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

6
python/sglang/srt/models/qwen2_5_vl.py
View File

@@ -499,12 +499,6 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
return pattern.pad_input_tokens(input_ids, mm_inputs)
def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
if any(item.precomputed_features is not None for item in items):
if not all(item.precomputed_features is not None for item in items):
raise NotImplementedError(
"MM inputs where only some items are precomputed."
)
return torch.concat([item.precomputed_features for item in items])
# in qwen-vl, last dim is the same
pixel_values = torch.cat([item.pixel_values for item in items], dim=0).type(
self.visual.dtype