[Serve LLM] Fix enable_log_requests=False not being forwarded to vLLM AsyncLLM

[kouroshHakha]

Summary

Forward enable_log_requests from engine args to the AsyncLLM constructor so that enable_log_requests=False actually suppresses per-request log messages.

Pre-requisit:

• [Serve] Fix enable_access_log=False not suppressing access logs on stderr #60822

Reproduction

from ray.serve.llm import LLMConfig

config = LLMConfig(
    model_loading_config={"model_id": "my-model", "model_source": "org/model"},
    engine_kwargs={
        "tensor_parallel_size": 4,
        "enable_log_requests": False,
    },
)

Expected: No Added request cmpl-... lines in stderr.

Actual:

INFO ... [async_llm.py:428] Added request cmpl-bench-174196d2-1487-0-afeeac71.

Root Cause

In _start_async_llm_engine() (vllm_engine.py), log_requests is never passed to AsyncLLM, so it defaults to True:

# Before (bug)
engine_client = AsyncLLM(
    vllm_config=vllm_engine_config,
    executor_class=executor_class,
    log_stats=not vllm_engine_args.disable_log_stats,
    stat_loggers=custom_stat_loggers,
)

Fix

# After
engine_client = AsyncLLM(
    vllm_config=vllm_engine_config,
    executor_class=executor_class,
    log_requests=vllm_engine_args.enable_log_requests,
    log_stats=not vllm_engine_args.disable_log_stats,
    stat_loggers=custom_stat_loggers,
)

Also removes the deprecated disable_log_requests compat shim in vllm_models.py since vLLM has fully dropped that flag.

[gemini-code-assist]

Code Review

This pull request introduces two important logging-related fixes. The primary change correctly forwards the enable_log_requests argument to the vLLM engine, allowing users to suppress per-request logs as intended. This also includes a cleanup to remove a deprecated compatibility shim for disable_log_requests. Additionally, a secondary fix ensures that setting enable_access_log=False in Ray Serve's logging configuration correctly suppresses access logs in stderr, not just in log files. A new regression test has been added to verify this behavior. The changes are well-implemented and improve the logging control and consistency. The code looks good.

[eicherseiji]

Is this meant for this PR?