[diffusion] fix: fix bugs to let LTX-2 pipeline support latest Sglang Args pipelines (#19295)

python/sglang/multimodal_gen/configs/pipeline_configs/base.py

@@ -153,6 +153,7 @@ class PipelineConfig:
     """The base configuration class for a generation pipeline."""
 
     task_type: ModelTaskType = ModelTaskType.I2I
+    skip_input_image_preprocess: bool = False
 
     model_path: str = ""
     pipeline_config_path: str | None = None

python/sglang/multimodal_gen/configs/pipeline_configs/ltx_2.py

@@ -115,7 +115,8 @@ def _gemma_postprocess_func(
 class LTX2PipelineConfig(PipelineConfig):
     """Configuration for LTX-Video pipeline."""
 
-    task_type: ModelTaskType = ModelTaskType.T2V
+    task_type: ModelTaskType = ModelTaskType.TI2V
+    skip_input_image_preprocess: bool = True
     dit_config: LTX2Config = field(default_factory=LTX2Config)
 
     # Model architecture

python/sglang/multimodal_gen/runtime/pipelines_core/stages/input_validation.py

@@ -140,6 +140,8 @@ class InputValidationStage(PipelineStage):
                 batch.height = height
 
         elif server_args.pipeline_config.task_type == ModelTaskType.TI2V:
+            if server_args.pipeline_config.skip_input_image_preprocess:
+                return
             # duplicate with vae_image_processor
             # further processing for ti2v task
             if isinstance(