diff --git a/vllm/model_executor/models/step3_vl.py b/vllm/model_executor/models/step3_vl.py index f057a6e3fd..1ab1c4ad66 100644 --- a/vllm/model_executor/models/step3_vl.py +++ b/vllm/model_executor/models/step3_vl.py @@ -142,8 +142,11 @@ class Step3VisionProcessor: class ImagePatcher: + def __init__(self, enable_patch: bool = True) -> None: + self.enable_patch = enable_patch + def determine_window_size(self, long: int, short: int) -> int: - if long <= 728: + if long < 728: return short if long / short > 1.5 else 0 return min(short, 504) if long / short > 4 else 504 @@ -241,7 +244,7 @@ class ImagePatcher: window_size = self.determine_window_size( max(img_height, img_width), min(img_height, img_width) ) - if window_size == 0: + if window_size == 0 or not self.enable_patch: return 0, 0 else: img_width, img_height = self.get_image_size_for_crop( @@ -277,7 +280,7 @@ class ImagePatcher: max(new_img_height, new_img_width), min(new_img_height, new_img_width) ) - if window_size == 0: + if window_size == 0 or not self.enable_patch: return img, [], None else: new_img_width, new_img_height = self.get_image_size_for_crop( @@ -327,7 +330,6 @@ class Step3VLProcessor: self.config = config self.tokenizer = tokenizer - self.image_size = 728 self.patch_size = 504 self.image_preprocessor = Step3VisionProcessor( @@ -340,7 +342,10 @@ class Step3VLProcessor: self.image_feature_placeholder = self.image_token * self.num_image_feature_size self.patch_feature_placeholder = self.image_token * self.num_patch_feature_size - self.patcher = ImagePatcher() + # Respect vision config switch to enable/disable patch extraction. + # For video understanding, it's preferable to disable patch. + enable_patch = getattr(self.config.vision_config, "enable_patch", True) + self.patcher = ImagePatcher(enable_patch=enable_patch) @property def image_token_id(self) -> int: