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 6, 13:15 UTC
[Bug] Fix FlashInfer MNNVL socket collisions under concurrent vLLM jobs (#36674)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Mar 17, 2026 at 15:19 UTC
bdb903bb5f4b943ad2a2d1c08f1f70d866e26496
1 file changed
+16
-9
vllm/distributed/device_communicators/flashinfer_all_reduce.py
+16
-9
@@ -3,6 +3,8 @@
3
4
5
import atexit
6
+import os
7
+import random
8
import threading
9
10
import torch
@@ -67,15 +69,20 @@ def initialize_fi_ar_workspace(
69
70
backend = envs.VLLM_FLASHINFER_ALLREDUCE_BACKEND
71
comm_backend = TorchDistBackend(group=group)
70
- _fi_ar_workspace = flashinfer_comm.create_allreduce_fusion_workspace(
71
- backend=backend,
72
- world_size=world_size,
73
- rank=rank,
74
- max_token_num=max_token_num,
75
- hidden_dim=hidden_dim,
76
- dtype=dtype,
77
- comm_backend=comm_backend,
78
- )
72
+ rng_state = random.getstate()
73
+ try:
74
+ random.seed(int.from_bytes(os.urandom(16), byteorder="big"))
75
+ _fi_ar_workspace = flashinfer_comm.create_allreduce_fusion_workspace(
76
+ backend=backend,
77
+ world_size=world_size,
78
+ rank=rank,
79
+ max_token_num=max_token_num,
80
+ hidden_dim=hidden_dim,
81
+ dtype=dtype,
82
+ comm_backend=comm_backend,
83
+ )
84
+ finally:
85
+ random.setstate(rng_state)
86
assert _fi_ar_workspace is not None
87
logger.debug(
88
"Initialized FlashInfer All Reduce workspace: backend=%s, "