[Bugfix] Fix config misalignment between offline and online diffusion inference (Wan2.2, Qwen-Image series) (#1979)

Signed-off-by: samithuang <285365963@qq.com>
Signed-off-by: Samit <285365963@qq.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

vllm_omni/diffusion/models/qwen_image/pipeline_qwen_image_layered.py

@@ -95,6 +95,9 @@ def get_qwen_image_layered_pre_process_func(
             else:
                 image = cast(PIL.Image.Image | torch.Tensor | np.ndarray, raw_image)
 
+            if isinstance(image, PIL.Image.Image) and image.mode != "RGBA":
+                image = image.convert("RGBA")
+
             # 1. calculate dimensions
             image_size = image.size
             assert request.sampling_params.resolution in [640, 1024], (
@@ -652,6 +655,8 @@ the image\n<|vision_start|><|image_pad|><|vision_end|><|im_end|>\n<|im_start|>as
             width = req.sampling_params.width
         else:
             # fallback to run pre-processing in pipeline (debug only)
+            if isinstance(image, PIL.Image.Image) and image.mode != "RGBA":
+                image = image.convert("RGBA")
             image_size = image[0].size if isinstance(image, list) else image.size
             assert resolution in [640, 1024], f"resolution must be either 640 or 1024, but got {resolution}"
             calculated_width, calculated_height = calculate_dimensions(

vllm_omni/diffusion/request.py

@@ -28,17 +28,23 @@ class OmniDiffusionRequest:
 
     def __post_init__(self):
         """Initialize dependent fields after dataclass initialization."""
+        # Detect whether user explicitly provided guidance_scale.
+        # The sentinel default is 0.0 (false-like); any truthy value means
+        # the caller set it intentionally.  We must resolve this BEFORE
+        # auto-filling guidance_scale_2, otherwise the sentinel leaks into
+        # guidance_scale_2.
+        if self.sampling_params.guidance_scale:
+            self.sampling_params.guidance_scale_provided = True
+        else:
+            self.sampling_params.guidance_scale = 1.0
+
         # Set do_classifier_free_guidance based on guidance scale and negative prompt
         if self.sampling_params.guidance_scale > 1.0 and any(
             (not isinstance(p, str) and p.get("negative_prompt")) for p in self.prompts
         ):
             self.sampling_params.do_classifier_free_guidance = True
+
+        # Auto-fill guidance_scale_2 from the (now-resolved) guidance_scale
+        # so downstream code always has a valid value.
         if self.sampling_params.guidance_scale_2 is None:
             self.sampling_params.guidance_scale_2 = self.sampling_params.guidance_scale
-
-        # The dataclass default value is 0 (false-like), used to detect whether user explicitly provides this value
-        # After this check is done, reset this value to old default 1
-        if self.sampling_params.guidance_scale:
-            self.sampling_params.guidance_scale_provided = True
-        else:
-            self.sampling_params.guidance_scale = 1.0

vllm_omni/diffusion/worker/diffusion_model_runner.py

@@ -221,6 +221,7 @@ class DiffusionModelRunner:
                 not getattr(req, "skip_cache_refresh", False)
                 and self.cache_backend is not None
                 and self.cache_backend.is_enabled()
+                and req.sampling_params.num_inference_steps is not None
             ):
                 self.cache_backend.refresh(self.pipeline, req.sampling_params.num_inference_steps)
 

vllm_omni/entrypoints/async_omni_diffusion.py

@@ -172,9 +172,6 @@ class AsyncOmniDiffusion:
         if request_id is None:
             request_id = f"diff-{uuid.uuid4().hex[:16]}"
 
-        if sampling_params.guidance_scale:
-            sampling_params.guidance_scale_provided = True
-
         if lora_request is not None:
             sampling_params.lora_request = lora_request
 

vllm_omni/entrypoints/openai/serving_chat.py

@@ -2050,20 +2050,26 @@ class OmniOpenAIServingChat(OpenAIServingChat, AudioMixin):
                 except ValueError:
                     logger.warning("Invalid size format: %s", extra_body.get("size"))
 
-            # Get request parameters from extra_body
-            # Text-to-image parameters (ref: text_to_image.py)
-            num_inference_steps = extra_body.get("num_inference_steps", 50)
+            # Get request parameters from extra_body.
+            # Avoid hardcoded defaults here — let each pipeline's forward()
+            # method apply its own model-specific default when the user does
+            # not provide a value.
+            num_inference_steps = extra_body.get("num_inference_steps")
             guidance_scale = extra_body.get("guidance_scale")
-            true_cfg_scale = extra_body.get("true_cfg_scale")  # Qwen-Image specific
+            true_cfg_scale = extra_body.get("true_cfg_scale") or extra_body.get("cfg_scale")
             seed = extra_body.get("seed")
             negative_prompt = extra_body.get("negative_prompt")
             num_outputs_per_prompt = extra_body.get("num_outputs_per_prompt", 1)
 
             # Text-to-video parameters (ref: text_to_video.py)
             num_frames = extra_body.get("num_frames")
-            guidance_scale_2 = extra_body.get("guidance_scale_2")  # For video high-noise CFG
+            guidance_scale_2 = extra_body.get("guidance_scale_2")
             lora_body = extra_body.get("lora")
 
+            # Qwen-Image-Layered parameters
+            layers = extra_body.get("layers")
+            resolution = extra_body.get("resolution")
+
             logger.info(
                 "Diffusion chat request %s: prompt=%r, ref_images=%d, params=%s",
                 request_id,
@@ -2087,25 +2093,27 @@ class OmniOpenAIServingChat(OpenAIServingChat, AudioMixin):
                 "negative_prompt": negative_prompt,
             }
             gen_params = OmniDiffusionSamplingParams(
-                num_inference_steps=num_inference_steps,
                 height=height,
                 width=width,
                 num_outputs_per_prompt=num_outputs_per_prompt,
                 seed=seed,
             )
 
+            # Only override defaults when the user explicitly provides values
+            if num_inference_steps is not None:
+                gen_params.num_inference_steps = num_inference_steps
             if guidance_scale is not None:
                 gen_params.guidance_scale = guidance_scale
-
-            # Add Qwen-Image specific parameter
             if true_cfg_scale is not None:
                 gen_params.true_cfg_scale = true_cfg_scale
-
-            # Add video generation parameters if set
             if num_frames is not None:
                 gen_params.num_frames = num_frames
             if guidance_scale_2 is not None:
                 gen_params.guidance_scale_2 = guidance_scale_2
+            if layers is not None:
+                gen_params.layers = layers
+            if resolution is not None:
+                gen_params.resolution = resolution
 
             # Parse per-request LoRA (works for both AsyncOmniDiffusion and AsyncOmni).
             if lora_body and isinstance(lora_body, dict):

vllm_omni/inputs/data.py

@@ -234,8 +234,9 @@ class OmniDiffusionSamplingParams:
     step_index: int | None = None
     boundary_ratio: float | None = None
 
-    # Scheduler parameters
-    num_inference_steps: int = 50
+    # Scheduler parameters – ``None`` means "not explicitly set by the caller";
+    # each pipeline's ``forward()`` decides its own model-specific default.
+    num_inference_steps: int | None = None
     guidance_scale: float = 0.0
     guidance_scale_provided: bool = False
     guidance_scale_2: float | None = None