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

model: support mllama4 (#5144)

Browse Source
This commit is contained in:
Mick
2025-04-10 00:28:44 +08:00
committed by GitHub
parent 87eddedfa2
commit fbebcb7aa4
7 changed files with 145 additions and 65 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
python/sglang/srt/managers/mm_utils.py
View File

@@ -148,7 +148,8 @@ def get_embedding_and_mask(
placeholder_tensor,
).unsqueeze(-1)
num_mm_tokens_in_input_ids = special_multimodal_mask.sum()
num_mm_tokens_in_input_ids = special_multimodal_mask.sum().item()
if num_mm_tokens_in_input_ids != num_mm_tokens_in_embedding:
logger.warning(
f"Number of tokens in multimodal embedding does not match those in the input text."
@@ -172,7 +173,7 @@ def get_embedding_and_mask(
embedding = embedding[-num_multimodal:, :]
else:
raise RuntimeError(
"Insufficient multimodal embedding length. This is an internal error"
f"Insufficient multimodal embedding length: {num_mm_tokens_in_input_ids=} vs {num_mm_tokens_in_embedding=}. This is an internal error"
)
return embedding, special_multimodal_mask