[None][fix] Switch AD AllReduce strategy to NCCL (#8979)

MrGeva · web-flow · commit 990e674b71bc · 2025-11-07T06:49:44.000+02:00
Signed-off-by: Eran Geva &lt;19514940+MrGeva@users.noreply.github.com&gt;
diff --git a/tensorrt_llm/_torch/auto_deploy/distributed/trtllm.py b/tensorrt_llm/_torch/auto_deploy/distributed/trtllm.py
@@ -28,7 +28,7 @@ def trtllm_allreduce(tensor, op, all_reduce_params=None):
             p_config = Mapping(world_size=world_size, tp_size=world_size, rank=rank)
             # Use Strategy.AUTO for optimal performance
             _allreduce_cache[cache_key] = AllReduce(
-                mapping=p_config, strategy=AllReduceStrategy.AUTO, dtype=tensor.dtype
+                mapping=p_config, strategy=AllReduceStrategy.NCCL, dtype=tensor.dtype
             )
 
         torch_op = _allreduce_cache[cache_key]

Original file line number	Diff line number	Diff line change
`@@ -28,7 +28,7 @@ def trtllm_allreduce(tensor, op, all_reduce_params=None):`
`28`	`28`	`p_config = Mapping(world_size=world_size, tp_size=world_size, rank=rank)`
`29`	`29`	`# Use Strategy.AUTO for optimal performance`
`30`	`30`	`_allreduce_cache[cache_key] = AllReduce(`
`31`		`- mapping=p_config, strategy=AllReduceStrategy.AUTO, dtype=tensor.dtype`
	`31`	`+ mapping=p_config, strategy=AllReduceStrategy.NCCL, dtype=tensor.dtype`
`32`	`32`	`)`
`33`	`33`
`34`	`34`	`torch_op = _allreduce_cache[cache_key]`