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 14, 05:30 UTC
[Bug] Fix shape issue for eplb expert weights (#27589)
Signed-off-by: yewentao256 <zhyanwentao@126.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
Wentao Ye committed
Oct 28, 2025 at 08:44 UTC
0484b6424894d785fb70f3e39c47aaee489340e3
1 file changed
+2
vllm/model_executor/layers/fused_moe/layer.py
+2
@@ -1959,6 +1959,8 @@ class FusedMoE(CustomOp):
1959
if name not in NON_EXPERT_WEIGHTS
1960
and weight.shape != torch.Size([])
1961
and not name.startswith("_shared_experts.")
1962
+ # exclude parameters from non-expert submodules (e.g. gate/shared)
1963
+ and not name.startswith("_gate.")
1964
]
1965
1966
def set_eplb_state(