[Misc] Fused MoE Marlin support for GPTQ (#8217)

vllm/model_executor/model_loader/utils.py

@@ -24,10 +24,18 @@ def get_model_architecture(
     # Special handling for quantized Mixtral.
     # FIXME(woosuk): This is a temporary hack.
     mixtral_supported = ["fp8", "compressed-tensors"]
+    # for gptq_marlin, only run fused MoE for int4
+    if model_config.quantization == "gptq_marlin":
+        hf_quant_config = getattr(model_config.hf_config,
+                                  "quantization_config", None)
+        if hf_quant_config and hf_quant_config.get("bits") == 4:
+            mixtral_supported.append("gptq_marlin")
+
     if (model_config.quantization is not None
             and model_config.quantization not in mixtral_supported
             and "MixtralForCausalLM" in architectures):
         architectures = ["QuantMixtralForCausalLM"]
+
     return ModelRegistry.resolve_model_cls(architectures)