infrastructure-distributed-synchron-in-distributed-inference-using-ray-if-a-worker-rai-92be13e4

THEN To avoid deadlock, use a separate thread to monitor Ray workers: one thread waits for worker exceptions via `ray.wait` while the main thread attempts the process group initialization. If any worker fails, abort the initialization and handle the error. Alternatively, ensure all prerequisites (e.g., GPU availability) are checked before starting workers.

import ray, threading

def init_with_monitor(handles, init_fn):
    failed = None
    def monitor():
        nonlocal failed
        try:
            ray.wait(handles, num_returns=len(handles), timeout=0)
        except Exception as e:
            failed = e
    t = threading.Thread(target=monitor)
    t.start()
    init_fn()
    t.join()
    if failed:
        raise failed