_norm": 1.3759201612209633, "skyrl.ai/learning_rate": 9.729729729729731e-05, "num_sequences": 128, "num_tokens": 38474, "learning_rate": 9.729729729729731e-05, "train_mean_nll": 2.7579760551452637, "progress": 0.02702702702702703, "time_total": 40.69628691673279}
+{"step": 3, "skyrl.ai/grad_norm": 0.9926018037640125, "skyrl.ai/learning_rate": 9.594594594594595e-05, "num_sequences": 128, "num_tokens": 39781, "learning_rate": 9.594594594594595e-05, "train_mean_nll": 2.6649603843688965, "progress": 0.04054054054054054, "time_total": 40.512256145477295}
+{"step": 4, "skyrl.ai/grad_norm": 1.0000684237912945, "skyrl.ai/learning_rate": 9.45945945945946e-05, "num_sequences": 128, "num_tokens": 37346, "learning_rate": 9.45945945945946e-05, "train_mean_nll": 2.6431944370269775, "progress": 0.05405405405405406, "time_total": 76.53921723365784}
+{"step": 5, "skyrl.ai/grad_norm": 0.8648871075824316, "skyrl.ai/learning_rate": 9.324324324324324e-05, "num_sequences": 128, "num_tokens": 38710, "learning_rate": 9.324324324324324e-05, "train_mean_nll": 2.497553586959839, "progress": 0.06756756756756757, "time_total": 77.63196992874146}
+{"step": 6, "skyrl.ai/grad_norm": 0.6221997711718907, "skyrl.ai/learning_rate": 9.189189189189189e-05, "num_sequences": 128, "num_tokens": 33356, "learning_rate": 9.189189189189189e-05, "train_mean_nll": 2.6639108657836914, "progress": 0.08108108108108109, "time_total": 76.48116183280945}
+{"step": 7, "skyrl.ai/grad_norm": 0.546004583585804, "skyrl.ai/learning_rate": 9.054054054054055e-05, "num_sequences": 128, "num_tokens": 39115, "learning_rate": 9.054054054054055e-05, "train_mean_nll": 2.4794111251831055, "progress": 0.0945945945945946, "time_total": 76.49954676628113}
+{"step": 8, "skyrl.ai/grad_norm": 0.5897014016684463, "skyrl.ai/learning_rate": 8.918918918918919e-05, "num_sequences": 128, "num_tokens": 38640, "learning_rate": 8.918918918918919e-05, "train_mean_nll": 2.446202278137207, "progress": 0.10810810810810811, "time_total": 76.49341583251953}
+{"step": 9, "skyrl.ai/grad_norm": 0.4815193103926899, "skyrl.ai/learning_rate": 8.783783783783784e-05, "num_sequences": 128, "num_tokens": 38863, "learning_rate": 8.783783783783784e-05, "train_mean_nll": 2.3817343711853027, "progress": 0.12162162162162163, "time_total": 76.62121725082397}
+{"step": 10, "skyrl.ai/grad_norm": 0.5768029587667796, "skyrl.ai/learning_rate": 8.64864864864865e-05, "num_sequences": 128, "num_tokens": 33155, "learning_rate": 8.64864864864865e-05, "train_mean_nll": 2.5206613540649414, "progress": 0.13513513513513514, "time_total": 76.48070955276489}
+{"step": 11, "skyrl.ai/grad_norm": 0.6411696654000828, "skyrl.ai/learning_rate": 8.513513513513514e-05, "num_sequences": 128, "num_tokens": 38406, "learning_rate": 8.513513513513514e-05, "train_mean_nll": 2.459423303604126, "progress": 0.14864864864864866, "time_total": 76.49941420555115}
+{"step": 12, "skyrl.ai/grad_norm": 0.5676292107649995, "skyrl.ai/learning_rate": 8.378378378378379e-05, "num_sequences": 128, "num_tokens": 38843, "learning_rate": 8.378378378378379e-05, "train_mean_nll": 2.3314881324768066, "progress": 0.16216216216216217, "time_total": 76.51964163780212}
+{"step": 13, "skyrl.ai/grad_norm": 0.5330967779616385, "skyrl.ai/learning_rate": 8.243243243243243e-05, "num_sequences": 128, "num_tokens": 35583, "learning_rate": 8.243243243243243e-05, "train_mean_nll": 2.370727300643921, "progress": 0.17567567567567569, "time_total": 76.48596978187561}
+{"step": 14, "skyrl.ai/grad_norm": 0.44244537584796306, "skyrl.ai/learning_rate": 8.108108108108108e-05, "num_sequences": 128, "num_tokens": 36717, "learning_rate": 8.108108108108108e-05, "train_mean_nll": 2.2848665714263916, "progress": 0.1891891891891892, "time_total": 76.49979639053345}
+{"step": 15, "skyrl.ai/grad_norm": 0.45961910755591384, "skyrl.ai/learning_rate": 7.972972972972974e-05, "num_sequences": 128, "num_tokens": 34002, "learning_rate": 7.972972972972974e-05, "train_mean_nll": 2.4119741916656494, "progress": 0.20270270270270271, "time_total": 76.71580243110657}
+{"step": 16, "skyrl.ai/grad_norm": 0.3437916340323792, "skyrl.ai/learning_rate": 7.837837837837838e-05, "num_sequences": 128, "num_tokens": 38620, "learning_rate": 7.837837837837838e-05, "train_mean_nll": 2.385261297225952, "progress": 0.21621621621621623, "time_total": 76.49681758880615}
+{"step": 17, "skyrl.ai/grad_norm": 0.32883653241641353, "skyrl.ai/learning_rate": 7.702702702702703e-05, "num_sequences": 128, "num_tokens": 37639, "learning_rate": 7.702702702702703e-05, "train_mean_nll": 2.256648540496826, "progress": 0.22972972972972974, "time_total": 75.49857974052429}
+{"step": 18, "skyrl.ai/grad_norm": 0.3842549687115808, "skyrl.ai/learning_rate": 7.567567567567568e-05, "num_sequences": 128, "num_tokens": 34118, "learning_rate": 7.567567567567568e-05, "train_mean_nll": 2.4064152240753174, "progress": 0.24324324324324326, "time_total": 77.49464082717896}
+{"step": 19, "skyrl.ai/grad_norm": 0.36468983411654887, "skyrl.ai/learning_rate": 7.432432432432433e-05, "num_sequences": 128, "num_tokens": 37178, "learning_rate": 7.432432432432433e-05, "train_mean_nll": 2.389909029006958, "progress": 0.25675675675675674, "time_total": 75.48325562477112}
+{"step": 20, "skyrl.ai/grad_norm": 0.3419431427929476, "skyrl.ai/learning_rate": 7.297297297297297e-05, "num_sequences": 128, "num_tokens": 35309, "learning_rate": 7.297297297297297e-05, "train_mean_nll": 2.354351758956909, "progress": 0.2702702702702703, "time_total": 118.78929805755615}
+{"step": 21, "skyrl.ai/grad_norm": 0.345487268590772, "skyrl.ai/learning_rate": 7.162162162162162e-05, "num_sequences": 128, "num_tokens": 37229, "learning_rate": 7.162162162162162e-05, "train_mean_nll": 2.3564682006835938, "progress": 0.28378378378378377, "time_total": 77.50107073783875}
+{"step": 22, "skyrl.ai/grad_norm": 0.4340724021435782, "skyrl.ai/learning_rate": 7.027027027027026e-05, "num_sequences": 128, "num_tokens": 36087, "learning_rate": 7.027027027027026e-05, "train_mean_nll": 2.349740743637085, "progress": 0.2972972972972973, "time_total": 76.56372880935669}
+{"step": 23, "skyrl.ai/grad_norm": 0.6663603128429356, "skyrl.ai/learning_rate": 6.891891891891892e-05, "num_sequences": 128, "num_tokens": 33348, "learning_rate": 6.891891891891892e-05, "train_mean_nll": 2.315340280532837, "progress": 0.3108108108108108, "time_total": 39.46789288520813}
+{"step": 24, "skyrl.ai/grad_norm": 0.3629089827149277, "skyrl.ai/learning_rate": 6.756756756756757e-05, "num_sequences": 128, "num_tokens": 35210, "learning_rate": 6.756756756756757e-05, "train_mean_nll": 2.381596088409424, "progress": 0.32432432432432434, "time_total": 76.47324466705322}
+{"step": 25, "skyrl.ai/grad_norm": 0.3893427785274499, "skyrl.ai/learning_rate": 6.621621621621621e-05, "num_sequences": 128, "num_tokens": 34833, "learning_rate": 6.621621621621621e-05, "train_mean_nll": 2.444774866104126, "progress": 0.33783783783783783, "time_total": 77.58800411224365}
+{"step": 26, "skyrl.ai/grad_norm": 0.3793604576491017, "skyrl.ai/learning_rate": 6.486486486486487e-05, "num_sequences": 128, "num_tokens": 37256, "learning_rate": 6.486486486486487e-05, "train_mean_nll": 2.40840482711792, "progress": 0.35135135135135137, "time_total": 76.49445819854736}
+{"step": 27, "skyrl.ai/grad_norm": 0.32772194994537934, "skyrl.ai/learning_rate": 6.351351351351352e-05, "num_sequences": 128, "num_tokens": 33947, "learning_rate": 6.351351351351352e-05, "train_mean_nll": 2.349726438522339, "progress": 0.36486486486486486, "time_total": 76.53575825691223}
+{"step": 28, "skyrl.ai/grad_norm": 0.3311138494060635, "skyrl.ai/learning_rate": 6.216216216216216e-05, "num_sequences": 128, "num_tokens": 32694, "learning_rate": 6.216216216216216e-05, "train_mean_nll": 2.4338009357452393, "progress": 0.3783783783783784, "time_total": 75.47068309783936}
+{"step": 29, "skyrl.ai/grad_norm": 0.3549697433593456, "skyrl.ai/learning_rate": 6.0810810810810814e-05, "num_sequences": 128, "num_tokens": 37440, "learning_rate": 6.0810810810810814e-05, "train_mean_nll": 2.1697309017181396, "progress": 0.3918918918918919, "time_total": 76.4883131980896}
+{"step": 30, "skyrl.ai/grad_norm": 0.3237251062653877, "skyrl.ai/learning_rate": 5.945945945945945e-05, "num_sequences": 128, "num_tokens": 39453, "learning_rate": 5.945945945945945e-05, "train_mean_nll": 2.282021999359131, "progress": 0.40540540540540543, "time_total": 77.6682481765747}
+{"step": 31, "skyrl.ai/grad_norm": 0.32631397007477897, "skyrl.ai/learning_rate": 5.810810810810812e-05, "num_sequences": 128, "num_tokens": 36143, "learning_rate": 5.810810810810812e-05, "train_mean_nll": 2.2800068855285645, "progress": 0.4189189189189189, "time_total": 77.47884702682495}
+{"step": 32, "skyrl.ai/grad_norm": 0.3856050658058465, "skyrl.ai/learning_rate": 5.6756756756756757e-05, "num_sequences": 128, "num_tokens": 34808, "learning_rate": 5.6756756756756757e-05, "train_mean_nll": 2.3551485538482666, "progress": 0.43243243243243246, "time_total": 76.5667634010315}
+{"step": 33, "skyrl.ai/grad_norm": 0.292981300085196, "skyrl.ai/learning_rate": 5.540540540540541e-05, "num_sequences": 128, "num_tokens": 40403, "learning_rate": 5.540540540540541e-05, "train_mean_nll": 2.2330923080444336, "progress": 0.44594594594594594, "time_total": 77.52463436126709}
+{"step": 34, "skyrl.ai/grad_norm": 0.3307628301574217, "skyrl.ai/learning_rate": 5.405405405405406e-05, "num_sequences": 128, "num_tokens": 33581, "learning_rate": 5.405405405405406e-05, "train_mean_nll": 2.357910394668579, "progress": 0.4594594594594595, "time_total": 76.47337627410889}
+{"step": 35, "skyrl.ai/grad_norm": 0.270800408171525, "skyrl.ai/learning_rate": 5.27027027027027e-05, "num_sequences": 128, "num_tokens": 35478, "learning_rate": 5.27027027027027e-05, "train_mean_nll": 2.3423290252685547, "progress": 0.47297297297297297, "time_total": 75.48386669158936}
+{"step": 36, "skyrl.ai/grad_norm": 0.44288095213023204, "skyrl.ai/learning_rate": 5.135135135135135e-05, "num_sequences": 128, "num_tokens": 37613, "learning_rate": 5.135135135135135e-05, "train_mean_nll": 2.2526679039001465, "progress": 0.4864864864864865, "time_total": 77.51296854019165}
+{"step": 37, "skyrl.ai/grad_norm": 0.3389692565050453, "skyrl.ai/learning_rate": 5e-05, "num_sequences": 128, "num_tokens": 37243, "learning_rate": 5e-05, "train_mean_nll": 2.455120801925659, "progress": 0.5, "time_total": 77.65897512435913}
+{"step": 38, "skyrl.ai/grad_norm": 0.3517002471639325, "skyrl.ai/learning_rate": 4.8648648648648654e-05, "num_sequences": 128, "num_tokens": 34551, "learning_rate": 4.8648648648648654e-05, "train_mean_nll": 2.2806174755096436, "progress": 0.5135135135135135, "time_total": 76.48372411727905}
+{"step": 39, "skyrl.ai/grad_norm": 0.32445883766323885, "skyrl.ai/learning_rate": 4.7297297297297306e-05, "num_sequences": 128, "num_tokens": 34720, "learning_rate": 4.7297297297297306e-05, "train_mean_nll": 2.223614454269409, "progress": 0.527027027027027, "time_total": 76.48590993881226}
+{"step": 40, "skyrl.ai/grad_norm": 0.29166733650857213, "skyrl.ai/learning_rate": 4.5945945945945944e-05, "num_sequences": 128, "num_tokens": 37150, "learning_rate": 4.5945945945945944e-05, "train_mean_nll": 2.3205010890960693, "progress": 0.5405405405405406, "time_total": 116.70233464241028}
+{"step": 41, "skyrl.ai/grad_norm": 0.33069804062642083, "skyrl.ai/learning_rate": 4.4594594594594596e-05, "num_sequences": 128, "num_tokens": 39832, "learning_rate": 4.4594594594594596e-05, "train_mean_nll": 2.2512245178222656, "progress": 0.5540540540540541, "time_total": 76.53500008583069}
+{"step": 42, "skyrl.ai/grad_norm": 0.3444839142267573, "skyrl.ai/learning_rate": 4.324324324324325e-05, "num_sequences": 128, "num_tokens": 41346, "learning_rate": 4.324324324324325e-05, "train_mean_nll": 2.1850860118865967, "progress": 0.5675675675675675, "time_total": 76.54952430725098}
+{"step": 43, "skyrl.ai/grad_norm": 0.30727129669347636, "skyrl.ai/learning_rate": 4.18918918918919e-05, "num_sequences": 128, "num_tokens": 35728, "learning_rate": 4.18918918918919e-05, "train_mean_nll": 2.3887956142425537, "progress": 0.581081081081081, "time_total": 40.50256276130676}
+{"step": 44, "skyrl.ai/grad_norm": 0.3043229538756859, "skyrl.ai/learning_rate": 4.054054054054054e-05, "num_sequences": 128, "num_tokens": 37469, "learning_rate": 4.054054054054054e-05, "train_mean_nll": 2.405290126800537, "progress": 0.5945945945945946, "time_total": 75.64009666442871}
+{"step": 45, "skyrl.ai/grad_norm": 0.29661968194443833, "skyrl.ai/learning_rate": 3.918918918918919e-05, "num_sequences": 128, "num_tokens": 36253, "learning_rate": 3.918918918918919e-05, "train_mean_nll": 2.3267242908477783, "progress": 0.6081081081081081, "time_total": 76.66661047935486}
+{"step": 46, "skyrl.ai/grad_norm": 0.33545840277169836, "skyrl.ai/learning_rate": 3.783783783783784e-05, "num_sequences": 128, "num_tokens": 39163, "learning_rate": 3.783783783783784e-05, "train_mean_nll": 2.4028143882751465, "progress": 0.6216216216216216, "time_total": 77.50019884109497}
diff --git a/tests/tinker/smoke_logs/sl_loop_b.log/checkpoints.jsonl b/tests/tinker/smoke_logs/sl_loop_b.log/checkpoints.jsonl
new file mode 100644
index 0000000000..28c58a12d3
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_b.log/checkpoints.jsonl
@@ -0,0 +1,2 @@
+{"name": "000020", "batch": 20, "state_path": "tinker://model_973e0c8b/weights/000020"}
+{"name": "000040", "batch": 40, "state_path": "tinker://model_973e0c8b/weights/000040"}
diff --git a/tests/tinker/smoke_logs/sl_loop_b.log/code.diff b/tests/tinker/smoke_logs/sl_loop_b.log/code.diff
new file mode 100644
index 0000000000..d8e559755f
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_b.log/code.diff
@@ -0,0 +1,3 @@
+### repo: /home/etang/tinker-cookbook @ dacb835b31fb3f7a012851a1bc64c950d0495de3
+modules: tinker_cookbook
+(no local changes)
diff --git a/tests/tinker/smoke_logs/sl_loop_b.log/config.json b/tests/tinker/smoke_logs/sl_loop_b.log/config.json
new file mode 100644
index 0000000000..b400cc9b82
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_b.log/config.json
@@ -0,0 +1,12 @@
+{
+  "base_url": "http://localhost:8000",
+  "log_path": "/tmp/sl_loop_b.log",
+  "model_name": "Qwen/Qwen3-0.6B",
+  "batch_size": 128,
+  "learning_rate": 0.0001,
+  "max_length": 32768,
+  "train_on_what": "last_assistant_message",
+  "lora_rank": 32,
+  "save_every": 20,
+  "ttl_seconds": 604800
+}
\ No newline at end of file
diff --git a/tests/tinker/smoke_logs/sl_loop_b.log/logs.log b/tests/tinker/smoke_logs/sl_loop_b.log/logs.log
new file mode 100644
index 0000000000..5a1ed0335d
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_b.log/logs.log
@@ -0,0 +1,674 @@
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/sl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B train_on_what=LAST_ASSISTANT_MESSAGE lora_rank=32 log_path=/tmp/sl_loop_b.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/sl_loop_b.log
+__main__:51 [INFO] Using renderer: qwen3
+__main__:54 [INFO] Loading dataset...
+__main__:64 [INFO] Dropping last 28 examples to keep batch size uniform at 128
+__main__:67 [INFO] Train batches: 74
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_e9ebd7a9
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/sl_loop_b.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/sl_loop_b.log
+tinker.lib.public_interfaces.service_client:159 [INFO] TrainingClient initialized for model model_973e0c8b
+__main__:87 [INFO] Training for 74 steps
+tinker_cookbook.supervised.common:188 [INFO] Weight reduction: 'mean' (token-mean loss)
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 0                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000100  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37834     │
+│ progress               │ 0.000000  │
+│ skyrl.ai/grad_norm     │ 1.419539  │
+│ skyrl.ai/learning_rate │ 0.000100  │
+│ time_total             │ 77.518479 │
+│ train_mean_nll         │ 2.937314  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 1                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000099  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35654     │
+│ progress               │ 0.013514  │
+│ skyrl.ai/grad_norm     │ 1.787320  │
+│ skyrl.ai/learning_rate │ 0.000099  │
+│ time_total             │ 77.489699 │
+│ train_mean_nll         │ 3.171597  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 2                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000097  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38474     │
+│ progress               │ 0.027027  │
+│ skyrl.ai/grad_norm     │ 1.387239  │
+│ skyrl.ai/learning_rate │ 0.000097  │
+│ time_total             │ 77.495156 │
+│ train_mean_nll         │ 2.826397  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 3                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000096  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39781     │
+│ progress               │ 0.040541  │
+│ skyrl.ai/grad_norm     │ 1.347174  │
+│ skyrl.ai/learning_rate │ 0.000096  │
+│ time_total             │ 76.504979 │
+│ train_mean_nll         │ 2.734261  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 4                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000095  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37346     │
+│ progress               │ 0.054054  │
+│ skyrl.ai/grad_norm     │ 1.056902  │
+│ skyrl.ai/learning_rate │ 0.000095  │
+│ time_total             │ 76.500813 │
+│ train_mean_nll         │ 2.713544  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 5                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000093  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38710     │
+│ progress               │ 0.067568  │
+│ skyrl.ai/grad_norm     │ 1.037518  │
+│ skyrl.ai/learning_rate │ 0.000093  │
+│ time_total             │ 76.510272 │
+│ train_mean_nll         │ 2.544317  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 6                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000092  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33356     │
+│ progress               │ 0.081081  │
+│ skyrl.ai/grad_norm     │ 0.760348  │
+│ skyrl.ai/learning_rate │ 0.000092  │
+│ time_total             │ 76.468495 │
+│ train_mean_nll         │ 2.698561  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 7                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000091  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39115     │
+│ progress               │ 0.094595  │
+│ skyrl.ai/grad_norm     │ 0.621467  │
+│ skyrl.ai/learning_rate │ 0.000091  │
+│ time_total             │ 76.525711 │
+│ train_mean_nll         │ 2.502498  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 8                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000089  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38640     │
