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 14, 22:35 UTC
[misc] skip p2p check by default (#21904)
youkaichao committed
Jul 30, 2025 at 22:05 UTC
e91d3c9cda69b9770241c79fbf94f81f5576e7f4
1 file changed
+7
-5
vllm/envs.py
+7
-5
@@ -668,12 +668,14 @@ environment_variables: dict[str, Callable[[], Any]] = {
668
(os.environ.get("VLLM_ALLOW_RUNTIME_LORA_UPDATING", "0").strip().lower() in
669
("1", "true")),
670
671
- # By default, vLLM will check the peer-to-peer capability itself,
672
- # in case of broken drivers. See https://github.com/vllm-project/vllm/blob/a9b15c606fea67a072416ea0ea115261a2756058/vllm/distributed/device_communicators/custom_all_reduce_utils.py#L101-L108 for details. # noqa
673
- # If this env var is set to 1, vLLM will skip the peer-to-peer check,
674
- # and trust the driver's peer-to-peer capability report.
671
+ # We assume drivers can report p2p status correctly.
672
+ # If the program hangs when using custom allreduce,
673
+ # potantially caused by a bug in the driver (535 series),
674
+ # if might be helpful to set VLLM_SKIP_P2P_CHECK=0
675
+ # so that vLLM can verify if p2p is actually working.
676
+ # See https://github.com/vllm-project/vllm/blob/a9b15c606fea67a072416ea0ea115261a2756058/vllm/distributed/device_communicators/custom_all_reduce_utils.py#L101-L108 for details. # noqa
677
"VLLM_SKIP_P2P_CHECK":
676
- lambda: os.getenv("VLLM_SKIP_P2P_CHECK", "0") == "1",
678
+ lambda: os.getenv("VLLM_SKIP_P2P_CHECK", "1") == "1",
679
680
# List of quantization kernels that should be disabled, used for testing
681
# and performance comparisons. Currently only affects MPLinearKernel