A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 11, 15:00 UTC
[CI/Build] Use spawn subprocess for ROCm (#30272)
Signed-off-by: Ryan Rock <ryan.rock@amd.com>
Ryan Rock committed
Dec 11, 2025 at 21:33 UTC
197473c4e71c99025a0fd3925d0f130bdbfa1e42
1 file changed
+6
examples/offline_inference/data_parallel.py
+6
@@ -33,6 +33,7 @@ import os
33
from time import sleep
34
35
from vllm import LLM, SamplingParams
36
+from vllm.platforms import current_platform
37
from vllm.utils.network_utils import get_open_port
38
39
@@ -222,6 +223,11 @@ if __name__ == "__main__":
223
224
from multiprocessing import Process
225
226
+ if current_platform.is_rocm():
227
+ from multiprocessing import set_start_method
228
+
229
+ set_start_method("spawn", force=True)
230
+
231
procs = []
232
for local_dp_rank, global_dp_rank in enumerate(
233
range(node_rank * dp_per_node, (node_rank + 1) * dp_per_node)