[VLM][Core] Support profiling with multiple multi-modal inputs per prompt (#7126)

This commit is contained in:
Cyrus Leung
2024-08-14 17:55:42 +00:00
committed by GitHub
parent 70b746efcf
commit 3f674a49b5
38 changed files with 572 additions and 216 deletions
+4 -3
View File
@@ -31,13 +31,13 @@ def get_blip_num_patches(*, image_size: int, patch_size: int) -> int:
def get_blip_image_feature_size(
hf_config: Union[BlipVisionConfig, Blip2VisionConfig], ) -> int:
hf_config: Union[BlipVisionConfig, Blip2VisionConfig]) -> int:
return get_blip_num_patches(image_size=hf_config.image_size,
patch_size=hf_config.patch_size)
def get_max_blip_image_tokens(
hf_config: Union[BlipVisionConfig, Blip2VisionConfig], ) -> int:
hf_config: Union[BlipVisionConfig, Blip2VisionConfig]) -> int:
return get_blip_image_feature_size(hf_config)
@@ -60,6 +60,7 @@ def dummy_seq_data_for_blip(
def dummy_image_for_blip(
hf_config: Union[BlipVisionConfig, Blip2VisionConfig],
num_images: int,
*,
image_width_override: Optional[int] = None,
image_height_override: Optional[int] = None,
@@ -71,7 +72,7 @@ def dummy_image_for_blip(
height = image_height_override
image = Image.new("RGB", (width, height), color=0)
return {"image": image}
return {"image": image if num_images == 1 else [image] * num_images}
def input_processor_for_blip(