[WideEP] Change DeepEPv2 to auto select hybrid mode by default (#57991)

Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Codex <noreply@openai.com>

vllm/distributed/device_communicators/all2all.py

@@ -1065,7 +1065,9 @@ class DeepEPV2All2AllManager(All2AllManagerBase):
         if os.environ.get("EP_DISABLE_GIN", "0") != "0":
             return
 
-        gin_type = query_nccl_gin_type(group)
+        gin_type = query_nccl_gin_type(
+            group, railed=envs.VLLM_DEEPEP_V2_ALLOW_HYBRID_MODE
+        )
         if gin_type is None:
             raise RuntimeError(
                 "DeepEPv2 communicator properties query failed; "

vllm/envs.py

@@ -2009,9 +2009,13 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL": lambda: bool(
         int(os.getenv("VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL", "0"))
     ),
-    # DeepEP v2: enable two-tier NVLink+RDMA hybrid mode
+    # DeepEP v2: use NVLink+RDMA hybrid communication across NVLink domains.
+    #   * hybrid mode = multi-node RDMA
+    #   * non hybrid mode = nvlink
+    # Set to 1 (the default) to allow DeepEP to autoselect hybrid vs non-hybrid.
+    # Set to 0 to disable hybrid mode.
     "VLLM_DEEPEP_V2_ALLOW_HYBRID_MODE": lambda: bool(
-        int(os.getenv("VLLM_DEEPEP_V2_ALLOW_HYBRID_MODE", "0"))
+        int(os.getenv("VLLM_DEEPEP_V2_ALLOW_HYBRID_MODE", "1"))
     ),
     # DeepEP v2: use fewer SMs at slight throughput cost
     "VLLM_DEEPEP_V2_PREFER_OVERLAP": lambda: bool(

vllm/utils/nccl.py

@@ -83,8 +83,10 @@ def find_nccl_library_paths() -> list[str] | None:
     return paths or None
 
 
-def query_nccl_gin_type(group: torch.distributed.ProcessGroup) -> int | None:
-    """Return the GIN type for an initialized group, or ``None`` on failure."""
+def query_nccl_gin_type(
+    group: torch.distributed.ProcessGroup, *, railed: bool = False
+) -> int | None:
+    """Return the full or railed GIN type, or ``None`` on query failure."""
     from vllm.distributed.device_communicators.pynccl_wrapper import (
         NCCL_COMM_PROPERTIES_LAYOUT_VERSION,
         NCCLLibrary,
@@ -126,4 +128,4 @@ def query_nccl_gin_type(group: torch.distributed.ProcessGroup) -> int | None:
     if result != 0:
         logger.warning("ncclCommQueryProperties returned error %d", result)
         return None
-    return props.ginType
+    return props.railedGinType if railed else props.ginType