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

support clip embedding model (#4506)

Browse Source
This commit is contained in:
Pan Lyu
2025-03-27 15:18:15 +08:00
committed by GitHub
parent 1afe3d0798
commit c913ed4046
8 changed files with 746 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
python/sglang/srt/openai_api/adapter.py
View File

@@ -1651,18 +1651,19 @@ def v1_embedding_request(all_requests, tokenizer_manager):
elif isinstance(prompt, list) and isinstance(
prompt[0], MultimodalEmbeddingInput
):
assert (
chat_template_name is not None
), "chat_template_name is required for multimodal inputs"
texts = []
images = []
for item in prompt:
texts.append(item.text if item.text is not None else None)
# TODO simply use padding for text, we should use a better way to handle this
texts.append(item.text if item.text is not None else "padding")
images.append(item.image if item.image is not None else None)
convs = generate_embedding_convs(texts, images, chat_template_name)
generate_prompts = []
for conv in convs:
generate_prompts.append(conv.get_prompt())
if chat_template_name is not None:
convs = generate_embedding_convs(texts, images, chat_template_name)
for conv in convs:
generate_prompts.append(conv.get_prompt())
else:
generate_prompts = texts
if len(generate_prompts) == 1:
prompt_kwargs = {"text": generate_prompts[0], "image_data": images[0]}
else: