minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

fix(qwen3.5): restrict MoE weights to local PP layers (#32022)

Browse Source 
Co-authored-by: chunjiang.hc <chunjiang.hc@alibaba-inc.com>
Co-authored-by: Yihao Wang <42559837+AgainstEntropy@users.noreply.github.com>
This commit is contained in:
ICENacl
2026-07-29 12:01:09 -07:00
committed by GitHub
co-authored by chunjiang.hc Yihao Wang
parent f73d6f2789
commit e4f7f7b380
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/qwen3_5.py
+3 -3
View File
@@ -2288,9 +2288,9 @@ class Qwen3_5MoeForConditionalGeneration(Qwen3VLForConditionalGeneration):
self._routed_experts_weights_of_layer = LazyValue( self._routed_experts_weights_of_layer = LazyValue(
lambda: { lambda: {
layer_id: layer.mlp.get_moe_weights() layer_id: self.model.layers[layer_id].mlp.get_moe_weights()
for layer_id, layer in enumerate(self.model.layers) for layer_id in range(self.model.start_layer, self.model.end_layer)
if isinstance(layer.mlp, Qwen2MoeSparseMoeBlock) if isinstance(self.model.layers[layer_id].mlp, Qwen2MoeSparseMoeBlock)
} }
) )