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

[MoE Refactor][5/N] Isolate zero expert to LongCatFlash (#28891)

Browse Source 
Signed-off-by: baonudesifeizhai <85092850+baonudesifeizhai@users.noreply.github.com>
Signed-off-by: Dongjie Zou <85092850+baonudesifeizhai@users.noreply.github.com>
Signed-off-by: baonudesifeizhai <baonudesifeizhai@gmail.com>
Signed-off-by: Robert Shaw <robertgshaw2@gmail.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Robert Shaw <robertgshaw2@gmail.com>
This commit is contained in:
baonudesifeizhai
2025-12-20 13:22:04 -05:00
committed by GitHub
parent 560ae9638c
commit 54c8924384
19 changed files with 264 additions and 109 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
vllm/model_executor/layers/quantization/gptq_marlin.py
View File

@@ -900,7 +900,7 @@ class GPTQMarlinMoEMethod(FusedMoEMethodBase):
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
assert layer.activation == "silu", "Only SiLU activation is supported."
topk_weights, topk_ids, _ = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)