[Bugfix]: Fix unquantized gpt-oss weight loading broken by FusedMoE r… (#45818)
Signed-off-by: priyansh jain <priyansh.jain2@amd.com> Co-authored-by: Li, Jiang <jiang1.li@intel.com>
Priyansh Jain committed
Jun 23, 2026 at 22:26 UTC
f4d5f73ffa402569ee76e5a2ade05ce7f8b5a843
1 file changed
+5
-1
vllm/model_executor/models/gpt_oss.py
+5
-1
@@ -979,7 +979,11 @@ class GptOssModel(nn.Module, EagleModelMixin):
979
tp_rank_start = tp_rank * per_rank_intermediate_size
980
tp_rank_end = min((tp_rank + 1) * per_rank_intermediate_size, intermediate_size)
981
982
- for name, weight in weights:
982
+ # Use centralized weight remapping for MoE expert parameters.
983
+ # The FusedMoE refactor moved expert params under
984
+ # `mlp.experts.routed_experts.*`; this remaps checkpoint names so
985
+ # MoE weight/bias keys resolve against params_dict.
986
+ for name, weight in remap_moe_expert_weights(weights, params_dict):
987
# Skip layers on other devices.
988
if is_pp_missing_parameter(name, self):
989
continue