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, 04:30 UTC
[V1][VLM] Fix edge case bug for InternVL2 (#11165)
Signed-off-by: Roger Wang <ywang@roblox.com>
Roger Wang committed
Dec 13, 2024 at 03:09 UTC
969da7d70bc0539f6be12027b71bef758325a61a
1 file changed
+4
-1
vllm/model_executor/models/internvl.py
+4
-1
@@ -669,8 +669,11 @@ class InternVLChatModel(nn.Module, SupportsMultiModal, SupportsPP):
669
image_embeds = self.extract_feature(image_input["data"])
670
671
patches_per_image = image_input["patches_per_image"]
672
+
673
+ # Only one image in the current batch
674
if len(patches_per_image) == 1:
673
- image_embeds = image_embeds.unsqueeze(0)
675
+ image_embeds = image_embeds.view(
676
+ -1, self.config.text_config.hidden_size).unsqueeze(0)
677
return image_embeds
678
679
# NOTE: Image embeddings are split into separate tensors for each image