[Distributed] Enable FlashInfer all-reduce by default (#52998)
Signed-off-by: Woosuk Kwon <woosuk@inferact.ai> Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
co-authored by
OpenAI Codex
parent
58e5ee0158
commit
6a962071bd
@@ -58,7 +58,10 @@ class CudaCommunicator(DeviceCommunicatorBase):
|
||||
|
||||
use_custom_allreduce = _ENABLE_CUSTOM_ALL_REDUCE
|
||||
use_torch_symm_mem = envs.VLLM_ALLREDUCE_USE_SYMM_MEM
|
||||
use_flashinfer_allreduce = envs.VLLM_ALLREDUCE_USE_FLASHINFER
|
||||
# FlashInfer all-reduce does not provide a fixed reduction order.
|
||||
use_flashinfer_allreduce = (
|
||||
envs.VLLM_ALLREDUCE_USE_FLASHINFER and not envs.VLLM_BATCH_INVARIANT
|
||||
)
|
||||
use_aiter_allreduce = use_custom_allreduce and bool(
|
||||
rocm_aiter_ops.is_custom_all_reduce_enabled()
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user