[diffusion]: Add model detectors and warning for quantized diffusion models (#18041)

python/sglang/multimodal_gen/registry.py

@@ -390,6 +390,17 @@ def get_model_info(
     # 1. Discover all available pipeline classes and cache them
     _discover_and_register_pipelines()
 
+    # Detect quantized models and fallback to diffusers
+    is_quantized = any(q in model_path.lower() for q in ["-4bit", "-awq", "-gptq"])
+    if is_quantized and backend != Backend.DIFFUSERS:
+        logger.info(
+            "Detected a quantized model format ('%s'). "
+            "The native sglang-diffusion engine currently only supports BF16/FP16. "
+            "Falling back to diffusers backend.",
+            model_path,
+        )
+        return _get_diffusers_model_info(model_path)
+
     # 2. Get pipeline class - check non-diffusers models first
     pipeline_class_name = get_non_diffusers_pipeline_name(model_path)
     if pipeline_class_name:
@@ -656,34 +667,49 @@ def _register_configs():
         sampling_param_cls=QwenImageSamplingParams,
         pipeline_config_cls=QwenImagePipelineConfig,
         hf_model_paths=["Qwen/Qwen-Image"],
+        model_detectors=[
+            lambda hf_id: "qwen-image" in hf_id.lower()
+            and "edit" not in hf_id.lower()
+            and "layered" not in hf_id.lower()
+            and "2512" not in hf_id.lower()
+        ],
     )
     register_configs(
         sampling_param_cls=QwenImage2512SamplingParams,
         pipeline_config_cls=QwenImagePipelineConfig,
         hf_model_paths=["Qwen/Qwen-Image-2512"],
+        model_detectors=[lambda hf_id: "qwen-image-2512" in hf_id.lower()],
     )
     register_configs(
         sampling_param_cls=QwenImageSamplingParams,
         pipeline_config_cls=QwenImageEditPipelineConfig,
         hf_model_paths=["Qwen/Qwen-Image-Edit"],
+        model_detectors=[
+            lambda hf_id: "qwen-image-edit" in hf_id.lower()
+            and "2509" not in hf_id.lower()
+            and "2511" not in hf_id.lower()
+        ],
     )
 
     register_configs(
         sampling_param_cls=QwenImageEditPlusSamplingParams,
         pipeline_config_cls=QwenImageEditPlusPipelineConfig,
         hf_model_paths=["Qwen/Qwen-Image-Edit-2509"],
+        model_detectors=[lambda hf_id: "qwen-image-edit-2509" in hf_id.lower()],
     )
 
     register_configs(
         sampling_param_cls=QwenImageEditPlusSamplingParams,
         pipeline_config_cls=QwenImageEditPlus_2511_PipelineConfig,
         hf_model_paths=["Qwen/Qwen-Image-Edit-2511"],
+        model_detectors=[lambda hf_id: "qwen-image-edit-2511" in hf_id.lower()],
     )
 
     register_configs(
         sampling_param_cls=QwenImageLayeredSamplingParams,
         pipeline_config_cls=QwenImageLayeredPipelineConfig,
         hf_model_paths=["Qwen/Qwen-Image-Layered"],
+        model_detectors=[lambda hf_id: "qwen-image-layered" in hf_id.lower()],
     )
 
     register_configs(