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 13, 13:15 UTC
[Bugfix][EPLB] Disabled shared expert overlap when EPLB is enabled (#28377)
Signed-off-by: Sage Moore <sage@neuralmagic.com> Signed-off-by: Sage Moore <sagemoore@utexas.edu> Signed-off-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Sage Moore committed
Nov 10, 2025 at 12:39 UTC
40d33264c680a8c725b93db6ccce608f99e5c7da
1 file changed
+10
-5
vllm/model_executor/layers/fused_moe/shared_fused_moe.py
+10
-5
@@ -28,13 +28,18 @@ class SharedFusedMoE(FusedMoE):
28
super().__init__(**kwargs)
29
self._shared_experts = shared_experts
30
31
- # Disable shared expert overlap if we are not using
32
- # flashinfer + DP since there is nothing to be gained in this case.
33
- # Disabling the overlap optimization also prevents the shared experts
34
- # from being hidden from torch.compile.
31
+ # Disable shared expert overlap if we are using eplb, because of
32
+ # correctness issues, or if using flashinfer with DP, since there
33
+ # is nothing to be gained in this case. Disabling the overlap
34
+ # optimization also prevents the shared experts from being hidden
35
+ # from torch.compile.
36
self.use_overlapped = (
37
use_overlapped
37
- and not (self.use_flashinfer_cutlass_kernels and self.dp_size > 1)
38
+ and not (
39
+ # TODO(wentao): find the root cause and remove this condition
40
+ self.enable_eplb
41
+ or (self.use_flashinfer_cutlass_kernels and self.dp_size > 1)
42
+ )
43
and self._shared_experts is not None
44
)
45