[misc] tune some env vars for GB200 (vllm-project#16992)

Signed-off-by: youkaichao <[email protected]>

Author: youkaichao

vllm/env_override.py

@@ -8,8 +8,21 @@
 # that interact with vllm workers.
 # they are executed whenever `import vllm` is called.
 
-# see https://github.com/NVIDIA/nccl/issues/1234
-os.environ['NCCL_CUMEM_ENABLE'] = '0'
+if not os.path.exists('/dev/nvidia-caps-imex-channels'):
+    # normally, we disable NCCL_CUMEM_ENABLE because it
+    # will cost 1~2 GiB GPU memory with cudagraph+allreduce,
+    # see https://github.com/NVIDIA/nccl/issues/1234
+    # for more details.
+    # However, NCCL requires NCCL_CUMEM_ENABLE to work with
+    # multi-node NVLink, typically on GB200-NVL72 systems.
+    # The ultimate way to detect multi-node NVLink is to use
+    # NVML APIs, which are too expensive to call here.
+    # As an approximation, we check the existence of
+    # /dev/nvidia-caps-imex-channels, used by
+    # multi-node NVLink to communicate across nodes.
+    # This will still cost some GPU memory, but it is worthwhile
+    # because we can get very fast cross-node bandwidth with NVLink.
+    os.environ['NCCL_CUMEM_ENABLE'] = '0'
 
 # see https://github.com/vllm-project/vllm/pull/15951
 # it avoids unintentional cuda initialization from torch.cuda.is_available()