+│ progress               │ 0.108108  │
+│ skyrl.ai/grad_norm     │ 0.597963  │
+│ skyrl.ai/learning_rate │ 0.000089  │
+│ time_total             │ 76.506974 │
+│ train_mean_nll         │ 2.464289  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 9                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000088  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38863     │
+│ progress               │ 0.121622  │
+│ skyrl.ai/grad_norm     │ 0.515755  │
+│ skyrl.ai/learning_rate │ 0.000088  │
+│ time_total             │ 76.510010 │
+│ train_mean_nll         │ 2.393878  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 10                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000086  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33155     │
+│ progress               │ 0.135135  │
+│ skyrl.ai/grad_norm     │ 0.678250  │
+│ skyrl.ai/learning_rate │ 0.000086  │
+│ time_total             │ 76.477246 │
+│ train_mean_nll         │ 2.540268  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 11                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000085  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38406     │
+│ progress               │ 0.148649  │
+│ skyrl.ai/grad_norm     │ 0.630785  │
+│ skyrl.ai/learning_rate │ 0.000085  │
+│ time_total             │ 76.500259 │
+│ train_mean_nll         │ 2.475491  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 12                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000084  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38843     │
+│ progress               │ 0.162162  │
+│ skyrl.ai/grad_norm     │ 0.590551  │
+│ skyrl.ai/learning_rate │ 0.000084  │
+│ time_total             │ 76.506439 │
+│ train_mean_nll         │ 2.344424  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 13                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000082  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35583     │
+│ progress               │ 0.175676  │
+│ skyrl.ai/grad_norm     │ 0.546812  │
+│ skyrl.ai/learning_rate │ 0.000082  │
+│ time_total             │ 76.591751 │
+│ train_mean_nll         │ 2.381466  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 14                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000081  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 36717     │
+│ progress               │ 0.189189  │
+│ skyrl.ai/grad_norm     │ 0.475584  │
+│ skyrl.ai/learning_rate │ 0.000081  │
+│ time_total             │ 76.374654 │
+│ train_mean_nll         │ 2.290986  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 15                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000080  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34002     │
+│ progress               │ 0.202703  │
+│ skyrl.ai/grad_norm     │ 0.517601  │
+│ skyrl.ai/learning_rate │ 0.000080  │
+│ time_total             │ 76.461699 │
+│ train_mean_nll         │ 2.419916  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 16                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000078  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38620     │
+│ progress               │ 0.216216  │
+│ skyrl.ai/grad_norm     │ 0.359259  │
+│ skyrl.ai/learning_rate │ 0.000078  │
+│ time_total             │ 75.604731 │
+│ train_mean_nll         │ 2.389113  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 17                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000077  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37639     │
+│ progress               │ 0.229730  │
+│ skyrl.ai/grad_norm     │ 0.338165  │
+│ skyrl.ai/learning_rate │ 0.000077  │
+│ time_total             │ 42.506935 │
+│ train_mean_nll         │ 2.259868  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 18                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000076  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34118     │
+│ progress               │ 0.243243  │
+│ skyrl.ai/grad_norm     │ 0.416603  │
+│ skyrl.ai/learning_rate │ 0.000076  │
+│ time_total             │ 76.720900 │
+│ train_mean_nll         │ 2.409774  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 19                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000074  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37178     │
+│ progress               │ 0.256757  │
+│ skyrl.ai/grad_norm     │ 0.370585  │
+│ skyrl.ai/learning_rate │ 0.000074  │
+│ time_total             │ 76.489900 │
+│ train_mean_nll         │ 2.392253  │
+└────────────────────────┴───────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_973e0c8b/weights/000020'}
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 20                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value      ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
+│ learning_rate          │ 0.000073   │
+│ num_sequences          │ 128        │
+│ num_tokens             │ 35309      │
+│ progress               │ 0.270270   │
+│ skyrl.ai/grad_norm     │ 0.348055   │
+│ skyrl.ai/learning_rate │ 0.000073   │
+│ time_total             │ 116.662111 │
+│ train_mean_nll         │ 2.357068   │
+└────────────────────────┴────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 21                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000072  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37229     │
+│ progress               │ 0.283784  │
+│ skyrl.ai/grad_norm     │ 0.349533  │
+│ skyrl.ai/learning_rate │ 0.000072  │
+│ time_total             │ 76.505576 │
+│ train_mean_nll         │ 2.358591  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 22                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000070  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 36087     │
+│ progress               │ 0.297297  │
+│ skyrl.ai/grad_norm     │ 0.437823  │
+│ skyrl.ai/learning_rate │ 0.000070  │
+│ time_total             │ 77.491332 │
+│ train_mean_nll         │ 2.352475  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 23                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000069  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33348     │
+│ progress               │ 0.310811  │
+│ skyrl.ai/grad_norm     │ 0.675538  │
+│ skyrl.ai/learning_rate │ 0.000069  │
+│ time_total             │ 76.467921 │
+│ train_mean_nll         │ 2.315181  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 24                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000068  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35210     │
+│ progress               │ 0.324324  │
+│ skyrl.ai/grad_norm     │ 0.366941  │
+│ skyrl.ai/learning_rate │ 0.000068  │
+│ time_total             │ 75.492159 │
+│ train_mean_nll         │ 2.383331  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 25                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000066  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34833     │
+│ progress               │ 0.337838  │
+│ skyrl.ai/grad_norm     │ 0.393730  │
+│ skyrl.ai/learning_rate │ 0.000066  │
+│ time_total             │ 76.488087 │
+│ train_mean_nll         │ 2.446757  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 26                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000065  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37256     │
+│ progress               │ 0.351351  │
+│ skyrl.ai/grad_norm     │ 0.384646  │
+│ skyrl.ai/learning_rate │ 0.000065  │
+│ time_total             │ 76.638779 │
+│ train_mean_nll         │ 2.410129  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 27                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000064  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33947     │
+│ progress               │ 0.364865  │
+│ skyrl.ai/grad_norm     │ 0.334691  │
+│ skyrl.ai/learning_rate │ 0.000064  │
+│ time_total             │ 77.462294 │
+│ train_mean_nll         │ 2.350893  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 28                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000062  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 32694     │
+│ progress               │ 0.378378  │
+│ skyrl.ai/grad_norm     │ 0.336949  │
+│ skyrl.ai/learning_rate │ 0.000062  │
+│ time_total             │ 77.451585 │
+│ train_mean_nll         │ 2.434802  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 29                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000061  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37440     │
+│ progress               │ 0.391892  │
+│ skyrl.ai/grad_norm     │ 0.361362  │
+│ skyrl.ai/learning_rate │ 0.000061  │
+│ time_total             │ 77.477361 │
+│ train_mean_nll         │ 2.170828  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 30                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000059  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39453     │
+│ progress               │ 0.405405  │
+│ skyrl.ai/grad_norm     │ 0.326734  │
+│ skyrl.ai/learning_rate │ 0.000059  │
+│ time_total             │ 76.520573 │
+│ train_mean_nll         │ 2.284935  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 31                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000058  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 36143     │
+│ progress               │ 0.418919  │
+│ skyrl.ai/grad_norm     │ 0.329552  │
+│ skyrl.ai/learning_rate │ 0.000058  │
+│ time_total             │ 76.565412 │
+│ train_mean_nll         │ 2.280433  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 32                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000057  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34808     │
+│ progress               │ 0.432432  │
+│ skyrl.ai/grad_norm     │ 0.382980  │
+│ skyrl.ai/learning_rate │ 0.000057  │
+│ time_total             │ 76.469333 │
+│ train_mean_nll         │ 2.354665  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 33                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000055  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 40403     │
+│ progress               │ 0.445946  │
+│ skyrl.ai/grad_norm     │ 0.291227  │
+│ skyrl.ai/learning_rate │ 0.000055  │
+│ time_total             │ 77.651549 │
+│ train_mean_nll         │ 2.232926  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 34                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000054  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33581     │
+│ progress               │ 0.459459  │
+│ skyrl.ai/grad_norm     │ 0.329446  │
+│ skyrl.ai/learning_rate │ 0.000054  │
+│ time_total             │ 76.468443 │
+│ train_mean_nll         │ 2.358539  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 35                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000053  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35478     │
+│ progress               │ 0.472973  │
+│ skyrl.ai/grad_norm     │ 0.268784  │
+│ skyrl.ai/learning_rate │ 0.000053  │
+│ time_total             │ 76.480491 │
+│ train_mean_nll         │ 2.342692  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 36                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000051  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37613     │
+│ progress               │ 0.486486  │
+│ skyrl.ai/grad_norm     │ 0.436201  │
+│ skyrl.ai/learning_rate │ 0.000051  │
+│ time_total             │ 76.597426 │
+│ train_mean_nll         │ 2.253095  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 37                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000050  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37243     │
+│ progress               │ 0.500000  │
+│ skyrl.ai/grad_norm     │ 0.328390  │
+│ skyrl.ai/learning_rate │ 0.000050  │
+│ time_total             │ 40.495865 │
+│ train_mean_nll         │ 2.455515  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 38                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000049  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34551     │
+│ progress               │ 0.513514  │
+│ skyrl.ai/grad_norm     │ 0.355869  │
+│ skyrl.ai/learning_rate │ 0.000049  │
+│ time_total             │ 76.466278 │
+│ train_mean_nll         │ 2.281497  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 39                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000047  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34720     │
+│ progress               │ 0.527027  │
+│ skyrl.ai/grad_norm     │ 0.324688  │
+│ skyrl.ai/learning_rate │ 0.000047  │
+│ time_total             │ 76.547050 │
+│ train_mean_nll         │ 2.224408  │
+└────────────────────────┴───────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_973e0c8b/weights/000040'}
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 40                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value      ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
+│ learning_rate          │ 0.000046   │
+│ num_sequences          │ 128        │
+│ num_tokens             │ 37150      │
+│ progress               │ 0.540541   │
+│ skyrl.ai/grad_norm     │ 0.290177   │
+│ skyrl.ai/learning_rate │ 0.000046   │
+│ time_total             │ 116.764540 │
+│ train_mean_nll         │ 2.320617   │
+└────────────────────────┴────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 41                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000045  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39832     │
+│ progress               │ 0.554054  │
+│ skyrl.ai/grad_norm     │ 0.329142  │
+│ skyrl.ai/learning_rate │ 0.000045  │
+│ time_total             │ 75.531128 │
+│ train_mean_nll         │ 2.251125  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 42                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000043  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 41346     │
+│ progress               │ 0.567568  │
+│ skyrl.ai/grad_norm     │ 0.342816  │
+│ skyrl.ai/learning_rate │ 0.000043  │
+│ time_total             │ 77.515983 │
+│ train_mean_nll         │ 2.186094  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_b.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 43                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000042  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35728     │
+│ progress               │ 0.581081  │
+│ skyrl.ai/grad_norm     │ 0.308010  │
+│ skyrl.ai/learning_rate │ 0.000042  │
+│ time_total             │ 76.476230 │
+│ train_mean_nll         │ 2.389423  │
+└────────────────────────┴───────────┘
diff --git a/tests/tinker/smoke_logs/sl_loop_b.log/metrics.jsonl b/tests/tinker/smoke_logs/sl_loop_b.log/metrics.jsonl
new file mode 100644
index 0000000000..8e0b9aa553
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_b.log/metrics.jsonl
@@ -0,0 +1,44 @@
+{"step": 0, "skyrl.ai/grad_norm": 1.4195386534746055, "skyrl.ai/learning_rate": 0.0001, "num_sequences": 128, "num_tokens": 37834, "learning_rate": 0.0001, "train_mean_nll": 2.9373137950897217, "progress": 0.0, "time_total": 77.51847863197327}
+{"step": 1, "skyrl.ai/grad_norm": 1.7873204841370076, "skyrl.ai/learning_rate": 9.864864864864865e-05, "num_sequences": 128, "num_tokens": 35654, "learning_rate": 9.864864864864865e-05, "train_mean_nll": 3.1715967655181885, "progress": 0.013513513513513514, "time_total": 77.48969888687134}
+{"step": 2, "skyrl.ai/grad_norm": 1.3872389986086984, "skyrl.ai/learning_rate": 9.729729729729731e-05, "num_sequences": 128, "num_tokens": 38474, "learning_rate": 9.729729729729731e-05, "train_mean_nll": 2.826397180557251, "progress": 0.02702702702702703, "time_total": 77.49515557289124}
+{"step": 3, "skyrl.ai/grad_norm": 1.347173808998758, "skyrl.ai/learning_rate": 9.594594594594595e-05, "num_sequences": 128, "num_tokens": 39781, "learning_rate": 9.594594594594595e-05, "train_mean_nll": 2.7342612743377686, "progress": 0.04054054054054054, "time_total": 76.50497913360596}
+{"step": 4, "skyrl.ai/grad_norm": 1.0569022766058442, "skyrl.ai/learning_rate": 9.45945945945946e-05, "num_sequences": 128, "num_tokens": 37346, "learning_rate": 9.45945945945946e-05, "train_mean_nll": 2.7135438919067383, "progress": 0.05405405405405406, "time_total": 76.50081324577332}
+{"step": 5, "skyrl.ai/grad_norm": 1.0375181495273973, "skyrl.ai/learning_rate": 9.324324324324324e-05, "num_sequences": 128, "num_tokens": 38710, "learning_rate": 9.324324324324324e-05, "train_mean_nll": 2.544316530227661, "progress": 0.06756756756756757, "time_total": 76.51027178764343}
+{"step": 6, "skyrl.ai/grad_norm": 0.7603481383776933, "skyrl.ai/learning_rate": 9.189189189189189e-05, "num_sequences": 128, "num_tokens": 33356, "learning_rate": 9.189189189189189e-05, "train_mean_nll": 2.698561191558838, "progress": 0.08108108108108109, "time_total": 76.46849489212036}
+{"step": 7, "skyrl.ai/grad_norm": 0.6214665187953508, "skyrl.ai/learning_rate": 9.054054054054055e-05, "num_sequences": 128, "num_tokens": 39115, "learning_rate": 9.054054054054055e-05, "train_mean_nll": 2.502498149871826, "progress": 0.0945945945945946, "time_total": 76.52571058273315}
+{"step": 8, "skyrl.ai/grad_norm": 0.5979632168990555, "skyrl.ai/learning_rate": 8.918918918918919e-05, "num_sequences": 128, "num_tokens": 38640, "learning_rate": 8.918918918918919e-05, "train_mean_nll": 2.4642891883850098, "progress": 0.10810810810810811, "time_total": 76.50697350502014}
+{"step": 9, "skyrl.ai/grad_norm": 0.5157547700747354, "skyrl.ai/learning_rate": 8.783783783783784e-05, "num_sequences": 128, "num_tokens": 38863, "learning_rate": 8.783783783783784e-05, "train_mean_nll": 2.393878221511841, "progress": 0.12162162162162163, "time_total": 76.510009765625}
+{"step": 10, "skyrl.ai/grad_norm": 0.6782504928072451, "skyrl.ai/learning_rate": 8.64864864864865e-05, "num_sequences": 128, "num_tokens": 33155, "learning_rate": 8.64864864864865e-05, "train_mean_nll": 2.5402684211730957, "progress": 0.13513513513513514, "time_total": 76.4772458076477}
+{"step": 11, "skyrl.ai/grad_norm": 0.630784871807005, "skyrl.ai/learning_rate": 8.513513513513514e-05, "num_sequences": 128, "num_tokens": 38406, "learning_rate": 8.513513513513514e-05, "train_mean_nll": 2.4754910469055176, "progress": 0.14864864864864866, "time_total": 76.50025868415833}
+{"step": 12, "skyrl.ai/grad_norm": 0.5905507883748928, "skyrl.ai/learning_rate": 8.378378378378379e-05, "num_sequences": 128, "num_tokens": 38843, "learning_rate": 8.378378378378379e-05, "train_mean_nll": 2.34442400932312, "progress": 0.16216216216216217, "time_total": 76.50643873214722}
