[diffusion] fix: fix diffusers backend issues in diffusion ci gt workflow (#20173)

.github/workflows/diffusion-ci-gt-gen.yml

@@ -29,7 +29,7 @@ jobs:
     strategy:
       matrix:
         part: [0, 1]
-    timeout-minutes: 60
+    timeout-minutes: 150
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -63,7 +63,7 @@ jobs:
     strategy:
       matrix:
         part: [0, 1]
-    timeout-minutes: 60
+    timeout-minutes: 150
     steps:
       - name: Checkout code
         uses: actions/checkout@v4

python/sglang/multimodal_gen/registry.py

@@ -336,11 +336,17 @@ class ModelInfo:
     pipeline_config_cls: Type[PipelineConfig]
 
 
-def _get_diffusers_model_info() -> ModelInfo:
+def _get_diffusers_model_info(
+    model_path: Optional[str] = None,
+    model_id: Optional[str] = None,
+) -> ModelInfo:
     """
     Get model info for diffusers backend.
 
     Returns a ModelInfo with DiffusersPipeline and generic configs.
+    When model_path is provided and has a registered native config,
+    inherits task_type from it so that validation (e.g. accepts_image_input)
+    works correctly even under the diffusers backend.
     """
     from sglang.multimodal_gen.configs.pipeline_configs.diffusers_generic import (
         DiffusersGenericPipelineConfig,
@@ -352,10 +358,34 @@ def _get_diffusers_model_info() -> ModelInfo:
         DiffusersPipeline,
     )
 
+    pipeline_config_cls = DiffusersGenericPipelineConfig
+
+    # If there is a registered native config for this model, inherit its task_type
+    if model_path is not None:
+        config_info = _get_config_info(model_path, model_id=model_id)
+        if config_info is not None:
+            native_task_type = config_info.pipeline_config_cls.task_type
+            if native_task_type != DiffusersGenericPipelineConfig.task_type:
+                pipeline_config_cls = dataclasses.make_dataclass(
+                    "DiffusersGenericPipelineConfig",
+                    [
+                        (
+                            "task_type",
+                            type(native_task_type),
+                            dataclasses.field(default=native_task_type),
+                        )
+                    ],
+                    bases=(DiffusersGenericPipelineConfig,),
+                )
+                logger.debug(
+                    "Inherited task_type=%s from native config for diffusers backend",
+                    native_task_type.name,
+                )
+
     return ModelInfo(
         pipeline_cls=DiffusersPipeline,
         sampling_param_cls=DiffusersGenericSamplingParams,
-        pipeline_config_cls=DiffusersGenericPipelineConfig,
+        pipeline_config_cls=pipeline_config_cls,
     )
 
 
@@ -392,7 +422,7 @@ def get_model_info(
         logger.info(
             "Using diffusers backend for model '%s' (explicitly requested)", model_path
         )
-        return _get_diffusers_model_info()
+        return _get_diffusers_model_info(model_path=model_path, model_id=model_id)
 
     # For AUTO or SGLANG backend, try native implementation first
     # 1. Discover all available pipeline classes and cache them
@@ -407,7 +437,7 @@ def get_model_info(
             "Falling back to diffusers backend.",
             model_path,
         )
-        return _get_diffusers_model_info(model_path)
+        return _get_diffusers_model_info(model_path=model_path, model_id=model_id)
 
     # 2. Get pipeline class - check non-diffusers models first
     pipeline_class_name = get_non_diffusers_pipeline_name(model_path)
@@ -427,7 +457,9 @@ def get_model_info(
             logger.error(f"Could not read model config for '{model_path}': {e}")
             if backend == Backend.AUTO:
                 logger.info("Falling back to diffusers backend")
-                return _get_diffusers_model_info()
+                return _get_diffusers_model_info(
+                    model_path=model_path, model_id=model_id
+                )
             return None
 
         pipeline_class_name = config.get("_class_name")
@@ -437,7 +469,9 @@ def get_model_info(
             )
             if backend == Backend.AUTO:
                 logger.info("Falling back to diffusers backend")
-                return _get_diffusers_model_info()
+                return _get_diffusers_model_info(
+                    model_path=model_path, model_id=model_id
+                )
             return None
 
     pipeline_cls = _PIPELINE_REGISTRY.get(pipeline_class_name)
@@ -447,7 +481,7 @@ def get_model_info(
                 f"Pipeline class '{pipeline_class_name}' specified in '{model_path}' has no native sglang support. "
                 f"Falling back to diffusers backend."
             )
-            return _get_diffusers_model_info()
+            return _get_diffusers_model_info(model_path=model_path, model_id=model_id)
         else:
             logger.error(
                 f"Pipeline class '{pipeline_class_name}' specified in '{model_path}' is not a registered EntryClass in the framework. "
@@ -464,7 +498,7 @@ def get_model_info(
                 f"Could not resolve native configuration for model '{model_path}'. "
                 f"Falling back to diffusers backend."
             )
-            return _get_diffusers_model_info()
+            return _get_diffusers_model_info(model_path=model_path, model_id=model_id)
         else:
             logger.error(
                 f"Could not resolve configuration for model '{model_path}'. "

python/sglang/multimodal_gen/runtime/pipelines/diffusers_pipeline.py

@@ -456,7 +456,22 @@ class DiffusersPipeline(ComposedPipelineBase):
             else:
                 raise
 
-        pipe = pipe.to(get_local_torch_device())
+        # Use CPU offload (all-or-nothing in diffusers) if any component offload is requested.
+        any_offload = (
+            server_args.dit_cpu_offload
+            or server_args.text_encoder_cpu_offload
+            or server_args.image_encoder_cpu_offload
+            or server_args.vae_cpu_offload
+        )
+        if any_offload:
+            device = get_local_torch_device()
+            gpu_id = device.index if device.index is not None else 0
+            pipe.enable_model_cpu_offload(gpu_id=gpu_id)
+            logger.info(
+                "Enabled model CPU offload for diffusers pipeline (gpu_id=%d)", gpu_id
+            )
+        else:
+            pipe = pipe.to(get_local_torch_device())
         # Apply VAE memory optimizations from pipeline config
         self._apply_vae_optimizations(pipe, server_args)
         # Apply attention backend if specified