diff --git a/vllm/model_executor/models/qwen3_5_mtp.py b/vllm/model_executor/models/qwen3_5_mtp.py index 0eca47492c..e49806365e 100644 --- a/vllm/model_executor/models/qwen3_5_mtp.py +++ b/vllm/model_executor/models/qwen3_5_mtp.py @@ -75,13 +75,22 @@ class Qwen3_5MultiTokenPredictor(nn.Module): config.hidden_size, ) + # Workaround: mtp.fc is stored as BF16 in NVFP4 checkpoints but is + # missing from hf_quant_config.json exclude_modules. Force unquantized. + # Ref: https://github.com/vllm-project/vllm/pull/38650 + # Ref: https://github.com/NVIDIA/Model-Optimizer/pull/1124 + fc_quant = ( + None + if (quant_config and quant_config.get_name() == "modelopt_fp4") + else quant_config + ) self.fc = ColumnParallelLinear( self.config.hidden_size * 2, self.config.hidden_size, gather_output=True, bias=False, return_bias=False, - quant_config=quant_config, + quant_config=fc_quant, prefix=f"{prefix}.fc", )