From c7742d61134783b50098ab249f6815051a4c4a2a Mon Sep 17 00:00:00 2001 From: Rui Qiao <161574667+ruisearch42@users.noreply.github.com> Date: Fri, 25 Jul 2025 17:08:30 -0700 Subject: [PATCH] [Bugfix] Always set RAY_ADDRESS for Ray actor before spawn (#21540) Signed-off-by: Rui Qiao --- vllm/utils/__init__.py | 19 ++++++++++--------- 1 file changed, 10 insertions(+), 9 deletions(-) diff --git a/vllm/utils/__init__.py b/vllm/utils/__init__.py index 9f4140ac64..054037b893 100644 --- a/vllm/utils/__init__.py +++ b/vllm/utils/__init__.py @@ -2883,26 +2883,27 @@ def _maybe_force_spawn(): if os.environ.get("VLLM_WORKER_MULTIPROC_METHOD") == "spawn": return - reason = None - if cuda_is_initialized(): - reason = "CUDA is initialized" - elif xpu_is_initialized(): - reason = "XPU is initialized" - elif is_in_ray_actor(): + reasons = [] + if is_in_ray_actor(): # even if we choose to spawn, we need to pass the ray address # to the subprocess so that it knows how to connect to the ray cluster. # env vars are inherited by subprocesses, even if we use spawn. import ray os.environ["RAY_ADDRESS"] = ray.get_runtime_context().gcs_address - reason = "In a Ray actor and can only be spawned" + reasons.append("In a Ray actor and can only be spawned") - if reason is not None: + if cuda_is_initialized(): + reasons.append("CUDA is initialized") + elif xpu_is_initialized(): + reasons.append("XPU is initialized") + + if reasons: logger.warning( "We must use the `spawn` multiprocessing start method. " "Overriding VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. " "See https://docs.vllm.ai/en/latest/usage/" "troubleshooting.html#python-multiprocessing " - "for more information. Reason: %s", reason) + "for more information. Reasons: %s", "; ".join(reasons)) os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"