fix run error

vllm/config/model.py

@@ -912,6 +912,7 @@ class ModelConfig:
                 # imports during override detection (e.g., MXFP4 imports Triton)
                 "mxfp4",
                 "cpu_awq",
+                "slimquant_marlin",
                 "slimquant_w4a8_marlin",
                 "slimquant_compressed_tensors_marlin",
             ]

vllm/config/speculative.py

@@ -371,7 +371,7 @@ class SpeculativeConfig:
                     tokenizer_revision=self.target_model_config.tokenizer_revision,
                     spec_target_max_model_len=self.target_model_config.max_model_len,
                     quantization=self.quantization,
-                    enforce_eager=True if envs.VLLM_SPEC_DECODE_EAGER else self.target_model_config.enforce_eager,
+                    enforce_eager=self.target_model_config.enforce_eager,
                     max_logprobs=self.target_model_config.max_logprobs,
                     hf_overrides=SpeculativeConfig.hf_config_override,
                     config_format=self.target_model_config.config_format,

vllm/model_executor/models/deepseek_v2.py

@@ -263,9 +263,14 @@ class DeepseekV2MoE(nn.Module):
             prefix=f"{prefix}.gate",
         )
         if getattr(config, "topk_method", None) == "noaux_tc":
-            self.gate.e_score_correction_bias = nn.Parameter(
-                torch.empty(config.n_routed_experts, dtype=torch.float32)
-            )
+            if envs.VLLM_ENABLE_MOE_FUSED_GATE:
+                # avoid moe_fused_gate precision error
+                self.gate.e_score_correction_bias = nn.Parameter(
+                torch.empty(config.n_routed_experts))
+            else:
+                self.gate.e_score_correction_bias = nn.Parameter(
+                    torch.empty(config.n_routed_experts, dtype=torch.float32)
+                )
         else:
             self.gate.e_score_correction_bias = None