A high-throughput and memory-efficient inference and serving engine for LLMs
[Bugfix]: Fix Quark gpt-oss weight loading broken by FusedMoe refactor (#45067)
Signed-off-by: Rohan138 <rohanpotdar138@gmail.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com>
Rohan Potdar committed
Jun 10, 2026 at 20:17 UTC
7920ccb97c2d27d0a1a822e42d11ba369c2f255c
1 file changed
+7
vllm/model_executor/models/gpt_oss.py
+7
@@ -635,6 +635,13 @@ class GptOssModel(nn.Module, EagleModelMixin):
635
"an unexpected condition. Please open an issue if encountered."
636
)
637
638
+ # The MoE refactor (#41184) moved expert params under
639
+ # `mlp.experts.routed_experts.*`; remap the legacy checkpoint
640
+ # name so keys like w2_bias resolve against params_dict.
641
+ fused_name = fused_name.replace(
642
+ ".mlp.experts.", ".mlp.experts.routed_experts."
643
+ )
644
+
645
moe_quant_method = _get_moe_weight_dtype(layer_id=layer_id)
646
647
if (