Deprecated vLLM parameters in V1 engine

[giladfrid009]

Currently in your EngineArgs class there are several vllm engine arguments that have been deprecated in V1 engine, and therefore cause crashes since they don't exist anymore.

class EngineArgs(TypedDict, total=False):
    model: str
    served_model_name: str | list[str] | None
    tokenizer: str | None
    # task: TaskOption
    skip_tokenizer_init: bool
    tokenizer_mode: str
...

For example:

The result of running your temporal clue example with the following config:

model = art.TrainableModel(
    name="066", project="temporal-clue", base_model="Qwen/Qwen2.5-7B-Instruct",
    _internal_config=InternalModelConfig(engine_args=EngineArgs(num_scheduler_steps=1)),
)

The num_scheduler_steps parameter does not exist anymore.

File ~/source/AgentDaC/.venv/lib/python3.11/site-packages/art/unsloth/service.py:1020, in llm()
   1018 for key in ["enable_log_requests", "disable_log_requests"]:
   1019     engine_args.pop(key, None)
-> [1020](https://vscode-remote+tunnel-002bgalileo1.vscode-resource.vscode-cdn.net/home/fre.gilad/source/AgentDaC/notebooks/examples/temporal_clue/~/source/AgentDaC/.venv/lib/python3.11/site-packages/art/unsloth/service.py:1020) return asyncio.create_task(get_llm(AsyncEngineArgs(**engine_args)))

TypeError: AsyncEngineArgs.__init__() got an unexpected keyword argument 'num_scheduler_steps'