biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[All Reduce] Change default backend of Flashinfer All Reduce to trtllm (#35793)

Browse Source 
Signed-off-by: hjjq <hanjieq@nvidia.com>
This commit is contained in:
Hanjie Qiu
2026-03-02 18:57:38 -05:00
committed by GitHub
parent 1b82b433fc
commit 96fc09503a
1 changed files with 5 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/envs.py
View File

@@ -168,7 +168,7 @@ if TYPE_CHECKING:
VLLM_FLASHINFER_MOE_BACKEND: Literal["throughput", "latency", "masked_gemm"] = ( VLLM_FLASHINFER_MOE_BACKEND: Literal["throughput", "latency", "masked_gemm"] = (
"latency" "latency"
) )
VLLM_FLASHINFER_ALLREDUCE_BACKEND: Literal["auto", "trtllm", "mnnvl"] = "auto" VLLM_FLASHINFER_ALLREDUCE_BACKEND: Literal["auto", "trtllm", "mnnvl"] = "trtllm"
VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE: int = 394 * 1024 * 1024 VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE: int = 394 * 1024 * 1024
VLLM_XGRAMMAR_CACHE_MB: int = 0 VLLM_XGRAMMAR_CACHE_MB: int = 0
VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256 VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
@@ -1297,9 +1297,12 @@ environment_variables: dict[str, Callable[[], Any]] = {
# Flashinfer fused allreduce backend. # Flashinfer fused allreduce backend.
# "auto" will default to "mnnvl", which performs mostly same/better than "trtllm". # "auto" will default to "mnnvl", which performs mostly same/better than "trtllm".
# But "mnnvl" backend does not support fuse with quantization. # But "mnnvl" backend does not support fuse with quantization.
# TODO: Default is "trtllm" right now because "mnnvl" has issues with cudagraph:
# https://github.com/vllm-project/vllm/issues/35772
# Should switch back to "auto" if the issue is resolved.
"VLLM_FLASHINFER_ALLREDUCE_BACKEND": env_with_choices( "VLLM_FLASHINFER_ALLREDUCE_BACKEND": env_with_choices(
"VLLM_FLASHINFER_ALLREDUCE_BACKEND", "VLLM_FLASHINFER_ALLREDUCE_BACKEND",
"auto", "trtllm",
["auto", "trtllm", "mnnvl"], ["auto", "trtllm", "mnnvl"],
), ),
# Control the workspace buffer size for the FlashInfer backend. # Control the workspace buffer size for the FlashInfer backend.