[Models] Remove GPU-CPU sync when do_pan_and_scan=false in Gemma3

vllm/model_executor/models/gemma3_mm.py

@@ -8,6 +8,7 @@
 from torch import nn
 from transformers import BatchFeature, Gemma3Config, Gemma3Processor
 from transformers.models.gemma3.processing_gemma3 import Gemma3ProcessorKwargs
+from typing_extensions import NotRequired
 
 import vllm.envs as envs
 from vllm.config import VllmConfig
@@ -52,7 +53,7 @@ class Gemma3ImagePixelInputs(TypedDict):
     over each image over each prompt in the batch.
     """
 
-    num_patches: torch.Tensor
+    num_patches: NotRequired[torch.Tensor]
     """Shape: `(batch_size * num_images)`"""
 
 
@@ -266,6 +267,12 @@ def _call_hf_processor(
             mm_kwargs,
         )
 
+        hf_processor = self.info.get_hf_processor(**mm_kwargs)
+        images_kwargs = self.info._resolve_image_kwargs(
+            hf_processor, {"do_pan_and_scan"})
+        if not images_kwargs["do_pan_and_scan"]:
+            return processed_outputs
+
         # HF processor pops the `num_crops` kwarg, which is needed by vLLM
         if (images := mm_data.get("images")) is not None:
             parsed_images = (self._get_data_parser().parse_mm_data({
@@ -276,7 +283,6 @@ def _call_hf_processor(
                 parsed_images.get_image_size(i)
                 for i in range(len(parsed_images))
             ]
-            hf_processor = self.info.get_hf_processor(**mm_kwargs)
 
             num_crops = [
                 self.info.get_num_crops(image_width=size.width,
@@ -293,6 +299,8 @@ def _get_mm_fields_config(
         hf_inputs: BatchFeature,
         hf_processor_mm_kwargs: Mapping[str, object],
     ) -> Mapping[str, MultiModalFieldConfig]:
+        if "num_crops" not in hf_inputs:
+            return dict(pixel_values=MultiModalFieldConfig.batched("image"))
         num_crops = hf_inputs.get("num_crops", torch.empty(0))
 
         return dict(
@@ -535,17 +543,22 @@ def _parse_and_validate_image_input(
         if not isinstance(pixel_values, (torch.Tensor, list)):
             raise ValueError("Incorrect type of pixel values. "
                              f"Got type: {type(pixel_values)}")
+        pixel_values = flatten_bn(pixel_values, concat=True)
+        pixel_values = self._validate_pixel_values(pixel_values)
+
+        if num_crops is None:
+            return Gemma3ImagePixelInputs(type="pixel_values",
+                                          pixel_values=pixel_values)
 
         if not isinstance(num_crops, (torch.Tensor, list)):
             raise ValueError("Incorrect type of num_crops. "
                              f"Got type: {type(num_crops)}")
 
-        pixel_values = flatten_bn(pixel_values, concat=True)
         num_crops = flatten_bn(num_crops, concat=True)
 
         return Gemma3ImagePixelInputs(
             type="pixel_values",
-            pixel_values=self._validate_pixel_values(pixel_values),
+            pixel_values=pixel_values,
             num_patches=num_crops + 1,
         )
 
@@ -563,14 +576,16 @@ def _process_image_input(
         assert self.vision_tower is not None
 
         pixel_values = image_input["pixel_values"]
-        num_patches = image_input["num_patches"]
+        num_patches = image_input.get("num_patches")
 
         image_features = self._image_pixels_to_features(
             self.vision_tower,
             pixel_values,
         )
         image_embeds = self.multi_modal_projector(image_features)
 
+        if num_patches is None:
+            return image_embeds
         return [
             e.flatten(0, 1) for e in image_embeds.split(num_patches.tolist())
         ]