xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-01-11 03:54:28 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Feat] Enable eplb with default all2all backend (#30559)

Browse Source 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye 2025-12-16 10:33:52 -05:00 committed by GitHub
parent 00a8d7628c
commit 59bd5f6a71
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 4 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/model_executor/layers/fused_moe/shared_fused_moe.py
View File

@ -29,14 +29,14 @@ class SharedFusedMoE(FusedMoE):
self._shared_experts = shared_experts
# Disable shared expert overlap if:
# - we are using eplb, because of correctness issues
# - we are using flashinfer with DP, since there nothing to gain
# - we are using eplb with non-default backend, because of correctness issues
# - we are using flashinfer with DP, since there nothint to gain
# - we are using marlin kernels
backend = self.moe_parallel_config.all2all_backend
self.use_overlapped = (
use_overlapped
and not (
# TODO(wentao): find the root cause and remove this condition
self.enable_eplb
(self.enable_eplb and backend != "allgather_reducescatter")
or (self.moe_config.use_flashinfer_cutlass_kernels and self.dp_size > 1)
)
and self._shared_experts is not None