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

[feat] Support session control for vision language models (#2210)

Browse Source
This commit is contained in:
Ying Sheng
2024-11-27 00:03:29 -08:00
committed by GitHub
parent c754652fcd
commit 37c8a5761f
7 changed files with 265 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
python/sglang/srt/models/llava.py
View File

@@ -49,7 +49,13 @@ class LlavaBaseForCausalLM(nn.Module):
image_sizes, pad_values = image_inputs.image_sizes, image_inputs.pad_values
# hardcode for spatial_unpad + anyres
image_aspect_ratio = "anyres" if len(image_sizes) == 1 else "pad"
if image_inputs.modalities is not None and (
"multi-images" in image_inputs.modalities
or "video" in image_inputs.modalities
):
image_aspect_ratio = "pad"
else:
image_aspect_ratio = "anyres"
offset_list = []
for image_s in image_sizes:
if len(image_sizes) > 16: