From 55c21c88363811feb2aeb5aecac3cd48683e4705 Mon Sep 17 00:00:00 2001
From: Micah Williamson <micah.williamson@amd.com>
Date: Sat, 22 Nov 2025 23:05:00 -0600
Subject: [PATCH] [ROCm][CI] Fix "Cannot re-initialize CUDA in forked
 subprocess" in test_pynccl.py  (#29119)

Signed-off-by: Micah Williamson <micah.williamson@amd.com>
---
 requirements/rocm-test.txt       | 3 +++
 tests/distributed/test_pynccl.py | 8 +++++---
 2 files changed, 8 insertions(+), 3 deletions(-)

diff --git a/requirements/rocm-test.txt b/requirements/rocm-test.txt
index eabb5065bfce..2d57e7e16786 100644
--- a/requirements/rocm-test.txt
+++ b/requirements/rocm-test.txt
@@ -40,5 +40,8 @@ mteb[bm25s]>=1.38.11, <2
 # Required for eval tests
 lm-eval[api] @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d
 
+# Required for multiprocessed tests that use spawn method
+multiprocess==0.70.16
+
 # Plugins test
 terratorch @ git+https://github.com/IBM/terratorch.git@07184fcf91a1324f831ff521dd238d97fe350e3e
diff --git a/tests/distributed/test_pynccl.py b/tests/distributed/test_pynccl.py
index c3085beeb356..c7c9d0602def 100644
--- a/tests/distributed/test_pynccl.py
+++ b/tests/distributed/test_pynccl.py
@@ -1,9 +1,9 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
-import multiprocessing
 import os
 
+import multiprocess as mp
 import numpy as np
 import pytest
 import torch
@@ -20,10 +20,12 @@ from vllm.distributed.parallel_state import (
 )
 from vllm.utils.system_utils import update_environment_variables
 
+mp.set_start_method("spawn", force=True)
+
 
 def distributed_run(fn, world_size):
     number_of_processes = world_size
-    processes: list[multiprocessing.Process] = []
+    processes: list[mp.Process] = []
     for i in range(number_of_processes):
         env: dict[str, str] = {}
         env["RANK"] = str(i)
@@ -32,7 +34,7 @@ def distributed_run(fn, world_size):
         env["LOCAL_WORLD_SIZE"] = str(number_of_processes)
         env["MASTER_ADDR"] = "localhost"
         env["MASTER_PORT"] = "12345"
-        p = multiprocessing.Process(target=fn, args=(env,))
+        p = mp.Process(target=fn, args=(env,))
         processes.append(p)
         p.start()