+{"step": 13, "skyrl.ai/grad_norm": 0.546812244629307, "skyrl.ai/learning_rate": 8.243243243243243e-05, "num_sequences": 128, "num_tokens": 35583, "learning_rate": 8.243243243243243e-05, "train_mean_nll": 2.3814659118652344, "progress": 0.17567567567567569, "time_total": 76.59175086021423}
+{"step": 14, "skyrl.ai/grad_norm": 0.47558408889567894, "skyrl.ai/learning_rate": 8.108108108108108e-05, "num_sequences": 128, "num_tokens": 36717, "learning_rate": 8.108108108108108e-05, "train_mean_nll": 2.2909858226776123, "progress": 0.1891891891891892, "time_total": 76.3746542930603}
+{"step": 15, "skyrl.ai/grad_norm": 0.5176009550997458, "skyrl.ai/learning_rate": 7.972972972972974e-05, "num_sequences": 128, "num_tokens": 34002, "learning_rate": 7.972972972972974e-05, "train_mean_nll": 2.4199159145355225, "progress": 0.20270270270270271, "time_total": 76.46169948577881}
+{"step": 16, "skyrl.ai/grad_norm": 0.3592592550557549, "skyrl.ai/learning_rate": 7.837837837837838e-05, "num_sequences": 128, "num_tokens": 38620, "learning_rate": 7.837837837837838e-05, "train_mean_nll": 2.389112949371338, "progress": 0.21621621621621623, "time_total": 75.6047306060791}
+{"step": 17, "skyrl.ai/grad_norm": 0.33816451069248143, "skyrl.ai/learning_rate": 7.702702702702703e-05, "num_sequences": 128, "num_tokens": 37639, "learning_rate": 7.702702702702703e-05, "train_mean_nll": 2.2598676681518555, "progress": 0.22972972972972974, "time_total": 42.50693464279175}
+{"step": 18, "skyrl.ai/grad_norm": 0.41660331800817935, "skyrl.ai/learning_rate": 7.567567567567568e-05, "num_sequences": 128, "num_tokens": 34118, "learning_rate": 7.567567567567568e-05, "train_mean_nll": 2.409773588180542, "progress": 0.24324324324324326, "time_total": 76.72089982032776}
+{"step": 19, "skyrl.ai/grad_norm": 0.37058514324435654, "skyrl.ai/learning_rate": 7.432432432432433e-05, "num_sequences": 128, "num_tokens": 37178, "learning_rate": 7.432432432432433e-05, "train_mean_nll": 2.3922526836395264, "progress": 0.25675675675675674, "time_total": 76.48990035057068}
+{"step": 20, "skyrl.ai/grad_norm": 0.3480550321913909, "skyrl.ai/learning_rate": 7.297297297297297e-05, "num_sequences": 128, "num_tokens": 35309, "learning_rate": 7.297297297297297e-05, "train_mean_nll": 2.357067823410034, "progress": 0.2702702702702703, "time_total": 116.66211080551147}
+{"step": 21, "skyrl.ai/grad_norm": 0.34953308338610417, "skyrl.ai/learning_rate": 7.162162162162162e-05, "num_sequences": 128, "num_tokens": 37229, "learning_rate": 7.162162162162162e-05, "train_mean_nll": 2.358590841293335, "progress": 0.28378378378378377, "time_total": 76.50557613372803}
+{"step": 22, "skyrl.ai/grad_norm": 0.4378227848312563, "skyrl.ai/learning_rate": 7.027027027027026e-05, "num_sequences": 128, "num_tokens": 36087, "learning_rate": 7.027027027027026e-05, "train_mean_nll": 2.352475166320801, "progress": 0.2972972972972973, "time_total": 77.49133205413818}
+{"step": 23, "skyrl.ai/grad_norm": 0.6755377711219428, "skyrl.ai/learning_rate": 6.891891891891892e-05, "num_sequences": 128, "num_tokens": 33348, "learning_rate": 6.891891891891892e-05, "train_mean_nll": 2.3151814937591553, "progress": 0.3108108108108108, "time_total": 76.4679205417633}
+{"step": 24, "skyrl.ai/grad_norm": 0.36694108527031516, "skyrl.ai/learning_rate": 6.756756756756757e-05, "num_sequences": 128, "num_tokens": 35210, "learning_rate": 6.756756756756757e-05, "train_mean_nll": 2.383330821990967, "progress": 0.32432432432432434, "time_total": 75.49215888977051}
+{"step": 25, "skyrl.ai/grad_norm": 0.3937296892407543, "skyrl.ai/learning_rate": 6.621621621621621e-05, "num_sequences": 128, "num_tokens": 34833, "learning_rate": 6.621621621621621e-05, "train_mean_nll": 2.4467570781707764, "progress": 0.33783783783783783, "time_total": 76.48808670043945}
+{"step": 26, "skyrl.ai/grad_norm": 0.3846464408202079, "skyrl.ai/learning_rate": 6.486486486486487e-05, "num_sequences": 128, "num_tokens": 37256, "learning_rate": 6.486486486486487e-05, "train_mean_nll": 2.4101293087005615, "progress": 0.35135135135135137, "time_total": 76.63877940177917}
+{"step": 27, "skyrl.ai/grad_norm": 0.33469089711136596, "skyrl.ai/learning_rate": 6.351351351351352e-05, "num_sequences": 128, "num_tokens": 33947, "learning_rate": 6.351351351351352e-05, "train_mean_nll": 2.350893020629883, "progress": 0.36486486486486486, "time_total": 77.46229386329651}
+{"step": 28, "skyrl.ai/grad_norm": 0.336948956880145, "skyrl.ai/learning_rate": 6.216216216216216e-05, "num_sequences": 128, "num_tokens": 32694, "learning_rate": 6.216216216216216e-05, "train_mean_nll": 2.4348015785217285, "progress": 0.3783783783783784, "time_total": 77.451584815979}
+{"step": 29, "skyrl.ai/grad_norm": 0.36136163092248136, "skyrl.ai/learning_rate": 6.0810810810810814e-05, "num_sequences": 128, "num_tokens": 37440, "learning_rate": 6.0810810810810814e-05, "train_mean_nll": 2.170828342437744, "progress": 0.3918918918918919, "time_total": 77.47736096382141}
+{"step": 30, "skyrl.ai/grad_norm": 0.32673394434664593, "skyrl.ai/learning_rate": 5.945945945945945e-05, "num_sequences": 128, "num_tokens": 39453, "learning_rate": 5.945945945945945e-05, "train_mean_nll": 2.284935235977173, "progress": 0.40540540540540543, "time_total": 76.52057337760925}
+{"step": 31, "skyrl.ai/grad_norm": 0.32955249716070606, "skyrl.ai/learning_rate": 5.810810810810812e-05, "num_sequences": 128, "num_tokens": 36143, "learning_rate": 5.810810810810812e-05, "train_mean_nll": 2.280432939529419, "progress": 0.4189189189189189, "time_total": 76.56541204452515}
+{"step": 32, "skyrl.ai/grad_norm": 0.38298011535688614, "skyrl.ai/learning_rate": 5.6756756756756757e-05, "num_sequences": 128, "num_tokens": 34808, "learning_rate": 5.6756756756756757e-05, "train_mean_nll": 2.3546645641326904, "progress": 0.43243243243243246, "time_total": 76.4693329334259}
+{"step": 33, "skyrl.ai/grad_norm": 0.29122662415652056, "skyrl.ai/learning_rate": 5.540540540540541e-05, "num_sequences": 128, "num_tokens": 40403, "learning_rate": 5.540540540540541e-05, "train_mean_nll": 2.2329256534576416, "progress": 0.44594594594594594, "time_total": 77.65154910087585}
+{"step": 34, "skyrl.ai/grad_norm": 0.32944563359840584, "skyrl.ai/learning_rate": 5.405405405405406e-05, "num_sequences": 128, "num_tokens": 33581, "learning_rate": 5.405405405405406e-05, "train_mean_nll": 2.35853910446167, "progress": 0.4594594594594595, "time_total": 76.46844339370728}
+{"step": 35, "skyrl.ai/grad_norm": 0.26878365206169, "skyrl.ai/learning_rate": 5.27027027027027e-05, "num_sequences": 128, "num_tokens": 35478, "learning_rate": 5.27027027027027e-05, "train_mean_nll": 2.342691659927368, "progress": 0.47297297297297297, "time_total": 76.48049068450928}
+{"step": 36, "skyrl.ai/grad_norm": 0.4362009722094959, "skyrl.ai/learning_rate": 5.135135135135135e-05, "num_sequences": 128, "num_tokens": 37613, "learning_rate": 5.135135135135135e-05, "train_mean_nll": 2.2530946731567383, "progress": 0.4864864864864865, "time_total": 76.59742593765259}
+{"step": 37, "skyrl.ai/grad_norm": 0.3283900938474904, "skyrl.ai/learning_rate": 5e-05, "num_sequences": 128, "num_tokens": 37243, "learning_rate": 5e-05, "train_mean_nll": 2.455514669418335, "progress": 0.5, "time_total": 40.49586510658264}
+{"step": 38, "skyrl.ai/grad_norm": 0.3558688165645349, "skyrl.ai/learning_rate": 4.8648648648648654e-05, "num_sequences": 128, "num_tokens": 34551, "learning_rate": 4.8648648648648654e-05, "train_mean_nll": 2.2814974784851074, "progress": 0.5135135135135135, "time_total": 76.46627807617188}
+{"step": 39, "skyrl.ai/grad_norm": 0.32468846777146654, "skyrl.ai/learning_rate": 4.7297297297297306e-05, "num_sequences": 128, "num_tokens": 34720, "learning_rate": 4.7297297297297306e-05, "train_mean_nll": 2.224407911300659, "progress": 0.527027027027027, "time_total": 76.54704976081848}
+{"step": 40, "skyrl.ai/grad_norm": 0.2901767827597186, "skyrl.ai/learning_rate": 4.5945945945945944e-05, "num_sequences": 128, "num_tokens": 37150, "learning_rate": 4.5945945945945944e-05, "train_mean_nll": 2.3206169605255127, "progress": 0.5405405405405406, "time_total": 116.76453971862793}
+{"step": 41, "skyrl.ai/grad_norm": 0.3291420954072253, "skyrl.ai/learning_rate": 4.4594594594594596e-05, "num_sequences": 128, "num_tokens": 39832, "learning_rate": 4.4594594594594596e-05, "train_mean_nll": 2.251124858856201, "progress": 0.5540540540540541, "time_total": 75.5311279296875}
+{"step": 42, "skyrl.ai/grad_norm": 0.34281551613606653, "skyrl.ai/learning_rate": 4.324324324324325e-05, "num_sequences": 128, "num_tokens": 41346, "learning_rate": 4.324324324324325e-05, "train_mean_nll": 2.18609356880188, "progress": 0.5675675675675675, "time_total": 77.51598334312439}
+{"step": 43, "skyrl.ai/grad_norm": 0.30801024983275643, "skyrl.ai/learning_rate": 4.18918918918919e-05, "num_sequences": 128, "num_tokens": 35728, "learning_rate": 4.18918918918919e-05, "train_mean_nll": 2.389422655105591, "progress": 0.581081081081081, "time_total": 76.47622966766357}
diff --git a/tests/tinker/smoke_logs/sl_loop_c.log/checkpoints.jsonl b/tests/tinker/smoke_logs/sl_loop_c.log/checkpoints.jsonl
new file mode 100644
index 0000000000..4192b1f4ce
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_c.log/checkpoints.jsonl
@@ -0,0 +1,2 @@
+{"name": "000020", "batch": 20, "state_path": "tinker://model_3cb4546d/weights/000020"}
+{"name": "000040", "batch": 40, "state_path": "tinker://model_3cb4546d/weights/000040"}
diff --git a/tests/tinker/smoke_logs/sl_loop_c.log/code.diff b/tests/tinker/smoke_logs/sl_loop_c.log/code.diff
new file mode 100644
index 0000000000..d8e559755f
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_c.log/code.diff
@@ -0,0 +1,3 @@
+### repo: /home/etang/tinker-cookbook @ dacb835b31fb3f7a012851a1bc64c950d0495de3
+modules: tinker_cookbook
+(no local changes)
diff --git a/tests/tinker/smoke_logs/sl_loop_c.log/config.json b/tests/tinker/smoke_logs/sl_loop_c.log/config.json
new file mode 100644
index 0000000000..495b342f3a
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_c.log/config.json
@@ -0,0 +1,12 @@
+{
+  "base_url": "http://localhost:8000",
+  "log_path": "/tmp/sl_loop_c.log",
+  "model_name": "Qwen/Qwen3-0.6B",
+  "batch_size": 128,
+  "learning_rate": 0.0001,
+  "max_length": 32768,
+  "train_on_what": "last_assistant_message",
+  "lora_rank": 32,
+  "save_every": 20,
+  "ttl_seconds": 604800
+}
\ No newline at end of file
diff --git a/tests/tinker/smoke_logs/sl_loop_c.log/logs.log b/tests/tinker/smoke_logs/sl_loop_c.log/logs.log
new file mode 100644
index 0000000000..44c742f0a5
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_c.log/logs.log
@@ -0,0 +1,674 @@
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/sl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B train_on_what=LAST_ASSISTANT_MESSAGE lora_rank=32 log_path=/tmp/sl_loop_c.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/sl_loop_c.log
+__main__:51 [INFO] Using renderer: qwen3
+__main__:54 [INFO] Loading dataset...
+__main__:64 [INFO] Dropping last 28 examples to keep batch size uniform at 128
+__main__:67 [INFO] Train batches: 74
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_6fdc40ca
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/sl_loop_c.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/sl_loop_c.log
+tinker.lib.public_interfaces.service_client:159 [INFO] TrainingClient initialized for model model_3cb4546d
+__main__:87 [INFO] Training for 74 steps
+tinker_cookbook.supervised.common:188 [INFO] Weight reduction: 'mean' (token-mean loss)
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 0                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000100  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37834     │
+│ progress               │ 0.000000  │
+│ skyrl.ai/grad_norm     │ 1.419494  │
+│ skyrl.ai/learning_rate │ 0.000100  │
+│ time_total             │ 87.546265 │
+│ train_mean_nll         │ 2.937314  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 1                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000099  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35654     │
+│ progress               │ 0.013514  │
+│ skyrl.ai/grad_norm     │ 1.749522  │
+│ skyrl.ai/learning_rate │ 0.000099  │
+│ time_total             │ 40.494420 │
+│ train_mean_nll         │ 3.141417  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 2                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000097  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38474     │
+│ progress               │ 0.027027  │
+│ skyrl.ai/grad_norm     │ 1.374225  │
+│ skyrl.ai/learning_rate │ 0.000097  │
+│ time_total             │ 77.749164 │
+│ train_mean_nll         │ 2.757701  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 3                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000096  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39781     │
+│ progress               │ 0.040541  │
+│ skyrl.ai/grad_norm     │ 1.003446  │
+│ skyrl.ai/learning_rate │ 0.000096  │
+│ time_total             │ 77.538416 │
+│ train_mean_nll         │ 2.669955  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 4                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000095  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37346     │
+│ progress               │ 0.054054  │
+│ skyrl.ai/grad_norm     │ 1.014955  │
+│ skyrl.ai/learning_rate │ 0.000095  │
+│ time_total             │ 77.584343 │
+│ train_mean_nll         │ 2.654062  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 5                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000093  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38710     │
+│ progress               │ 0.067568  │
+│ skyrl.ai/grad_norm     │ 0.996195  │
+│ skyrl.ai/learning_rate │ 0.000093  │
+│ time_total             │ 76.493607 │
+│ train_mean_nll         │ 2.506083  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 6                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000092  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33356     │
+│ progress               │ 0.081081  │
+│ skyrl.ai/grad_norm     │ 0.641986  │
+│ skyrl.ai/learning_rate │ 0.000092  │
+│ time_total             │ 76.474288 │
+│ train_mean_nll         │ 2.669612  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 7                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000091  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39115     │
+│ progress               │ 0.094595  │
+│ skyrl.ai/grad_norm     │ 0.563422  │
+│ skyrl.ai/learning_rate │ 0.000091  │
+│ time_total             │ 76.499918 │
+│ train_mean_nll         │ 2.484150  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 8                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000089  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38640     │
+│ progress               │ 0.108108  │
+│ skyrl.ai/grad_norm     │ 0.586878  │
+│ skyrl.ai/learning_rate │ 0.000089  │
+│ time_total             │ 76.500674 │
+│ train_mean_nll         │ 2.451504  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 9                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000088  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38863     │
+│ progress               │ 0.121622  │
+│ skyrl.ai/grad_norm     │ 0.488777  │
+│ skyrl.ai/learning_rate │ 0.000088  │
+│ time_total             │ 76.483361 │
+│ train_mean_nll         │ 2.384386  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 10                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000086  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33155     │
+│ progress               │ 0.135135  │
+│ skyrl.ai/grad_norm     │ 0.620995  │
+│ skyrl.ai/learning_rate │ 0.000086  │
+│ time_total             │ 76.466245 │
+│ train_mean_nll         │ 2.525290  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 11                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000085  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38406     │
+│ progress               │ 0.148649  │
+│ skyrl.ai/grad_norm     │ 0.633048  │
+│ skyrl.ai/learning_rate │ 0.000085  │
+│ time_total             │ 76.584131 │
+│ train_mean_nll         │ 2.464255  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 12                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000084  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38843     │
+│ progress               │ 0.162162  │
+│ skyrl.ai/grad_norm     │ 0.579784  │
+│ skyrl.ai/learning_rate │ 0.000084  │
+│ time_total             │ 77.508669 │
+│ train_mean_nll         │ 2.335011  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 13                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000082  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35583     │
+│ progress               │ 0.175676  │
+│ skyrl.ai/grad_norm     │ 0.543443  │
+│ skyrl.ai/learning_rate │ 0.000082  │
+│ time_total             │ 77.479943 │
+│ train_mean_nll         │ 2.372522  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 14                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000081  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 36717     │
+│ progress               │ 0.189189  │
+│ skyrl.ai/grad_norm     │ 0.451880  │
+│ skyrl.ai/learning_rate │ 0.000081  │
+│ time_total             │ 77.489390 │
+│ train_mean_nll         │ 2.286112  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 15                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000080  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34002     │
+│ progress               │ 0.202703  │
+│ skyrl.ai/grad_norm     │ 0.472303  │
+│ skyrl.ai/learning_rate │ 0.000080  │
+│ time_total             │ 77.467145 │
+│ train_mean_nll         │ 2.414178  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 16                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000078  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38620     │
+│ progress               │ 0.216216  │
+│ skyrl.ai/grad_norm     │ 0.346693  │
+│ skyrl.ai/learning_rate │ 0.000078  │
+│ time_total             │ 77.586419 │
+│ train_mean_nll         │ 2.385986  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 17                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000077  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37639     │
+│ progress               │ 0.229730  │
+│ skyrl.ai/grad_norm     │ 0.330259  │
+│ skyrl.ai/learning_rate │ 0.000077  │
+│ time_total             │ 75.488438 │
+│ train_mean_nll         │ 2.257043  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 18                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000076  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34118     │
+│ progress               │ 0.243243  │
+│ skyrl.ai/grad_norm     │ 0.389382  │
+│ skyrl.ai/learning_rate │ 0.000076  │
+│ time_total             │ 77.459734 │
+│ train_mean_nll         │ 2.406650  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 19                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000074  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37178     │
+│ progress               │ 0.256757  │
+│ skyrl.ai/grad_norm     │ 0.367051  │
+│ skyrl.ai/learning_rate │ 0.000074  │
+│ time_total             │ 76.488872 │
+│ train_mean_nll         │ 2.391297  │
+└────────────────────────┴───────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_3cb4546d/weights/000020'}
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 20                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value      ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
+│ learning_rate          │ 0.000073   │
+│ num_sequences          │ 128        │
+│ num_tokens             │ 35309      │
+│ progress               │ 0.270270   │
+│ skyrl.ai/grad_norm     │ 0.345469   │
+│ skyrl.ai/learning_rate │ 0.000073   │
+│ time_total             │ 118.814218 │
+│ train_mean_nll         │ 2.355332   │
+└────────────────────────┴────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 21                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000072  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37229     │
+│ progress               │ 0.283784  │
+│ skyrl.ai/grad_norm     │ 0.347097  │
+│ skyrl.ai/learning_rate │ 0.000072  │
+│ time_total             │ 40.484439 │
+│ train_mean_nll         │ 2.357981  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 22                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000070  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 36087     │
+│ progress               │ 0.297297  │
+│ skyrl.ai/grad_norm     │ 0.436165  │
+│ skyrl.ai/learning_rate │ 0.000070  │
+│ time_total             │ 76.482713 │
+│ train_mean_nll         │ 2.351512  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 23                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000069  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33348     │
+│ progress               │ 0.310811  │
+│ skyrl.ai/grad_norm     │ 0.679087  │
+│ skyrl.ai/learning_rate │ 0.000069  │
+│ time_total             │ 75.449901 │
+│ train_mean_nll         │ 2.315460  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 24                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000068  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35210     │
+│ progress               │ 0.324324  │
+│ skyrl.ai/grad_norm     │ 0.366201  │
+│ skyrl.ai/learning_rate │ 0.000068  │
+│ time_total             │ 76.476222 │
+│ train_mean_nll         │ 2.382999  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 25                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000066  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34833     │
+│ progress               │ 0.337838  │
+│ skyrl.ai/grad_norm     │ 0.390126  │
+│ skyrl.ai/learning_rate │ 0.000066  │
+│ time_total             │ 76.471672 │
+│ train_mean_nll         │ 2.446065  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 26                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000065  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37256     │
+│ progress               │ 0.351351  │
+│ skyrl.ai/grad_norm     │ 0.381447  │
+│ skyrl.ai/learning_rate │ 0.000065  │
+│ time_total             │ 76.560245 │
+│ train_mean_nll         │ 2.409359  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 27                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000064  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33947     │
+│ progress               │ 0.364865  │
+│ skyrl.ai/grad_norm     │ 0.330323  │
+│ skyrl.ai/learning_rate │ 0.000064  │
+│ time_total             │ 76.468834 │
+│ train_mean_nll         │ 2.350033  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 28                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000062  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 32694     │
+│ progress               │ 0.378378  │
+│ skyrl.ai/grad_norm     │ 0.332719  │
+│ skyrl.ai/learning_rate │ 0.000062  │
+│ time_total             │ 76.454501 │
+│ train_mean_nll         │ 2.433611  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 29                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000061  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37440     │
+│ progress               │ 0.391892  │
+│ skyrl.ai/grad_norm     │ 0.358639  │
+│ skyrl.ai/learning_rate │ 0.000061  │
+│ time_total             │ 76.492569 │
+│ train_mean_nll         │ 2.170107  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 30                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000059  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39453     │
+│ progress               │ 0.405405  │
+│ skyrl.ai/grad_norm     │ 0.326137  │
+│ skyrl.ai/learning_rate │ 0.000059  │
+│ time_total             │ 76.509510 │
+│ train_mean_nll         │ 2.282948  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 31                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000058  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 36143     │
+│ progress               │ 0.418919  │
+│ skyrl.ai/grad_norm     │ 0.325924  │
+│ skyrl.ai/learning_rate │ 0.000058  │
+│ time_total             │ 75.544014 │
+│ train_mean_nll         │ 2.279570  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 32                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000057  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34808     │
+│ progress               │ 0.432432  │
+│ skyrl.ai/grad_norm     │ 0.386430  │
+│ skyrl.ai/learning_rate │ 0.000057  │
+│ time_total             │ 77.480128 │
+│ train_mean_nll         │ 2.355539  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 33                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000055  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 40403     │
+│ progress               │ 0.445946  │
+│ skyrl.ai/grad_norm     │ 0.293365  │
+│ skyrl.ai/learning_rate │ 0.000055  │
+│ time_total             │ 76.508817 │
+│ train_mean_nll         │ 2.232615  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 34                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000054  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33581     │
+│ progress               │ 0.459459  │
+│ skyrl.ai/grad_norm     │ 0.344365  │
+│ skyrl.ai/learning_rate │ 0.000054  │
+│ time_total             │ 76.445399 │
+│ train_mean_nll         │ 2.357985  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 35                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000053  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35478     │
+│ progress               │ 0.472973  │
+│ skyrl.ai/grad_norm     │ 0.271450  │
+│ skyrl.ai/learning_rate │ 0.000053  │
+│ time_total             │ 75.625342 │
+│ train_mean_nll         │ 2.342813  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 36                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000051  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37613     │
+│ progress               │ 0.486486  │
+│ skyrl.ai/grad_norm     │ 0.436700  │
+│ skyrl.ai/learning_rate │ 0.000051  │
+│ time_total             │ 76.592943 │
+│ train_mean_nll         │ 2.252346  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 37                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000050  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37243     │
+│ progress               │ 0.500000  │
+│ skyrl.ai/grad_norm     │ 0.323914  │
+│ skyrl.ai/learning_rate │ 0.000050  │
+│ time_total             │ 75.485813 │
+│ train_mean_nll         │ 2.455329  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 38                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000049  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34551     │
+│ progress               │ 0.513514  │
+│ skyrl.ai/grad_norm     │ 0.353730  │
+│ skyrl.ai/learning_rate │ 0.000049  │
+│ time_total             │ 76.462132 │
+│ train_mean_nll         │ 2.281997  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 39                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000047  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34720     │
+│ progress               │ 0.527027  │
+│ skyrl.ai/grad_norm     │ 0.325868  │
+│ skyrl.ai/learning_rate │ 0.000047  │
+│ time_total             │ 76.471973 │
+│ train_mean_nll         │ 2.224411  │
+└────────────────────────┴───────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_3cb4546d/weights/000040'}
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 40                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value      ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
+│ learning_rate          │ 0.000046   │
+│ num_sequences          │ 128        │
+│ num_tokens             │ 37150      │
+│ progress               │ 0.540541   │
+│ skyrl.ai/grad_norm     │ 0.291081   │
+│ skyrl.ai/learning_rate │ 0.000046   │
+│ time_total             │ 115.698803 │
+│ train_mean_nll         │ 2.320339   │
+└────────────────────────┴────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 41                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000045  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39832     │
+│ progress               │ 0.554054  │
+│ skyrl.ai/grad_norm     │ 0.330219  │
+│ skyrl.ai/learning_rate │ 0.000045  │
+│ time_total             │ 40.640421 │
+│ train_mean_nll         │ 2.251844  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 42                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000043  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 41346     │
+│ progress               │ 0.567568  │
+│ skyrl.ai/grad_norm     │ 0.348969  │
+│ skyrl.ai/learning_rate │ 0.000043  │
+│ time_total             │ 76.516095 │
+│ train_mean_nll         │ 2.186341  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_c.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 43                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000042  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35728     │
+│ progress               │ 0.581081  │
+│ skyrl.ai/grad_norm     │ 0.307291  │
+│ skyrl.ai/learning_rate │ 0.000042  │
+│ time_total             │ 76.478968 │
+│ train_mean_nll         │ 2.388552  │
+└────────────────────────┴───────────┘
diff --git a/tests/tinker/smoke_logs/sl_loop_c.log/metrics.jsonl b/tests/tinker/smoke_logs/sl_loop_c.log/metrics.jsonl
new file mode 100644
index 0000000000..16149b840f
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_c.log/metrics.jsonl
@@ -0,0 +1,44 @@
+{"step": 0, "skyrl.ai/grad_norm": 1.4194938927680651, "skyrl.ai/learning_rate": 0.0001, "num_sequences": 128, "num_tokens": 37834, "learning_rate": 0.0001, "train_mean_nll": 2.9373137950897217, "progress": 0.0, "time_total": 87.54626536369324}
+{"step": 1, "skyrl.ai/grad_norm": 1.7495224164784222, "skyrl.ai/learning_rate": 9.864864864864865e-05, "num_sequences": 128, "num_tokens": 35654, "learning_rate": 9.864864864864865e-05, "train_mean_nll": 3.141416549682617, "progress": 0.013513513513513514, "time_total": 40.49442005157471}
+{"step": 2, "skyrl.ai/grad_norm": 1.3742247910445407, "skyrl.ai/learning_rate": 9.729729729729731e-05, "num_sequences": 128, "num_tokens": 38474, "learning_rate": 9.729729729729731e-05, "train_mean_nll": 2.7577013969421387, "progress": 0.02702702702702703, "time_total": 77.74916410446167}
+{"step": 3, "skyrl.ai/grad_norm": 1.0034464811186297, "skyrl.ai/learning_rate": 9.594594594594595e-05, "num_sequences": 128, "num_tokens": 39781, "learning_rate": 9.594594594594595e-05, "train_mean_nll": 2.669954776763916, "progress": 0.04054054054054054, "time_total": 77.53841614723206}
+{"step": 4, "skyrl.ai/grad_norm": 1.014955270467524, "skyrl.ai/learning_rate": 9.45945945945946e-05, "num_sequences": 128, "num_tokens": 37346, "learning_rate": 9.45945945945946e-05, "train_mean_nll": 2.654061794281006, "progress": 0.05405405405405406, "time_total": 77.58434295654297}
+{"step": 5, "skyrl.ai/grad_norm": 0.9961948418275205, "skyrl.ai/learning_rate": 9.324324324324324e-05, "num_sequences": 128, "num_tokens": 38710, "learning_rate": 9.324324324324324e-05, "train_mean_nll": 2.5060834884643555, "progress": 0.06756756756756757, "time_total": 76.49360680580139}
+{"step": 6, "skyrl.ai/grad_norm": 0.6419861220214568, "skyrl.ai/learning_rate": 9.189189189189189e-05, "num_sequences": 128, "num_tokens": 33356, "learning_rate": 9.189189189189189e-05, "train_mean_nll": 2.669612407684326, "progress": 0.08108108108108109, "time_total": 76.47428822517395}
+{"step": 7, "skyrl.ai/grad_norm": 0.5634222153184144, "skyrl.ai/learning_rate": 9.054054054054055e-05, "num_sequences": 128, "num_tokens": 39115, "learning_rate": 9.054054054054055e-05, "train_mean_nll": 2.484149694442749, "progress": 0.0945945945945946, "time_total": 76.49991822242737}
+{"step": 8, "skyrl.ai/grad_norm": 0.5868782375274831, "skyrl.ai/learning_rate": 8.918918918918919e-05, "num_sequences": 128, "num_tokens": 38640, "learning_rate": 8.918918918918919e-05, "train_mean_nll": 2.4515039920806885, "progress": 0.10810810810810811, "time_total": 76.50067448616028}
+{"step": 9, "skyrl.ai/grad_norm": 0.48877668042056244, "skyrl.ai/learning_rate": 8.783783783783784e-05, "num_sequences": 128, "num_tokens": 38863, "learning_rate": 8.783783783783784e-05, "train_mean_nll": 2.384385585784912, "progress": 0.12162162162162163, "time_total": 76.48336100578308}
+{"step": 10, "skyrl.ai/grad_norm": 0.6209951118501068, "skyrl.ai/learning_rate": 8.64864864864865e-05, "num_sequences": 128, "num_tokens": 33155, "learning_rate": 8.64864864864865e-05, "train_mean_nll": 2.525290012359619, "progress": 0.13513513513513514, "time_total": 76.4662446975708}
+{"step": 11, "skyrl.ai/grad_norm": 0.6330475311859658, "skyrl.ai/learning_rate": 8.513513513513514e-05, "num_sequences": 128, "num_tokens": 38406, "learning_rate": 8.513513513513514e-05, "train_mean_nll": 2.46425461769104, "progress": 0.14864864864864866, "time_total": 76.58413076400757}
+{"step": 12, "skyrl.ai/grad_norm": 0.5797838691336037, "skyrl.ai/learning_rate": 8.378378378378379e-05, "num_sequences": 128, "num_tokens": 38843, "learning_rate": 8.378378378378379e-05, "train_mean_nll": 2.3350110054016113, "progress": 0.16216216216216217, "time_total": 77.50866913795471}
+{"step": 13, "skyrl.ai/grad_norm": 0.5434433763130382, "skyrl.ai/learning_rate": 8.243243243243243e-05, "num_sequences": 128, "num_tokens": 35583, "learning_rate": 8.243243243243243e-05, "train_mean_nll": 2.3725218772888184, "progress": 0.17567567567567569, "time_total": 77.47994303703308}
+{"step": 14, "skyrl.ai/grad_norm": 0.45187986236742045, "skyrl.ai/learning_rate": 8.108108108108108e-05, "num_sequences": 128, "num_tokens": 36717, "learning_rate": 8.108108108108108e-05, "train_mean_nll": 2.286111831665039, "progress": 0.1891891891891892, "time_total": 77.48939037322998}
+{"step": 15, "skyrl.ai/grad_norm": 0.4723029116416624, "skyrl.ai/learning_rate": 7.972972972972974e-05, "num_sequences": 128, "num_tokens": 34002, "learning_rate": 7.972972972972974e-05, "train_mean_nll": 2.414177894592285, "progress": 0.20270270270270271, "time_total": 77.46714520454407}
+{"step": 16, "skyrl.ai/grad_norm": 0.3466929473490256, "skyrl.ai/learning_rate": 7.837837837837838e-05, "num_sequences": 128, "num_tokens": 38620, "learning_rate": 7.837837837837838e-05, "train_mean_nll": 2.385986089706421, "progress": 0.21621621621621623, "time_total": 77.58641862869263}
+{"step": 17, "skyrl.ai/grad_norm": 0.3302586078980289, "skyrl.ai/learning_rate": 7.702702702702703e-05, "num_sequences": 128, "num_tokens": 37639, "learning_rate": 7.702702702702703e-05, "train_mean_nll": 2.2570433616638184, "progress": 0.22972972972972974, "time_total": 75.48843789100647}
+{"step": 18, "skyrl.ai/grad_norm": 0.3893822355807422, "skyrl.ai/learning_rate": 7.567567567567568e-05, "num_sequences": 128, "num_tokens": 34118, "learning_rate": 7.567567567567568e-05, "train_mean_nll": 2.4066498279571533, "progress": 0.24324324324324326, "time_total": 77.45973420143127}
+{"step": 19, "skyrl.ai/grad_norm": 0.3670506729696611, "skyrl.ai/learning_rate": 7.432432432432433e-05, "num_sequences": 128, "num_tokens": 37178, "learning_rate": 7.432432432432433e-05, "train_mean_nll": 2.3912971019744873, "progress": 0.25675675675675674, "time_total": 76.48887205123901}
+{"step": 20, "skyrl.ai/grad_norm": 0.3454687541752905, "skyrl.ai/learning_rate": 7.297297297297297e-05, "num_sequences": 128, "num_tokens": 35309, "learning_rate": 7.297297297297297e-05, "train_mean_nll": 2.355332136154175, "progress": 0.2702702702702703, "time_total": 118.81421756744385}
+{"step": 21, "skyrl.ai/grad_norm": 0.3470970650120757, "skyrl.ai/learning_rate": 7.162162162162162e-05, "num_sequences": 128, "num_tokens": 37229, "learning_rate": 7.162162162162162e-05, "train_mean_nll": 2.3579814434051514, "progress": 0.28378378378378377, "time_total": 40.48443937301636}
+{"step": 22, "skyrl.ai/grad_norm": 0.4361649818702482, "skyrl.ai/learning_rate": 7.027027027027026e-05, "num_sequences": 128, "num_tokens": 36087, "learning_rate": 7.027027027027026e-05, "train_mean_nll": 2.3515121936798096, "progress": 0.2972972972972973, "time_total": 76.4827127456665}
+{"step": 23, "skyrl.ai/grad_norm": 0.6790872312870737, "skyrl.ai/learning_rate": 6.891891891891892e-05, "num_sequences": 128, "num_tokens": 33348, "learning_rate": 6.891891891891892e-05, "train_mean_nll": 2.315459728240967, "progress": 0.3108108108108108, "time_total": 75.44990110397339}
+{"step": 24, "skyrl.ai/grad_norm": 0.3662012124359752, "skyrl.ai/learning_rate": 6.756756756756757e-05, "num_sequences": 128, "num_tokens": 35210, "learning_rate": 6.756756756756757e-05, "train_mean_nll": 2.3829991817474365, "progress": 0.32432432432432434, "time_total": 76.47622203826904}
+{"step": 25, "skyrl.ai/grad_norm": 0.39012587129983894, "skyrl.ai/learning_rate": 6.621621621621621e-05, "num_sequences": 128, "num_tokens": 34833, "learning_rate": 6.621621621621621e-05, "train_mean_nll": 2.4460654258728027, "progress": 0.33783783783783783, "time_total": 76.47167229652405}
+{"step": 26, "skyrl.ai/grad_norm": 0.3814466008615295, "skyrl.ai/learning_rate": 6.486486486486487e-05, "num_sequences": 128, "num_tokens": 37256, "learning_rate": 6.486486486486487e-05, "train_mean_nll": 2.4093592166900635, "progress": 0.35135135135135137, "time_total": 76.56024527549744}
+{"step": 27, "skyrl.ai/grad_norm": 0.33032280695631916, "skyrl.ai/learning_rate": 6.351351351351352e-05, "num_sequences": 128, "num_tokens": 33947, "learning_rate": 6.351351351351352e-05, "train_mean_nll": 2.3500328063964844, "progress": 0.36486486486486486, "time_total": 76.468834400177}
+{"step": 28, "skyrl.ai/grad_norm": 0.3327186649193775, "skyrl.ai/learning_rate": 6.216216216216216e-05, "num_sequences": 128, "num_tokens": 32694, "learning_rate": 6.216216216216216e-05, "train_mean_nll": 2.433610677719116, "progress": 0.3783783783783784, "time_total": 76.45450067520142}
+{"step": 29, "skyrl.ai/grad_norm": 0.35863921265811594, "skyrl.ai/learning_rate": 6.0810810810810814e-05, "num_sequences": 128, "num_tokens": 37440, "learning_rate": 6.0810810810810814e-05, "train_mean_nll": 2.170107126235962, "progress": 0.3918918918918919, "time_total": 76.49256873130798}
+{"step": 30, "skyrl.ai/grad_norm": 0.3261373695807218, "skyrl.ai/learning_rate": 5.945945945945945e-05, "num_sequences": 128, "num_tokens": 39453, "learning_rate": 5.945945945945945e-05, "train_mean_nll": 2.2829484939575195, "progress": 0.40540540540540543, "time_total": 76.50950980186462}
+{"step": 31, "skyrl.ai/grad_norm": 0.3259241339666833, "skyrl.ai/learning_rate": 5.810810810810812e-05, "num_sequences": 128, "num_tokens": 36143, "learning_rate": 5.810810810810812e-05, "train_mean_nll": 2.2795701026916504, "progress": 0.4189189189189189, "time_total": 75.5440137386322}
+{"step": 32, "skyrl.ai/grad_norm": 0.3864299967711568, "skyrl.ai/learning_rate": 5.6756756756756757e-05, "num_sequences": 128, "num_tokens": 34808, "learning_rate": 5.6756756756756757e-05, "train_mean_nll": 2.355539083480835, "progress": 0.43243243243243246, "time_total": 77.48012828826904}
+{"step": 33, "skyrl.ai/grad_norm": 0.29336461343164494, "skyrl.ai/learning_rate": 5.540540540540541e-05, "num_sequences": 128, "num_tokens": 40403, "learning_rate": 5.540540540540541e-05, "train_mean_nll": 2.232614517211914, "progress": 0.44594594594594594, "time_total": 76.50881671905518}
+{"step": 34, "skyrl.ai/grad_norm": 0.34436466791746617, "skyrl.ai/learning_rate": 5.405405405405406e-05, "num_sequences": 128, "num_tokens": 33581, "learning_rate": 5.405405405405406e-05, "train_mean_nll": 2.357985496520996, "progress": 0.4594594594594595, "time_total": 76.44539904594421}
+{"step": 35, "skyrl.ai/grad_norm": 0.2714500680008173, "skyrl.ai/learning_rate": 5.27027027027027e-05, "num_sequences": 128, "num_tokens": 35478, "learning_rate": 5.27027027027027e-05, "train_mean_nll": 2.3428127765655518, "progress": 0.47297297297297297, "time_total": 75.62534189224243}
+{"step": 36, "skyrl.ai/grad_norm": 0.43669954370291814, "skyrl.ai/learning_rate": 5.135135135135135e-05, "num_sequences": 128, "num_tokens": 37613, "learning_rate": 5.135135135135135e-05, "train_mean_nll": 2.2523462772369385, "progress": 0.4864864864864865, "time_total": 76.59294319152832}
+{"step": 37, "skyrl.ai/grad_norm": 0.3239142930227886, "skyrl.ai/learning_rate": 5e-05, "num_sequences": 128, "num_tokens": 37243, "learning_rate": 5e-05, "train_mean_nll": 2.455328941345215, "progress": 0.5, "time_total": 75.48581290245056}
+{"step": 38, "skyrl.ai/grad_norm": 0.35372994180702944, "skyrl.ai/learning_rate": 4.8648648648648654e-05, "num_sequences": 128, "num_tokens": 34551, "learning_rate": 4.8648648648648654e-05, "train_mean_nll": 2.281996726989746, "progress": 0.5135135135135135, "time_total": 76.46213150024414}
+{"step": 39, "skyrl.ai/grad_norm": 0.32586765378478905, "skyrl.ai/learning_rate": 4.7297297297297306e-05, "num_sequences": 128, "num_tokens": 34720, "learning_rate": 4.7297297297297306e-05, "train_mean_nll": 2.2244105339050293, "progress": 0.527027027027027, "time_total": 76.47197270393372}
+{"step": 40, "skyrl.ai/grad_norm": 0.29108088996807113, "skyrl.ai/learning_rate": 4.5945945945945944e-05, "num_sequences": 128, "num_tokens": 37150, "learning_rate": 4.5945945945945944e-05, "train_mean_nll": 2.3203389644622803, "progress": 0.5405405405405406, "time_total": 115.69880318641663}
+{"step": 41, "skyrl.ai/grad_norm": 0.3302190919782092, "skyrl.ai/learning_rate": 4.4594594594594596e-05, "num_sequences": 128, "num_tokens": 39832, "learning_rate": 4.4594594594594596e-05, "train_mean_nll": 2.2518436908721924, "progress": 0.5540540540540541, "time_total": 40.64042091369629}
+{"step": 42, "skyrl.ai/grad_norm": 0.34896948663718, "skyrl.ai/learning_rate": 4.324324324324325e-05, "num_sequences": 128, "num_tokens": 41346, "learning_rate": 4.324324324324325e-05, "train_mean_nll": 2.1863410472869873, "progress": 0.5675675675675675, "time_total": 76.51609539985657}
+{"step": 43, "skyrl.ai/grad_norm": 0.3072908759780453, "skyrl.ai/learning_rate": 4.18918918918919e-05, "num_sequences": 128, "num_tokens": 35728, "learning_rate": 4.18918918918919e-05, "train_mean_nll": 2.388551950454712, "progress": 0.581081081081081, "time_total": 76.47896790504456}
diff --git a/tests/tinker/smoke_logs/sl_loop_d.log/checkpoints.jsonl b/tests/tinker/smoke_logs/sl_loop_d.log/checkpoints.jsonl
new file mode 100644
index 0000000000..cbdbaddb1f
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_d.log/checkpoints.jsonl
@@ -0,0 +1,2 @@
+{"name": "000020", "batch": 20, "state_path": "tinker://model_a46bd29e/weights/000020"}
+{"name": "000040", "batch": 40, "state_path": "tinker://model_a46bd29e/weights/000040"}
diff --git a/tests/tinker/smoke_logs/sl_loop_d.log/code.diff b/tests/tinker/smoke_logs/sl_loop_d.log/code.diff
new file mode 100644
index 0000000000..d8e559755f
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_d.log/code.diff
@@ -0,0 +1,3 @@
+### repo: /home/etang/tinker-cookbook @ dacb835b31fb3f7a012851a1bc64c950d0495de3
+modules: tinker_cookbook
+(no local changes)
diff --git a/tests/tinker/smoke_logs/sl_loop_d.log/config.json b/tests/tinker/smoke_logs/sl_loop_d.log/config.json
new file mode 100644
index 0000000000..6da5f67bb4
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_d.log/config.json
@@ -0,0 +1,12 @@
+{
+  "base_url": "http://localhost:8000",
+  "log_path": "/tmp/sl_loop_d.log",
+  "model_name": "Qwen/Qwen3-0.6B",
+  "batch_size": 128,
+  "learning_rate": 0.0001,
+  "max_length": 32768,
+  "train_on_what": "last_assistant_message",
+  "lora_rank": 32,
+  "save_every": 20,
+  "ttl_seconds": 604800
+}
\ No newline at end of file
diff --git a/tests/tinker/smoke_logs/sl_loop_d.log/logs.log b/tests/tinker/smoke_logs/sl_loop_d.log/logs.log
new file mode 100644
index 0000000000..1f11432f6a
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_d.log/logs.log
@@ -0,0 +1,644 @@
+root:680 [INFO] Command line invocation: /home/etang/tinker-cookbook/tinker_cookbook/recipes/sl_loop.py base_url=http://localhost:8000 model_name=Qwen/Qwen3-0.6B train_on_what=LAST_ASSISTANT_MESSAGE lora_rank=32 log_path=/tmp/sl_loop_d.log
+tinker_cookbook.utils.ml_log:618 [INFO] Logging to: /tmp/sl_loop_d.log
+__main__:51 [INFO] Using renderer: qwen3
+__main__:54 [INFO] Loading dataset...
+__main__:64 [INFO] Dropping last 28 examples to keep batch size uniform at 128
+__main__:67 [INFO] Train batches: 74
+tinker.lib.public_interfaces.service_client:75 [INFO] ServiceClient initialized for session session_ead9aee6
+tinker_cookbook.checkpoint_utils:395 [INFO] No checkpoints found at /tmp/sl_loop_d.log/checkpoints.jsonl
+tinker_cookbook.checkpoint_utils:426 [INFO] No checkpoints found with key state_path in /tmp/sl_loop_d.log
+tinker.lib.public_interfaces.service_client:159 [INFO] TrainingClient initialized for model model_a46bd29e
+__main__:87 [INFO] Training for 74 steps
+tinker_cookbook.supervised.common:188 [INFO] Weight reduction: 'mean' (token-mean loss)
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 0                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000100  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37834     │
+│ progress               │ 0.000000  │
+│ skyrl.ai/grad_norm     │ 1.419466  │
+│ skyrl.ai/learning_rate │ 0.000100  │
+│ time_total             │ 77.674056 │
+│ train_mean_nll         │ 2.937314  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 1                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000099  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35654     │
+│ progress               │ 0.013514  │
+│ skyrl.ai/grad_norm     │ 1.766389  │
+│ skyrl.ai/learning_rate │ 0.000099  │
+│ time_total             │ 77.489237 │
+│ train_mean_nll         │ 3.162877  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 2                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000097  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38474     │
+│ progress               │ 0.027027  │
+│ skyrl.ai/grad_norm     │ 1.380754  │
+│ skyrl.ai/learning_rate │ 0.000097  │
+│ time_total             │ 38.622305 │
+│ train_mean_nll         │ 2.814990  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 3                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000096  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39781     │
+│ progress               │ 0.040541  │
+│ skyrl.ai/grad_norm     │ 1.256220  │
+│ skyrl.ai/learning_rate │ 0.000096  │
+│ time_total             │ 76.494820 │
+│ train_mean_nll         │ 2.717247  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 4                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000095  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37346     │
+│ progress               │ 0.054054  │
+│ skyrl.ai/grad_norm     │ 1.049826  │
+│ skyrl.ai/learning_rate │ 0.000095  │
+│ time_total             │ 76.496100 │
+│ train_mean_nll         │ 2.699676  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 5                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000093  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38710     │
+│ progress               │ 0.067568  │
+│ skyrl.ai/grad_norm     │ 1.067086  │
+│ skyrl.ai/learning_rate │ 0.000093  │
+│ time_total             │ 76.584077 │
+│ train_mean_nll         │ 2.537158  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 6                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000092  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33356     │
+│ progress               │ 0.081081  │
+│ skyrl.ai/grad_norm     │ 0.736274  │
+│ skyrl.ai/learning_rate │ 0.000092  │
+│ time_total             │ 76.462485 │
+│ train_mean_nll         │ 2.691379  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 7                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000091  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39115     │
+│ progress               │ 0.094595  │
+│ skyrl.ai/grad_norm     │ 0.610632  │
+│ skyrl.ai/learning_rate │ 0.000091  │
+│ time_total             │ 76.602791 │
+│ train_mean_nll         │ 2.499062  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 8                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000089  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38640     │
+│ progress               │ 0.108108  │
+│ skyrl.ai/grad_norm     │ 0.598359  │
+│ skyrl.ai/learning_rate │ 0.000089  │
+│ time_total             │ 76.482659 │
+│ train_mean_nll         │ 2.461901  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 9                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000088  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38863     │
+│ progress               │ 0.121622  │
+│ skyrl.ai/grad_norm     │ 0.513877  │
+│ skyrl.ai/learning_rate │ 0.000088  │
+│ time_total             │ 76.514921 │
+│ train_mean_nll         │ 2.393402  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 10                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000086  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33155     │
+│ progress               │ 0.135135  │
+│ skyrl.ai/grad_norm     │ 0.675169  │
+│ skyrl.ai/learning_rate │ 0.000086  │
+│ time_total             │ 77.464031 │
+│ train_mean_nll         │ 2.537810  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 11                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000085  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38406     │
+│ progress               │ 0.148649  │
+│ skyrl.ai/grad_norm     │ 0.626751  │
+│ skyrl.ai/learning_rate │ 0.000085  │
+│ time_total             │ 77.508308 │
+│ train_mean_nll         │ 2.474285  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 12                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000084  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38843     │
+│ progress               │ 0.162162  │
+│ skyrl.ai/grad_norm     │ 0.587304  │
+│ skyrl.ai/learning_rate │ 0.000084  │
+│ time_total             │ 77.612279 │
+│ train_mean_nll         │ 2.344043  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 13                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000082  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35583     │
+│ progress               │ 0.175676  │
+│ skyrl.ai/grad_norm     │ 0.545789  │
+│ skyrl.ai/learning_rate │ 0.000082  │
+│ time_total             │ 77.486650 │
+│ train_mean_nll         │ 2.380935  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 14                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000081  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 36717     │
+│ progress               │ 0.189189  │
+│ skyrl.ai/grad_norm     │ 0.474083  │
+│ skyrl.ai/learning_rate │ 0.000081  │
+│ time_total             │ 77.482110 │
+│ train_mean_nll         │ 2.289858  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 15                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000080  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34002     │
+│ progress               │ 0.202703  │
+│ skyrl.ai/grad_norm     │ 0.511856  │
+│ skyrl.ai/learning_rate │ 0.000080  │
+│ time_total             │ 76.445762 │
+│ train_mean_nll         │ 2.420629  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 16                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000078  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 38620     │
+│ progress               │ 0.216216  │
+│ skyrl.ai/grad_norm     │ 0.358070  │
+│ skyrl.ai/learning_rate │ 0.000078  │
+│ time_total             │ 76.652211 │
+│ train_mean_nll         │ 2.389158  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 17                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000077  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37639     │
+│ progress               │ 0.229730  │
+│ skyrl.ai/grad_norm     │ 0.336696  │
+│ skyrl.ai/learning_rate │ 0.000077  │
+│ time_total             │ 76.554973 │
+│ train_mean_nll         │ 2.258745  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 18                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000076  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34118     │
+│ progress               │ 0.243243  │
+│ skyrl.ai/grad_norm     │ 0.420464  │
+│ skyrl.ai/learning_rate │ 0.000076  │
+│ time_total             │ 76.455621 │
+│ train_mean_nll         │ 2.409728  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 19                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000074  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37178     │
+│ progress               │ 0.256757  │
+│ skyrl.ai/grad_norm     │ 0.369806  │
+│ skyrl.ai/learning_rate │ 0.000074  │
+│ time_total             │ 42.479602 │
+│ train_mean_nll         │ 2.392418  │
+└────────────────────────┴───────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_a46bd29e/weights/000020'}
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 20                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value      ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
+│ learning_rate          │ 0.000073   │
+│ num_sequences          │ 128        │
+│ num_tokens             │ 35309      │
+│ progress               │ 0.270270   │
+│ skyrl.ai/grad_norm     │ 0.347546   │
+│ skyrl.ai/learning_rate │ 0.000073   │
+│ time_total             │ 116.676181 │
+│ train_mean_nll         │ 2.357243   │
+└────────────────────────┴────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 21                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000072  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37229     │
+│ progress               │ 0.283784  │
+│ skyrl.ai/grad_norm     │ 0.352529  │
+│ skyrl.ai/learning_rate │ 0.000072  │
+│ time_total             │ 76.480985 │
+│ train_mean_nll         │ 2.359369  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 22                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000070  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 36087     │
+│ progress               │ 0.297297  │
+│ skyrl.ai/grad_norm     │ 0.439494  │
+│ skyrl.ai/learning_rate │ 0.000070  │
+│ time_total             │ 75.470494 │
+│ train_mean_nll         │ 2.353091  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 23                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000069  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33348     │
+│ progress               │ 0.310811  │
+│ skyrl.ai/grad_norm     │ 0.682835  │
+│ skyrl.ai/learning_rate │ 0.000069  │
+│ time_total             │ 76.461006 │
+│ train_mean_nll         │ 2.316307  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 24                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000068  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35210     │
+│ progress               │ 0.324324  │
+│ skyrl.ai/grad_norm     │ 0.366110  │
+│ skyrl.ai/learning_rate │ 0.000068  │
+│ time_total             │ 76.467583 │
+│ train_mean_nll         │ 2.383547  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 25                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000066  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34833     │
+│ progress               │ 0.337838  │
+│ skyrl.ai/grad_norm     │ 0.395728  │
+│ skyrl.ai/learning_rate │ 0.000066  │
+│ time_total             │ 77.473791 │
+│ train_mean_nll         │ 2.448057  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 26                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000065  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37256     │
+│ progress               │ 0.351351  │
+│ skyrl.ai/grad_norm     │ 0.383532  │
+│ skyrl.ai/learning_rate │ 0.000065  │
+│ time_total             │ 76.483187 │
+│ train_mean_nll         │ 2.410551  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 27                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000064  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33947     │
+│ progress               │ 0.364865  │
+│ skyrl.ai/grad_norm     │ 0.334303  │
+│ skyrl.ai/learning_rate │ 0.000064  │
+│ time_total             │ 76.542615 │
+│ train_mean_nll         │ 2.351529  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 28                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000062  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 32694     │
+│ progress               │ 0.378378  │
+│ skyrl.ai/grad_norm     │ 0.336119  │
+│ skyrl.ai/learning_rate │ 0.000062  │
+│ time_total             │ 75.442138 │
+│ train_mean_nll         │ 2.435189  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 29                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000061  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37440     │
+│ progress               │ 0.391892  │
+│ skyrl.ai/grad_norm     │ 0.363025  │
+│ skyrl.ai/learning_rate │ 0.000061  │
+│ time_total             │ 76.492469 │
+│ train_mean_nll         │ 2.170725  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 30                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000059  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39453     │
+│ progress               │ 0.405405  │
+│ skyrl.ai/grad_norm     │ 0.329587  │
+│ skyrl.ai/learning_rate │ 0.000059  │
+│ time_total             │ 76.527954 │
+│ train_mean_nll         │ 2.284290  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 31                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000058  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 36143     │
+│ progress               │ 0.418919  │
+│ skyrl.ai/grad_norm     │ 0.324875  │
+│ skyrl.ai/learning_rate │ 0.000058  │
+│ time_total             │ 76.648657 │
+│ train_mean_nll         │ 2.280195  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 32                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000057  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34808     │
+│ progress               │ 0.432432  │
+│ skyrl.ai/grad_norm     │ 0.386744  │
+│ skyrl.ai/learning_rate │ 0.000057  │
+│ time_total             │ 76.552472 │
+│ train_mean_nll         │ 2.355594  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 33                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000055  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 40403     │
+│ progress               │ 0.445946  │
+│ skyrl.ai/grad_norm     │ 0.291313  │
+│ skyrl.ai/learning_rate │ 0.000055  │
+│ time_total             │ 76.512269 │
+│ train_mean_nll         │ 2.234097  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 34                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000054  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 33581     │
+│ progress               │ 0.459459  │
+│ skyrl.ai/grad_norm     │ 0.323469  │
+│ skyrl.ai/learning_rate │ 0.000054  │
+│ time_total             │ 75.450547 │
+│ train_mean_nll         │ 2.358581  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 35                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000053  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 35478     │
+│ progress               │ 0.472973  │
+│ skyrl.ai/grad_norm     │ 0.270337  │
+│ skyrl.ai/learning_rate │ 0.000053  │
+│ time_total             │ 76.495394 │
+│ train_mean_nll         │ 2.343688  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 36                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000051  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37613     │
+│ progress               │ 0.486486  │
+│ skyrl.ai/grad_norm     │ 0.438996  │
+│ skyrl.ai/learning_rate │ 0.000051  │
+│ time_total             │ 76.483823 │
+│ train_mean_nll         │ 2.253183  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 37                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000050  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 37243     │
+│ progress               │ 0.500000  │
+│ skyrl.ai/grad_norm     │ 0.328100  │
+│ skyrl.ai/learning_rate │ 0.000050  │
+│ time_total             │ 76.559658 │
+│ train_mean_nll         │ 2.456638  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 38                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000049  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34551     │
+│ progress               │ 0.513514  │
+│ skyrl.ai/grad_norm     │ 0.351431  │
+│ skyrl.ai/learning_rate │ 0.000049  │
+│ time_total             │ 75.463238 │
+│ train_mean_nll         │ 2.281233  │
+└────────────────────────┴───────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 39                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000047  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 34720     │
+│ progress               │ 0.527027  │
+│ skyrl.ai/grad_norm     │ 0.326205  │
+│ skyrl.ai/learning_rate │ 0.000047  │
+│ time_total             │ 40.466141 │
+│ train_mean_nll         │ 2.225522  │
+└────────────────────────┴───────────┘
+tinker_cookbook.checkpoint_utils:466 [INFO] Saved checkpoints: {'state_path': 'tinker://model_a46bd29e/weights/000040'}
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+                Step 40                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
+┃ Metric                 ┃ Value      ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
+│ learning_rate          │ 0.000046   │
+│ num_sequences          │ 128        │
+│ num_tokens             │ 37150      │
+│ progress               │ 0.540541   │
+│ skyrl.ai/grad_norm     │ 0.291013   │
+│ skyrl.ai/learning_rate │ 0.000046   │
+│ time_total             │ 115.610721 │
+│ train_mean_nll         │ 2.320696   │
+└────────────────────────┴────────────┘
+tinker_cookbook.utils.ml_log:206 [INFO] Wrote metrics to /tmp/sl_loop_d.log/metrics.jsonl
+tinker_cookbook.utils.ml_log:279 [INFO] 
+               Step 41                
+┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃ Metric                 ┃ Value     ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
+│ learning_rate          │ 0.000045  │
+│ num_sequences          │ 128       │
+│ num_tokens             │ 39832     │
+│ progress               │ 0.554054  │
+│ skyrl.ai/grad_norm     │ 0.331071  │
+│ skyrl.ai/learning_rate │ 0.000045  │
+│ time_total             │ 77.513973 │
+│ train_mean_nll         │ 2.251503  │
+└────────────────────────┴───────────┘
diff --git a/tests/tinker/smoke_logs/sl_loop_d.log/metrics.jsonl b/tests/tinker/smoke_logs/sl_loop_d.log/metrics.jsonl
new file mode 100644
index 0000000000..13e7652554
--- /dev/null
+++ b/tests/tinker/smoke_logs/sl_loop_d.log/metrics.jsonl
@@ -0,0 +1,42 @@
+{"step": 0, "skyrl.ai/grad_norm": 1.4194660110892312, "skyrl.ai/learning_rate": 0.0001, "num_sequences": 128, "num_tokens": 37834, "learning_rate": 0.0001, "train_mean_nll": 2.9373137950897217, "progress": 0.0, "time_total": 77.67405557632446}
+{"step": 1, "skyrl.ai/grad_norm": 1.7663889894650249, "skyrl.ai/learning_rate": 9.864864864864865e-05, "num_sequences": 128, "num_tokens": 35654, "learning_rate": 9.864864864864865e-05, "train_mean_nll": 3.162876844406128, "progress": 0.013513513513513514, "time_total": 77.48923659324646}
+{"step": 2, "skyrl.ai/grad_norm": 1.3807542213211028, "skyrl.ai/learning_rate": 9.729729729729731e-05, "num_sequences": 128, "num_tokens": 38474, "learning_rate": 9.729729729729731e-05, "train_mean_nll": 2.8149900436401367, "progress": 0.02702702702702703, "time_total": 38.622305393218994}
+{"step": 3, "skyrl.ai/grad_norm": 1.2562199812002224, "skyrl.ai/learning_rate": 9.594594594594595e-05, "num_sequences": 128, "num_tokens": 39781, "learning_rate": 9.594594594594595e-05, "train_mean_nll": 2.717247486114502, "progress": 0.04054054054054054, "time_total": 76.49482035636902}
+{"step": 4, "skyrl.ai/grad_norm": 1.049826376047764, "skyrl.ai/learning_rate": 9.45945945945946e-05, "num_sequences": 128, "num_tokens": 37346, "learning_rate": 9.45945945945946e-05, "train_mean_nll": 2.699676275253296, "progress": 0.05405405405405406, "time_total": 76.49610042572021}
+{"step": 5, "skyrl.ai/grad_norm": 1.067086083437052, "skyrl.ai/learning_rate": 9.324324324324324e-05, "num_sequences": 128, "num_tokens": 38710, "learning_rate": 9.324324324324324e-05, "train_mean_nll": 2.537158250808716, "progress": 0.06756756756756757, "time_total": 76.58407664299011}
+{"step": 6, "skyrl.ai/grad_norm": 0.7362744137224916, "skyrl.ai/learning_rate": 9.189189189189189e-05, "num_sequences": 128, "num_tokens": 33356, "learning_rate": 9.189189189189189e-05, "train_mean_nll": 2.6913793087005615, "progress": 0.08108108108108109, "time_total": 76.46248507499695}
+{"step": 7, "skyrl.ai/grad_norm": 0.610632113256186, "skyrl.ai/learning_rate": 9.054054054054055e-05, "num_sequences": 128, "num_tokens": 39115, "learning_rate": 9.054054054054055e-05, "train_mean_nll": 2.4990618228912354, "progress": 0.0945945945945946, "time_total": 76.60279083251953}
+{"step": 8, "skyrl.ai/grad_norm": 0.5983594111397604, "skyrl.ai/learning_rate": 8.918918918918919e-05, "num_sequences": 128, "num_tokens": 38640, "learning_rate": 8.918918918918919e-05, "train_mean_nll": 2.4619009494781494, "progress": 0.10810810810810811, "time_total": 76.48265886306763}
+{"step": 9, "skyrl.ai/grad_norm": 0.513877102427449, "skyrl.ai/learning_rate": 8.783783783783784e-05, "num_sequences": 128, "num_tokens": 38863, "learning_rate": 8.783783783783784e-05, "train_mean_nll": 2.393401622772217, "progress": 0.12162162162162163, "time_total": 76.51492071151733}
+{"step": 10, "skyrl.ai/grad_norm": 0.6751687024483758, "skyrl.ai/learning_rate": 8.64864864864865e-05, "num_sequences": 128, "num_tokens": 33155, "learning_rate": 8.64864864864865e-05, "train_mean_nll": 2.5378100872039795, "progress": 0.13513513513513514, "time_total": 77.46403121948242}
+{"step": 11, "skyrl.ai/grad_norm": 0.6267508302218245, "skyrl.ai/learning_rate": 8.513513513513514e-05, "num_sequences": 128, "num_tokens": 38406, "learning_rate": 8.513513513513514e-05, "train_mean_nll": 2.4742848873138428, "progress": 0.14864864864864866, "time_total": 77.50830841064453}
+{"step": 12, "skyrl.ai/grad_norm": 0.5873035274858249, "skyrl.ai/learning_rate": 8.378378378378379e-05, "num_sequences": 128, "num_tokens": 38843, "learning_rate": 8.378378378378379e-05, "train_mean_nll": 2.344043254852295, "progress": 0.16216216216216217, "time_total": 77.61227917671204}
+{"step": 13, "skyrl.ai/grad_norm": 0.5457892948565732, "skyrl.ai/learning_rate": 8.243243243243243e-05, "num_sequences": 128, "num_tokens": 35583, "learning_rate": 8.243243243243243e-05, "train_mean_nll": 2.380934953689575, "progress": 0.17567567567567569, "time_total": 77.48664999008179}
+{"step": 14, "skyrl.ai/grad_norm": 0.4740833931419579, "skyrl.ai/learning_rate": 8.108108108108108e-05, "num_sequences": 128, "num_tokens": 36717, "learning_rate": 8.108108108108108e-05, "train_mean_nll": 2.289858102798462, "progress": 0.1891891891891892, "time_total": 77.48211026191711}
+{"step": 15, "skyrl.ai/grad_norm": 0.5118555948866679, "skyrl.ai/learning_rate": 7.972972972972974e-05, "num_sequences": 128, "num_tokens": 34002, "learning_rate": 7.972972972972974e-05, "train_mean_nll": 2.4206290245056152, "progress": 0.20270270270270271, "time_total": 76.44576215744019}
+{"step": 16, "skyrl.ai/grad_norm": 0.358069682251165, "skyrl.ai/learning_rate": 7.837837837837838e-05, "num_sequences": 128, "num_tokens": 38620, "learning_rate": 7.837837837837838e-05, "train_mean_nll": 2.38915753364563, "progress": 0.21621621621621623, "time_total": 76.6522114276886}
+{"step": 17, "skyrl.ai/grad_norm": 0.33669552556994264, "skyrl.ai/learning_rate": 7.702702702702703e-05, "num_sequences": 128, "num_tokens": 37639, "learning_rate": 7.702702702702703e-05, "train_mean_nll": 2.258744955062866, "progress": 0.22972972972972974, "time_total": 76.55497312545776}
+{"step": 18, "skyrl.ai/grad_norm": 0.4204640990772262, "skyrl.ai/learning_rate": 7.567567567567568e-05, "num_sequences": 128, "num_tokens": 34118, "learning_rate": 7.567567567567568e-05, "train_mean_nll": 2.4097280502319336, "progress": 0.24324324324324326, "time_total": 76.45562148094177}
+{"step": 19, "skyrl.ai/grad_norm": 0.3698055193101831, "skyrl.ai/learning_rate": 7.432432432432433e-05, "num_sequences": 128, "num_tokens": 37178, "learning_rate": 7.432432432432433e-05, "train_mean_nll": 2.392418146133423, "progress": 0.25675675675675674, "time_total": 42.479602098464966}
+{"step": 20, "skyrl.ai/grad_norm": 0.34754613139614876, "skyrl.ai/learning_rate": 7.297297297297297e-05, "num_sequences": 128, "num_tokens": 35309, "learning_rate": 7.297297297297297e-05, "train_mean_nll": 2.3572428226470947, "progress": 0.2702702702702703, "time_total": 116.67618131637573}
+{"step": 21, "skyrl.ai/grad_norm": 0.35252946750760694, "skyrl.ai/learning_rate": 7.162162162162162e-05, "num_sequences": 128, "num_tokens": 37229, "learning_rate": 7.162162162162162e-05, "train_mean_nll": 2.3593690395355225, "progress": 0.28378378378378377, "time_total": 76.48098468780518}
+{"step": 22, "skyrl.ai/grad_norm": 0.4394941353205905, "skyrl.ai/learning_rate": 7.027027027027026e-05, "num_sequences": 128, "num_tokens": 36087, "learning_rate": 7.027027027027026e-05, "train_mean_nll": 2.353090763092041, "progress": 0.2972972972972973, "time_total": 75.4704942703247}
+{"step": 23, "skyrl.ai/grad_norm": 0.6828352725772945, "skyrl.ai/learning_rate": 6.891891891891892e-05, "num_sequences": 128, "num_tokens": 33348, "learning_rate": 6.891891891891892e-05, "train_mean_nll": 2.3163068294525146, "progress": 0.3108108108108108, "time_total": 76.46100568771362}
+{"step": 24, "skyrl.ai/grad_norm": 0.36611048019937215, "skyrl.ai/learning_rate": 6.756756756756757e-05, "num_sequences": 128, "num_tokens": 35210, "learning_rate": 6.756756756756757e-05, "train_mean_nll": 2.383546829223633, "progress": 0.32432432432432434, "time_total": 76.46758341789246}
+{"step": 25, "skyrl.ai/grad_norm": 0.39572772281935936, "skyrl.ai/learning_rate": 6.621621621621621e-05, "num_sequences": 128, "num_tokens": 34833, "learning_rate": 6.621621621621621e-05, "train_mean_nll": 2.4480574131011963, "progress": 0.33783783783783783, "time_total": 77.47379088401794}
+{"step": 26, "skyrl.ai/grad_norm": 0.3835315956991299, "skyrl.ai/learning_rate": 6.486486486486487e-05, "num_sequences": 128, "num_tokens": 37256, "learning_rate": 6.486486486486487e-05, "train_mean_nll": 2.410550832748413, "progress": 0.35135135135135137, "time_total": 76.4831874370575}
+{"step": 27, "skyrl.ai/grad_norm": 0.33430281733441997, "skyrl.ai/learning_rate": 6.351351351351352e-05, "num_sequences": 128, "num_tokens": 33947, "learning_rate": 6.351351351351352e-05, "train_mean_nll": 2.351529359817505, "progress": 0.36486486486486486, "time_total": 76.54261541366577}
+{"step": 28, "skyrl.ai/grad_norm": 0.3361192034416266, "skyrl.ai/learning_rate": 6.216216216216216e-05, "num_sequences": 128, "num_tokens": 32694, "learning_rate": 6.216216216216216e-05, "train_mean_nll": 2.4351885318756104, "progress": 0.3783783783783784, "time_total": 75.44213843345642}
+{"step": 29, "skyrl.ai/grad_norm": 0.36302524693495586, "skyrl.ai/learning_rate": 6.0810810810810814e-05, "num_sequences": 128, "num_tokens": 37440, "learning_rate": 6.0810810810810814e-05, "train_mean_nll": 2.170724630355835, "progress": 0.3918918918918919, "time_total": 76.49246883392334}
+{"step": 30, "skyrl.ai/grad_norm": 0.32958736842603464, "skyrl.ai/learning_rate": 5.945945945945945e-05, "num_sequences": 128, "num_tokens": 39453, "learning_rate": 5.945945945945945e-05, "train_mean_nll": 2.284289598464966, "progress": 0.40540540540540543, "time_total": 76.5279541015625}
+{"step": 31, "skyrl.ai/grad_norm": 0.3248749033717518, "skyrl.ai/learning_rate": 5.810810810810812e-05, "num_sequences": 128, "num_tokens": 36143, "learning_rate": 5.810810810810812e-05, "train_mean_nll": 2.2801947593688965, "progress": 0.4189189189189189, "time_total": 76.64865732192993}
+{"step": 32, "skyrl.ai/grad_norm": 0.386744315333098, "skyrl.ai/learning_rate": 5.6756756756756757e-05, "num_sequences": 128, "num_tokens": 34808, "learning_rate": 5.6756756756756757e-05, "train_mean_nll": 2.3555939197540283, "progress": 0.43243243243243246, "time_total": 76.55247235298157}
+{"step": 33, "skyrl.ai/grad_norm": 0.29131346700737404, "skyrl.ai/learning_rate": 5.540540540540541e-05, "num_sequences": 128, "num_tokens": 40403, "learning_rate": 5.540540540540541e-05, "train_mean_nll": 2.2340970039367676, "progress": 0.44594594594594594, "time_total": 76.51226878166199}
+{"step": 34, "skyrl.ai/grad_norm": 0.32346850076171424, "skyrl.ai/learning_rate": 5.405405405405406e-05, "num_sequences": 128, "num_tokens": 33581, "learning_rate": 5.405405405405406e-05, "train_mean_nll": 2.3585805892944336, "progress": 0.4594594594594595, "time_total": 75.45054650306702}
+{"step": 35, "skyrl.ai/grad_norm": 0.27033689587266685, "skyrl.ai/learning_rate": 5.27027027027027e-05, "num_sequences": 128, "num_tokens": 35478, "learning_rate": 5.27027027027027e-05, "train_mean_nll": 2.3436880111694336, "progress": 0.47297297297297297, "time_total": 76.4953944683075}
+{"step": 36, "skyrl.ai/grad_norm": 0.4389957845644059, "skyrl.ai/learning_rate": 5.135135135135135e-05, "num_sequences": 128, "num_tokens": 37613, "learning_rate": 5.135135135135135e-05, "train_mean_nll": 2.253182888031006, "progress": 0.4864864864864865, "time_total": 76.48382329940796}
+{"step": 37, "skyrl.ai/grad_norm": 0.32809987426108383, "skyrl.ai/learning_rate": 5e-05, "num_sequences": 128, "num_tokens": 37243, "learning_rate": 5e-05, "train_mean_nll": 2.4566383361816406, "progress": 0.5, "time_total": 76.55965781211853}
+{"step": 38, "skyrl.ai/grad_norm": 0.3514309849054616, "skyrl.ai/learning_rate": 4.8648648648648654e-05, "num_sequences": 128, "num_tokens": 34551, "learning_rate": 4.8648648648648654e-05, "train_mean_nll": 2.2812325954437256, "progress": 0.5135135135135135, "time_total": 75.46323847770691}
+{"step": 39, "skyrl.ai/grad_norm": 0.32620477795966657, "skyrl.ai/learning_rate": 4.7297297297297306e-05, "num_sequences": 128, "num_tokens": 34720, "learning_rate": 4.7297297297297306e-05, "train_mean_nll": 2.2255215644836426, "progress": 0.527027027027027, "time_total": 40.46614122390747}
+{"step": 40, "skyrl.ai/grad_norm": 0.29101311599366725, "skyrl.ai/learning_rate": 4.5945945945945944e-05, "num_sequences": 128, "num_tokens": 37150, "learning_rate": 4.5945945945945944e-05, "train_mean_nll": 2.3206958770751953, "progress": 0.5405405405405406, "time_total": 115.61072111129761}
+{"step": 41, "skyrl.ai/grad_norm": 0.3310711724332053, "skyrl.ai/learning_rate": 4.4594594594594596e-05, "num_sequences": 128, "num_tokens": 39832, "learning_rate": 4.4594594594594596e-05, "train_mean_nll": 2.2515032291412354, "progress": 0.5540540540540541, "time_total": 77.5139729976654}

