A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 03:15 UTC
[Bugfix] Validate and concatenate image embeddings in MiniCPMVBaseModel (#11631)
whyiug committed
Dec 31, 2024 at 02:51 UTC
36e76700453924c8d421db99af70a88a1df835cd
1 file changed
+6
vllm/model_executor/models/minicpmv.py
+6
@@ -487,6 +487,12 @@ class MiniCPMVBaseModel(nn.Module, SupportsMultiModal, SupportsPP):
487
image_embeds = kwargs.pop("image_embeds", None)
488
489
if image_embeds is not None:
490
+ if not isinstance(image_embeds, (torch.Tensor, list)):
491
+ raise ValueError(f"Incorrect type of image embeds. "
492
+ f"Got type: {type(image_embeds)}")
493
+ if isinstance(image_embeds, list):
494
+ image_embeds = torch.concat(image_embeds)
495
+
496
return MiniCPMVImageEmbeddingInputs(
497
image_bounds=self._get_image_bounds(input_ids, im_start_id,
498
im_end_id, slice_start_id,