Instructions to use google/paligemma2-3b-ft-docci-448 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use google/paligemma2-3b-ft-docci-448 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="google/paligemma2-3b-ft-docci-448")

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("google/paligemma2-3b-ft-docci-448")
model = AutoModelForMultimodalLM.from_pretrained("google/paligemma2-3b-ft-docci-448", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use google/paligemma2-3b-ft-docci-448 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "google/paligemma2-3b-ft-docci-448"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "google/paligemma2-3b-ft-docci-448",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/google/paligemma2-3b-ft-docci-448

SGLang

How to use google/paligemma2-3b-ft-docci-448 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "google/paligemma2-3b-ft-docci-448" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "google/paligemma2-3b-ft-docci-448",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "google/paligemma2-3b-ft-docci-448" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "google/paligemma2-3b-ft-docci-448",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use google/paligemma2-3b-ft-docci-448 with Docker Model Runner:
```
docker model run hf.co/google/paligemma2-3b-ft-docci-448
```

docci-896 3b, 10b, 28b?

by lucyknada - opened Dec 5, 2024

Discussion

lucyknada

Dec 5, 2024

hi there, just wondering was a docci train not done for 896 resolution for all the sizes and is 28b missing? or will that be uploaded too? thanks

pcuenq

Dec 6, 2024

Hi @lucyknada ! So far, Google only fine-tuned on DOCCI the 3B and 10B variants, at the 448 resolution. I think that 896 would be most helpful for tasks that benefit from finer details, such as OCR or text extraction. Just curious, is there a particular task you have in mind?

lucyknada

Dec 6, 2024

I agree, mostly finer details in general where the extra res could help; also non docci tunes seemed to only output a few words each so any task is affected by this pretty much; desktop screenshot descriptions and guidance, OCR, regular detailed image descriptions, large image details (e.g. traffic or manufacturing monitoring) and much more, possibly I wasn't using the non docci ones right? but even the huggingface demo was suffering from the same issue(s).

pcuenq

Dec 7, 2024

Yes, the demo is running on a quick VQAv2 fine-tune, whose answers are short. But as the DOCCI checkpoints demonstrate, it is indeed possible to fine-tune to much longer and detailed responses.

lucyknada

Dec 9, 2024

•

edited Dec 9, 2024

Thanks! not just the demo however, I got very short responses from the regular non-docci checkpoints too, hence why larger resolution with docci would be nice

lkv

Google org Aug 11, 2025

This comment has been hidden

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment