A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 11, 08:45 UTC
[Feat] Enable eplb with default all2all backend (#30559)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Dec 16, 2025 at 10:33 UTC
59bd5f6a718a309517343e126f5086e057227992
1 file changed
+4
-4
vllm/model_executor/layers/fused_moe/shared_fused_moe.py
+4
-4
@@ -29,14 +29,14 @@ class SharedFusedMoE(FusedMoE):
29
self._shared_experts = shared_experts
30
31
# Disable shared expert overlap if:
32
- # - we are using eplb, because of correctness issues
33
- # - we are using flashinfer with DP, since there nothing to gain
32
+ # - we are using eplb with non-default backend, because of correctness issues
33
+ # - we are using flashinfer with DP, since there nothint to gain
34
# - we are using marlin kernels
35
+ backend = self.moe_parallel_config.all2all_backend
36
self.use_overlapped = (
37
use_overlapped
38
and not (
38
- # TODO(wentao): find the root cause and remove this condition
39
- self.enable_eplb
39
+ (self.enable_eplb and backend != "allgather_reducescatter")
40
or (self.moe_config.use_flashinfer_cutlass_kernels and self.dp_size > 1)
41
)
42
and self._shared_experts is not None