From d178ca0fe26d2a611625e4eacbbd6464b3739d50 Mon Sep 17 00:00:00 2001
From: Eric Tang <erictang000@gmail.com>
Date: Thu, 7 May 2026 01:52:20 +0000
Subject: [PATCH 21/21] [multi-lora-rl] Reset PR #1579 test files to upstream
 main version

PR #1579's test files evolved between when our branch cherry-picked
them and when the PR was merged to main (model field is now optional
via _resolve_model). Reset them to main's version to keep the
multi_lora_rl diff strictly to RL functionality.
---
 skyrl/benchmarks/load_test_concurrency.py          |  2 +-
 .../test_new_inference_generation.py               |  4 +---
 .../test_remote_inference_client_chat_template.py  |  2 +-
 .../test_vlm_inference_generation.py               |  3 +--
 .../gpu/gpu_ci/megatron/test_router_replay.py      |  1 -
 .../gpu/gpu_ci/test_engine_generation.py           |  8 ++++----
 .../gpu_ci/test_pause_and_continue_generation.py   |  2 +-
 .../gpu/gpu_ci/test_skyrl_gym_generator.py         |  1 -
 .../gpu/gpu_ci/test_skyrl_vlm_gym_generator.py     |  1 -
 tests/train/generators/test_skyrl_gym_generator.py | 14 --------------
 .../test_skyrl_gym_generator_chat_templating.py    |  1 -
 tests/train/generators/test_skyrl_vlm_generator.py |  1 -
 12 files changed, 9 insertions(+), 31 deletions(-)

