[misc][distributed] add VLLM_SKIP_P2P_CHECK flag (#8911)

vllm/distributed/device_communicators/custom_all_reduce.py

@@ -28,6 +28,10 @@ def _can_p2p(rank: int, world_size: int) -> bool:
     for i in range(world_size):
         if i == rank:
             continue
+        if envs.VLLM_SKIP_P2P_CHECK:
+            logger.info(
+                "Skipping P2P check and trusting the driver's P2P report.")
+            return torch.cuda.can_device_access_peer(rank, i)
         if not gpu_p2p_access_check(rank, i):
             return False
     return True

vllm/envs.py

@@ -63,6 +63,7 @@ if TYPE_CHECKING:
     VLLM_USE_TRITON_AWQ: bool = False
     VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
     VLLM_ALLOW_DEPRECATED_BEAM_SEARCH: bool = False
+    VLLM_SKIP_P2P_CHECK: bool = False
 
 
 def get_default_cache_root():
@@ -423,6 +424,13 @@ environment_variables: Dict[str, Callable[[], Any]] = {
     lambda:
     (os.environ.get("VLLM_ALLOW_RUNTIME_LORA_UPDATING", "0").strip().lower() in
      ("1", "true")),
+
+    # By default, vLLM will check the peer-to-peer capability itself,
+    # 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
+    # If this env var is set to 1, vLLM will skip the peer-to-peer check,
+    # and trust the driver's peer-to-peer capability report.
+    "VLLM_SKIP_P2P_CHECK":
+    lambda: os.getenv("VLLM_SKIP_P2P_CHECK", "0") == "1",
 }
 
 # end-env-vars-definition