biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

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

Browse Source
This commit is contained in:
Cyrus Leung
2024-08-15 01:55:42 +08:00
committed by GitHub
parent 70b746efcf
commit 3f674a49b5
38 changed files with 572 additions and 216 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/model_executor/models/siglip.py
View File

@@ -52,6 +52,7 @@ def get_max_siglip_image_tokens(hf_config: SiglipVisionConfig) -> int:
def dummy_seq_data_for_siglip(
hf_config: SiglipVisionConfig,
seq_len: int,
num_images: int,
*,
image_token_id: int,
image_feature_size_override: Optional[int] = None,
@@ -61,13 +62,14 @@ def dummy_seq_data_for_siglip(
else:
image_feature_size = image_feature_size_override
token_ids = [image_token_id] * image_feature_size
token_ids += [0] * (seq_len - image_feature_size)
token_ids = [image_token_id] * image_feature_size * num_images
token_ids += [0] * (seq_len - image_feature_size * num_images)
return SequenceData(token_ids)
def dummy_image_for_siglip(
hf_config: SiglipVisionConfig,
num_images: int,
*,
image_width_override: Optional[int] = None,
image_height_override: Optional[int] = None,
@@ -79,7 +81,7 @@ def dummy_image_for_siglip(
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_siglip(