diff --git a/skyrl/benchmarks/load_test_concurrency.py b/skyrl/benchmarks/load_test_concurrency.py
index 88b36fbb8f..036011f89d 100644
--- a/skyrl/benchmarks/load_test_concurrency.py
+++ b/skyrl/benchmarks/load_test_concurrency.py
@@ -186,7 +186,7 @@ async def fire_client_generate(
 
     async def _call(idx: int):
         try:
-            return await client.generate(input_batch, model=client.model_name)
+            return await client.generate(input_batch)
         except Exception as e:
             raise RuntimeError(f"request {idx}: {type(e).__name__}: {e}") from e
 
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_new_inference_generation.py b/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_new_inference_generation.py
index 95338f7c0f..e622f132fc 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_new_inference_generation.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_new_inference_generation.py
@@ -563,7 +563,7 @@ async def test_client_generate(vllm_server: InferenceEngineState):
         sampling_params=sampling_params,
     )
 
-    output = await client.generate(engine_input, model=client.model_name)
+    output = await client.generate(engine_input)
 
     assert len(output["responses"]) == 1
     assert len(output["response_ids"]) == 1
@@ -595,7 +595,6 @@ async def test_client_tokenize_detokenize_roundtrip(vllm_server: InferenceEngine
 
 def _build_sample_payload(
     token_ids: List[int],
-    model: str = MODEL_QWEN2_5,
     num_samples: int = 1,
     sampling_params: Dict[str, Any] | None = None,
     session_id: str | None = None,
@@ -604,7 +603,6 @@ def _build_sample_payload(
 ) -> Dict[str, Any]:
     """Build a Tinker-format sample request payload."""
     body: Dict[str, Any] = {
-        "model": model,
         "prompt": {"chunks": [{"tokens": token_ids}]},
         "num_samples": num_samples,
         "sampling_params": sampling_params or {"temperature": 0.7, "max_tokens": 64},
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_remote_inference_client_chat_template.py b/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_remote_inference_client_chat_template.py
index 1ef6e10bc9..300fac372a 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_remote_inference_client_chat_template.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_remote_inference_client_chat_template.py
@@ -90,7 +90,7 @@ async def test_custom_chat_template(ray_init_fixture, use_custom_template: bool)
             prompt_token_ids=[prompt_token_ids],
             sampling_params={"max_tokens": 10},
         )
-        output = await client.generate(engine_input, model=client.model_name)
+        output = await client.generate(engine_input)
         assert len(output["responses"]) == 1
         assert isinstance(output["responses"][0], str)
 
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_vlm_inference_generation.py b/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_vlm_inference_generation.py
index 04f2448c24..c432fa6dcd 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_vlm_inference_generation.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_vlm_inference_generation.py
@@ -203,7 +203,6 @@ async def test_sample_with_multimodal_image(module_scoped_ray_init_fixture):
 
         request_payload = {
             "json": {
-                "model": MODEL_QWEN3_VL,
                 "prompt": prompt,
                 "num_samples": 1,
                 "sampling_params": {"temperature": 0.0, "max_tokens": 20},
@@ -410,7 +409,7 @@ async def test_generate_with_multimodal_features_red_square(module_scoped_ray_in
             "sampling_params": {"max_tokens": 64, "temperature": 0.0},
             "mm_features": [features],
         }
-        gen_result = await client.generate(input_batch, model=client.model_name)
+        gen_result = await client.generate(input_batch)
 
         # Structural assertions
         assert len(gen_result["responses"]) == 1
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/megatron/test_router_replay.py b/tests/backends/skyrl_train/gpu/gpu_ci/megatron/test_router_replay.py
index f543ebbb55..70be711746 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/megatron/test_router_replay.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/megatron/test_router_replay.py
@@ -150,7 +150,6 @@ async def test_logprobs(ray_init_fixture, tp, pp, cp, ep, etp, extra_tf_kwargs):
                 skyrl_gym_cfg=cfg.environment.skyrl_gym,
                 inference_engine_client=client,
                 tokenizer=tokenizer,
-                policy_model_name=client.model_name,
             )
 
             input_batch: GeneratorInput = get_test_generator_input(
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/test_engine_generation.py b/tests/backends/skyrl_train/gpu/gpu_ci/test_engine_generation.py
index fba1dd6dad..82a65a198c 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/test_engine_generation.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/test_engine_generation.py
@@ -41,7 +41,7 @@ def get_test_actor_config(model: str = MODEL) -> SkyRLTrainConfig:
 
 async def run_batch_generation(client, prompts, sampling_params):
     engine_input = InferenceEngineInput(prompts=prompts, sampling_params=sampling_params)
-    engine_output = await client.generate(engine_input, model=client.model_name)
+    engine_output = await client.generate(engine_input)
     return engine_output["responses"], engine_output["stop_reasons"]
 
 
@@ -49,7 +49,7 @@ async def run_single_generation(client, prompts, sampling_params):
     tasks = []
     for prompt in prompts:
         engine_input = InferenceEngineInput(prompts=[prompt], sampling_params=sampling_params)
-        task = client.generate(engine_input, model=client.model_name)
+        task = client.generate(engine_input)
         tasks.append(task)
 
     results = await asyncio.gather(*tasks)
@@ -65,7 +65,7 @@ async def run_single_generation(client, prompts, sampling_params):
 
 async def run_batch_generation_with_tokens(client, prompt_token_ids, sampling_params):
     engine_input = InferenceEngineInput(prompt_token_ids=prompt_token_ids, sampling_params=sampling_params)
-    engine_output = await client.generate(engine_input, model=client.model_name)
+    engine_output = await client.generate(engine_input)
     return engine_output["responses"], engine_output["stop_reasons"]
 
 
@@ -73,7 +73,7 @@ async def run_single_generation_with_tokens(client, prompt_token_ids, sampling_p
     tasks = []
     for tokens in prompt_token_ids:
         engine_input = InferenceEngineInput(prompt_token_ids=[tokens], sampling_params=sampling_params)
-        task = client.generate(engine_input, model=client.model_name)
+        task = client.generate(engine_input)
         tasks.append(task)
 
     results = await asyncio.gather(*tasks)
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/test_pause_and_continue_generation.py b/tests/backends/skyrl_train/gpu/gpu_ci/test_pause_and_continue_generation.py
index 55ec3f77d7..d3084482ba 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/test_pause_and_continue_generation.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/test_pause_and_continue_generation.py
@@ -234,7 +234,7 @@ async def one_req(i: int):
                     "sampling_params": dict(sampling_params),
                     "session_ids": [i],
                 }
-                return await client.generate(engine_input, model=client.model_name)
+                return await client.generate(engine_input)
 
             tasks = [asyncio.create_task(one_req(i)) for i in range(num_requests)]
             # Let requests start and enqueue; with max_num_seqs=2, 2 run and 1 wait per engine
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_gym_generator.py b/tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_gym_generator.py
index c66461b1b8..8a3a258a5e 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_gym_generator.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_gym_generator.py
@@ -171,7 +171,6 @@ async def run_generator_end_to_end(
             skyrl_gym_cfg=env_cfg,
             inference_engine_client=inference_engine_client,
             tokenizer=tokenizer,
-            policy_model_name=inference_engine_client.model_name,
         )
 
         input_batch: GeneratorInput = get_test_generator_input(
diff --git a/tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_vlm_gym_generator.py b/tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_vlm_gym_generator.py
index 03bfe9a3fe..a671e8d0f8 100644
--- a/tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_vlm_gym_generator.py
+++ b/tests/backends/skyrl_train/gpu/gpu_ci/test_skyrl_vlm_gym_generator.py
@@ -191,7 +191,6 @@ async def test_vlm_generator_color_classification(ray_init_fixture):
             skyrl_gym_cfg=env_cfg,
             inference_engine_client=inference_client,
             tokenizer=tokenizer,
-            policy_model_name=inference_client.model_name,
         )
 
         num_prompts = 2
diff --git a/tests/train/generators/test_skyrl_gym_generator.py b/tests/train/generators/test_skyrl_gym_generator.py
index 21da71dbbd..d8bee68a1a 100644
--- a/tests/train/generators/test_skyrl_gym_generator.py
+++ b/tests/train/generators/test_skyrl_gym_generator.py
@@ -283,7 +283,6 @@ def mock_generate(_, model=None):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
-        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []  # to make sure observation_ids are encoded correctly
 
@@ -335,7 +334,6 @@ async def test_generate_batched(mock_make, mock_tokenizer, mock_llm, mock_env, g
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
-        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []  # to make sure observation_ids are encoded correctly
 
@@ -378,7 +376,6 @@ async def test_generate_interface_compliance(
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
-        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []  # to make sure observation_ids are encoded correctly
 
@@ -501,7 +498,6 @@ def mock_generate(input_batch, model=None):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
-        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []  # to make sure observation_ids are encoded correctly
 
@@ -591,7 +587,6 @@ def mock_encode(text, **kwargs):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
-        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []  # to make sure observation_ids are encoded correctly
 
@@ -673,7 +668,6 @@ def mock_encode(text, **kwargs):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
-        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []  # to make sure observation_ids are encoded correctly
 
@@ -737,7 +731,6 @@ def mock_apply_chat_template(messages, **kwargs):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
-        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []  # to make sure observation_ids are encoded correctly
 
@@ -859,7 +852,6 @@ def mock_encode_or_tokenize(text):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
-        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []  # to make sure observation_ids are encoded correctly
 
@@ -959,7 +951,6 @@ def step(self, action):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
-        policy_model_name="mock-model",
     )
 
     # Run agent loop
@@ -1050,7 +1041,6 @@ def step(self, action):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
-        policy_model_name="mock-model",
     )
     # Ensure base_conversation_token_ids doesn't shift observation slicing in test
     generator.base_conversation_token_ids = []
@@ -1141,7 +1131,6 @@ def step(self, action):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
-        policy_model_name="mock-model",
     )
     # Force retokenize path regardless of model resolution logic if needed
     generator.custom_chat_template = "<custom>"
@@ -1230,7 +1219,6 @@ def mock_make_func(*args, **kwargs):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
-        policy_model_name="mock-model",
     )
 
     prompt = [{"role": "user", "content": "Q?"}]
@@ -1322,7 +1310,6 @@ def mock_make_func(*args, **kwargs):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
-        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []
 
@@ -1440,7 +1427,6 @@ def step(self, action):
         skyrl_gym_cfg=mock_env_cfg,
         inference_engine_client=mock_llm,
         tokenizer=mock_tokenizer,
-        policy_model_name="mock-model",
     )
     generator.base_conversation_token_ids = []
 
diff --git a/tests/train/generators/test_skyrl_gym_generator_chat_templating.py b/tests/train/generators/test_skyrl_gym_generator_chat_templating.py
index 5e7f48a1d7..2d06bb3983 100644
--- a/tests/train/generators/test_skyrl_gym_generator_chat_templating.py
+++ b/tests/train/generators/test_skyrl_gym_generator_chat_templating.py
@@ -101,7 +101,6 @@ def _build_generator(tokenizer, model_name: str, chat_template_config, extra_ove
         skyrl_gym_cfg=env_cfg,
         inference_engine_client=None,  # to be replaced per-test
         tokenizer=tokenizer,
-        policy_model_name="mock-model",
     )
 
 
diff --git a/tests/train/generators/test_skyrl_vlm_generator.py b/tests/train/generators/test_skyrl_vlm_generator.py
index 4b505f8595..675c212a99 100644
--- a/tests/train/generators/test_skyrl_vlm_generator.py
+++ b/tests/train/generators/test_skyrl_vlm_generator.py
@@ -90,7 +90,6 @@ def _build_vlm_generator(tokenizer):
         skyrl_gym_cfg=env_cfg,
         inference_engine_client=mock_client,
         tokenizer=tokenizer,
-        policy_model_name="mock-model",
     )
     return generator