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

[Bugfix][MoE] Fix 6-8% decode regression: prefer multi-stream shared expert overlap (#38990)

Browse Source 
Signed-off-by: Martin Vit <martin@voipmonitor.org>
Signed-off-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
This commit is contained in:
Martin Vit
2026-04-05 16:28:31 +02:00
committed by GitHub
parent 9a528260ef
commit 228023b3a5
1 changed files with 9 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
vllm/model_executor/layers/fused_moe/runner/shared_experts.py
View File

@@ -93,24 +93,24 @@ class SharedExperts:
)
@property
def _has_external_experts(self) -> bool:
def _use_external_experts(self) -> bool:
if self._use_dp_chunking:
return False
# Disable shared expert overlap if:
# - we are using eplb with non-default backend, because of correctness issues
# - we are using flashinfer with DP, since there nothing to gain
backend = self._moe_config.moe_parallel_config.all2all_backend
return not (
(
self._moe_config.moe_parallel_config.enable_eplb
and backend != "allgather_reducescatter"
)
or self._moe_config.moe_parallel_config.use_fi_nvl_two_sided_kernels
)
return (
self._moe_config.moe_parallel_config.enable_eplb
and backend != "allgather_reducescatter"
) or self._moe_config.moe_parallel_config.use_fi_nvl_two_sided_kernels
def _determine_shared_experts_order(
self,
hidden_states: torch.Tensor,
) -> SharedExpertsOrder:
if self._has_external_experts and not self._use_dp_chunking:
if self._use_external_experts:
return SharedExpertsOrder.EXTERNAL
if self._quant_method.mk_owns_shared_expert: