[tinker][megatron] Multi-LoRA Megatron + Tinker API RL Training#1621

Open

erictang000 wants to merge 22 commits intoNovaSky-AI:mainfrom

erictang000:multi_lora_rl

Commits on Apr 28, 2026

Commits on May 4, 2026

[docs] Add Multi-LoRA Megatron Tinker design doc (v1)

erictang000
and
claude
committed
[multi-lora] Add AdapterStore for per-worker LoRA slot bookkeeping

erictang000
and
claude
committed
[multi-lora] Wire AdapterStore into MegatronPolicyWorkerBase

erictang000
and
claude
committed
[multi-lora] Add ensure_active_adapter + model_id threading to dispatch

erictang000
and
claude
committed
[multi-lora] Allow multiple LoRA policy adapters in SkyRLTrainBackend

erictang000
and
claude
committed
[multi-lora] Add GPU-gated multi-LoRA integration test for Megatron

erictang000
and
claude
committed
[multi-lora] Add two-client smoke runbook

erictang000
and
claude
committed
[multi-lora] Fix _lora_signature_from to not read non-existent target_modules

erictang000
and
claude
committed
x
erictang000
committed
[multi-lora] Swap grad buffers along with params + optimizer state

erictang000
and
claude
committed
[multi-lora-rl] Wire model_id through the LoRA sync + sampling path

erictang000
and
claude
committed
[multi-lora-rl] Tolerate non-JSON error bodies in load/unload_lora_adapter

erictang000
and
claude
committed
[multi-lora-rl] Update design doc + RL two-client smoke runbook

erictang000
and
claude
committed
Merge remote-tracking branch 'origin/main' into multi_lora_rl
erictang000
committed

Commits on May 6, 2026