Skip to content

[Bugfix] Fixes for Phi3v and Ultravox Multimodal EmbeddingInputs Support #8979

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 5 commits into from
Oct 5, 2024
Merged

[Bugfix] Fixes for Phi3v and Ultravox Multimodal EmbeddingInputs Support #8979

Changes from 1 commit
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
515037b
fix: bug fixes for phi3v and ultravox image embedding support
hhzhang16 Sep 30, 2024
c5c5c02
lint: fixing lint for ultravox
hhzhang16 Sep 30, 2024
9a87aa6
feat: simplify ultravox input mapper
hhzhang16 Sep 30, 2024
1918192
fix: fixes for single image enbeds input
hhzhang16 Oct 1, 2024
f0a04ba
feat: raise error if phi3v image embeds are not batched 2D tensors
hhzhang16 Oct 3, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 11 additions & 2 deletions vllm/model_executor/models/phi3v.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -437,7 +437,8 @@ def input_processor_for_phi3v(ctx: InputContext,
input_height=h,
num_crops=num_crops))
elif isinstance(image_data, torch.Tensor):
num_images, image_feature_size, hidden_size = image_data.shape
image_feature_size = [image_data.shape[0]]
image_data = [image_data]
elif is_list_of(image_data, torch.Tensor):
image_feature_size = [item.shape[0] for item in image_data]
else:
Expand Down Expand Up @@ -613,7 +614,15 @@ def _process_image_input(
) -> torch.Tensor:

if image_input["type"] == "image_embeds":
return list(torch.unbind(image_input["data"], dim=0))
image_data = image_input["data"]
if is_list_of(image_data, torch.Tensor):
# it's already a list of tensors
return image_data
if len(image_data.shape) == 2:
# 2D tensor
return image_data
Copy link
Member

@DarkLight1337 DarkLight1337 Oct 2, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This should raise an error. We expect batched 2D tensors which can be either a list of 2D tensors or a single 3D tensor.

Copy link
Contributor Author

@hhzhang16 hhzhang16 Oct 3, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Gotcha, okay

Copy link
Contributor Author

@hhzhang16 hhzhang16 Oct 3, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed

# 3D tensor
return list(torch.unbind(image_data, dim=0))

assert self.vision_embed_tokens is not None
image_embeds = self.vision_embed_tokens(image_input["data"],
Expand Down