From 7920ccb97c2d27d0a1a822e42d11ba369c2f255c Mon Sep 17 00:00:00 2001 From: Rohan Potdar <66227218+Rohan138@users.noreply.github.com> Date: Wed, 10 Jun 2026 20:17:46 -0500 Subject: [PATCH] [Bugfix]: Fix Quark gpt-oss weight loading broken by FusedMoe refactor (#45067) Signed-off-by: Rohan138 Co-authored-by: Andreas Karatzas --- vllm/model_executor/models/gpt_oss.py | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/vllm/model_executor/models/gpt_oss.py b/vllm/model_executor/models/gpt_oss.py index 1e69b321cd..ddcaecb08b 100644 --- a/vllm/model_executor/models/gpt_oss.py +++ b/vllm/model_executor/models/gpt_oss.py @@ -635,6 +635,13 @@ class GptOssModel(nn.Module, EagleModelMixin): "an unexpected condition. Please open an issue if encountered." ) + # The MoE refactor (#41184) moved expert params under + # `mlp.experts.routed_experts.*`; remap the legacy checkpoint + # name so keys like w2_bias resolve against params_dict. + fused_name = fused_name.replace( + ".mlp.experts.", ".mlp.experts.routed_experts." + ) + moe_quant_method = _get_moe_weight_dtype(layer_id=layer_id) if (