[Bug]: The os.sched_yield() function may cause severe CPU host bound issues in certain scenarios.

[shaopeng-666]

Your current environment

The output of `python collect_env.py`

Your output of above commands here

🐛 Describe the bug

In certain NPU scenarios, os.sched_yield() can cause severe CPU host-bound performance issues (e.g., 800IA2, Qwen2.5-VL-72B W8A8 TP4 scenarios). Currently, you can avoid this problem by making the following manual modifications:
In the vllm/distributed/utils.py file within the vllm project, replace the implementation of the def sched_yield() function:

# Original implementation:
def sched_yield():
    if USE_SCHED_YIELD:
        os.sched_yield()
    else:
        time.sleep(0)
# Modified implementation:
def sched_yield():
    time.sleep(0)