EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Refactor vlm embedding routine to use precomputed feature (#6543)

Browse Source 
Signed-off-by: Xinyuan Tong <justinning0323@outlook.com>
This commit is contained in:
Xinyuan Tong
2025-05-24 18:39:21 -07:00
committed by GitHub
parent 0d47788025
commit 681fdc264b
8 changed files with 285 additions and 203 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
python/sglang/srt/managers/multimodal_processors/qwen_vl.py
View File

@@ -144,12 +144,11 @@ class Qwen2_5VLImageProcessor(SGLangBaseProcessor):
if base_output.images:
if images_are_preprocessed:
image_grid_thw = torch.concat(
[
torch.as_tensor(item.image_grid_thws)
for item in base_output.images
]
)
all_image_grid_thws = [
item.image_grid_thws
for item in base_output.images
if item.image_grid_thws is not None
]
all_pixel_values = [
item.pixel_values
for item in base_output.images
@@ -160,6 +159,9 @@ class Qwen2_5VLImageProcessor(SGLangBaseProcessor):
for item in base_output.images
if item.precomputed_features is not None
]
image_grid_thw = (
torch.concat(all_image_grid_thws) if all_image_grid_thws else None
)
pixel_values = (
torch.concat(all_pixel_values) if all_pixel_values else None
)