解决custom allreduce在dp情况下的其服务错误问题

912dc4c9 · xiabo · 16732666 · 912dc4c9
Commit 912dc4c9 authored Jan 07, 2026 by xiabo
Hide whitespace changes
Inline Side-by-side

Showing with 5 additions and 1 deletion

vllm/distributed/device_communicators/custom_all_reduce.py vllm/distributed/device_communicators/custom_all_reduce.py +5 -1

No files found.
--- a/vllm/distributed/device_communicators/custom_all_reduce.py
+++ b/vllm/distributed/device_communicators/custom_all_reduce.py
@@ -121,7 +121,11 @@ class CustomAllreduce:
        else:
            device_ids = list(range(cuda_device_count_stateless()))

-        physical_device_id = device_ids[device.index]
+        if (world_size == len(device_ids)):
+            physical_device_id = device_ids[device.index % world_size]
+        else:
+            physical_device_id = device_ids[device.index]
+            
        tensor = torch.tensor([physical_device_id],
                              dtype=torch.int,
                              device="cpu")