xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 07:15:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix GLM-4 PP Missing Layer When using with PP. (#21531)

Browse Source 
Signed-off-by: zRzRzRzRzRzRzR <2448370773@qq.com>
This commit is contained in:
Yuxuan Zhang 2025-07-25 11:07:38 +08:00 committed by GitHub
parent 633f6e804b
commit 2ce90e5b01
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 9 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/model_executor/models/glm4_moe.py
View File

@ -612,14 +612,20 @@ class Glm4MoeForCausalLM(nn.Module, SupportsPP):
self.num_expert_groups = config.n_group
self.moe_layers: list[FusedMoE] = []
example_moe = None
for layer in self.model.layers:
if isinstance(layer, PPMissingLayer):
continue
assert isinstance(layer, Glm4MoeDecoderLayer)
if isinstance(layer.mlp, Glm4MoE):
# Pick last one layer since the first ones may be dense layers.
example_moe = layer.mlp
self.moe_layers.append(layer.mlp.experts)
# Pick last one layer since the first ones may be dense layers.
example_moe = typing.cast(
Glm4MoE, self.model.layers[config.num_hidden_layers - 1].mlp)
if example_moe is None:
raise RuntimeError("No Glm4MoE layer found in model.layers.")
self.num_logical_experts = example_moe.n_logical_experts
self.num_physical_experts = example_moe.n_physical_experts
self.num_local_physical_experts = example_moe.n_local_physical_experts