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

Fix llama4 vision (#7840)

Browse Source 
Signed-off-by: Xinyuan Tong <justinning0323@outlook.com>
This commit is contained in:
Xinyuan Tong
2025-07-08 14:00:03 -07:00
committed by GitHub
parent 2e7ab862e3
commit 4bab50a6b5
3 changed files with 73 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/managers/mm_utils.py
View File

@@ -248,7 +248,9 @@ def _get_chunked_prefill_embedding(
) -> Optional[torch.Tensor]:
# Calculate embedding for each request, try to get it from cache to avoid repeated calculation
embedding_list = []
for i in range(len(items_size) - 1):
# FIXME(Xinyuan): temporary workaround for eagle3, which may have len(items_size) > len(prefix_length)
max_iterations = min(len(items_size) - 1, len(prefix_length))
for i in range(max_iterations):
if items_size[i] == items_size[i + 1]:
continue
embedding_items_per_req = embedding_items[items_size[i] : items_size[i + 1]]
@@ -269,7 +271,7 @@ def _get_chunked_prefill_embedding(
embedding_per_req_chunk, _, end_index = get_embedding_chunk(
embedding=embedding_per_req,
extend_prefix_len=prefix_length[i],
extend_seq_len=extend_length[i],
extend_seq_len=extend_length[i] if i < len(extend_length) else 0,
items_offset=items_offset,
)
# remove this item from cache if chunk reaches to the end