-
Notifications
You must be signed in to change notification settings - Fork 230
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.
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
[transformers snippet] Support pipeline VLMs #1012
base: main
Are you sure you want to change the base?
Conversation
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Has huggingface/transformers#34170 been released? If it hasn't, should we wait until the next transformers release before merging or is it ok?
`image_ny = "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"`, | ||
`image_chicago = "https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg"`, |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The example is super nice, but not all image-text-to-text models support multiple images reliably. I'd go for a simpler single-image example for now.
Agreed with this! another problem is that although most models use the However since you have this check: if (model.tags.includes("conversational") && model.config?.tokenizer_config?.chat_template) { It would be even better and simpler to use chat template for vlms also, with something like this: image_ny = "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
image_chicago = "https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg"
messages = [
{
"role": "user",
"content": [
{"type": "text", "text": "What’s the difference between these two images?"},
{"type": "image"},
{"type": "image"},
],
}
]
outputs = pipe([image_ny, image_chicago], text=messages) |
I completely agree with @yonigozlan best to have chat template |
Description
Rn, if you go to conversational VLM like Llama-3.2-11B-Vision-Instruct, you would not receive high-level pipeline snippet.
On the contrary, if you go to conversational LLM like Llama-3.1-8B-Instruct, you will receive high-level pipeline snippet.
huggingface/transformers#34170 was merged. Therefore, now we can add pipeline snippet for VLM.
Here is an example snippet, one would get