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

12
vllm/model_executor/layers/quantization/fp8.py
View File

@@ -1292,13 +1292,11 @@ class Fp8MoEMethod(FusedMoEMethodBase):
apply_router_weight_on_input=layer.apply_router_weight_on_input,
)
select_result = layer.select_experts(
topk_weights, topk_ids = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
topk_weights, topk_ids, zero_expert_result = select_result
if self.rocm_aiter_moe_enabled:
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import ( # noqa: E501
rocm_aiter_fused_experts,
@@ -1353,13 +1351,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
apply_router_weight_on_input=layer.apply_router_weight_on_input,
)
if layer.zero_expert_num != 0 and layer.zero_expert_type is not None:
assert not isinstance(result, tuple), (
"Shared + zero experts are mutually exclusive not yet supported"
)
return result, zero_expert_result
else:
return result
return result
class Fp8OnlineMoEMethod(Fp8MoEMethod):