[skyrl] Pass through profiler_config to vLLM engine

[pcmoritz]

Can be used like

"generator":{"inference_engine":{"tensor_parallel_size":4, "profiler_config":"{\"profiler\":\"torch\",\"torch_profiler_dir\":\"/mnt/shared_storage/vllm_profile\"}"}}}'

[gemini-code-assist]

Code Review

This pull request adds a profiler_config field to the InferenceEngineConfig class to pass profiling configurations to the vLLM engine. The review feedback points out that this addition is redundant because engine_init_kwargs already exists for this purpose. Additionally, the current implementation is likely to cause a TypeError at runtime because vLLM expects profiling parameters as top-level arguments rather than a nested dictionary, and there is a type mismatch between the defined Dict type hint and the JSON string usage.

[gemini-code-assist]

The addition of profiler_config appears redundant because engine_init_kwargs (line 487) is already designed to pass arbitrary arguments to the vLLM engine.

Furthermore, this implementation is likely to cause runtime issues:

1. vLLM Compatibility: vLLM expects profiling parameters (such as profiler and torch_profiler_dir) as top-level arguments. If profiler_config is passed as a dictionary key to vllm.LLM or vllm.AsyncEngineArgs (via **kwargs in the engine initialization), it will trigger a TypeError for an unexpected keyword argument.
2. Type Mismatch: The PR description shows this field being populated with a JSON string, but the type hint is Dict[str, Any]. This will lead to type errors during configuration parsing or when the code expects a dictionary.

Recommendation: Instead of adding a new field, use the existing engine_init_kwargs. If a dedicated field is strictly required for grouping, logic must be added to the engine initialization (e.g., in vllm_engine.py) to unpack these values into the top-level engine arguments.

[pcmoritz]

Replaced by #1622