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

[Minor] fix the style for multimodal models (#2257)

Browse Source
This commit is contained in:
Lianmin Zheng
2024-11-29 04:24:20 -08:00
committed by GitHub
parent f50a6cf443
commit afe1e46586
3 changed files with 14 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/models/llava.py
View File

@@ -134,7 +134,6 @@ class LlavaBaseForCausalLM(nn.Module):
image_inputs = forward_batch.image_inputs
if forward_batch.forward_mode.is_extend():
bs = forward_batch.batch_size
# Got List[List[str]] extend it to List[str]
# The length of the List should be equal to batch size
modalities_list = []
@@ -142,7 +141,7 @@ class LlavaBaseForCausalLM(nn.Module):
for im in image_inputs:
if im and im.modalities is not None:
modalities_list.extend(im.modalities)
if im and im.image_offsets is not None:
if im and im.image_offsets:
max_image_offset.append(max(im.image_offsets))
else:
max_image_offset.append(-1)
@@ -159,6 +158,7 @@ class LlavaBaseForCausalLM(nn.Module):
need_vision = start_positions <= np.array(max_image_offset)
if need_vision.any():
bs = forward_batch.batch_size
pixel_values = [
image_inputs[i].pixel_values for i in range(bs) if need_vision[i]
]