chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

vlm: fix tiny multimodal cache bug (#12984)

Browse Source
This commit is contained in:
Yuhao Yang
2025-11-10 21:30:19 +08:00
committed by GitHub
parent 5639145fac
commit 1240ac13b8
2 changed files with 1 additions and 60 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/managers/mm_utils.py
View File

@@ -380,7 +380,7 @@ def _get_chunked_prefill_embedding(
# if all items has been prefixed, we do not need to calculate embedding
if all([offset_end < prefix_length[i] for _, offset_end in items_offset]):
continue
item_hashes = [item.hash for item in embedding_items]
item_hashes = [item.hash for item in embedding_items_per_req]
embedding_items_hash = MultiModalStaticCache.combine_hashes(item_hashes)
embedding_per_req = embedding_cache.get(item_hashes)
if embedding_per_req is None: