chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Support precomputed multimodal features for Qwen-VL and Gemma3 models. (#6136)

Browse Source 
Co-authored-by: Yury Sulsky <ysulsky@tesla.com>
This commit is contained in:
Yury Sulsky
2025-05-16 12:26:15 -07:00
committed by GitHub
parent c23a7072b6
commit f19a9204cd
14 changed files with 592 additions and 125 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
python/sglang/srt/managers/mm_utils.py
View File

@@ -368,13 +368,13 @@ def general_mm_embed_routine(
input_ids: torch.Tensor,
forward_batch: ForwardBatch,
language_model: nn.Module,
image_data_embedding_func: Callable[
[List[MultimodalDataItem]], torch.Tensor
image_data_embedding_func: Optional[
Callable[[List[MultimodalDataItem]], torch.Tensor]
] = None,
audio_data_embedding_func: Callable[
[List[MultimodalDataItem]], torch.Tensor
audio_data_embedding_func: Optional[
Callable[[List[MultimodalDataItem]], torch.Tensor]
] = None,
placeholder_tokens: dict[Modality, List[int]] = None,
placeholder_tokens: Optional[dict[Modality, List[int]]] = None,
**kwargs,
) -> torch.Tensor:
"""
@@ -389,7 +389,6 @@ def general_mm_embed_routine(
forwarded hidden states
"""
assert hasattr(language_model, "get_input_embeddings")
embed_tokens = language_model.get_input_embeddings()
if (