Instructions to use google/paligemma-3b-pt-224 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use google/paligemma-3b-pt-224 with Transformers:

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

pipe = pipeline("image-text-to-text", model="google/paligemma-3b-pt-224")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("google/paligemma-3b-pt-224")
model = AutoModelForMultimodalLM.from_pretrained("google/paligemma-3b-pt-224", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use google/paligemma-3b-pt-224 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "google/paligemma-3b-pt-224"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "google/paligemma-3b-pt-224",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/google/paligemma-3b-pt-224

SGLang

How to use google/paligemma-3b-pt-224 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/paligemma-3b-pt-224" \
    --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/paligemma-3b-pt-224",
		"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/paligemma-3b-pt-224" \
        --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/paligemma-3b-pt-224",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use google/paligemma-3b-pt-224 with Docker Model Runner:
```
docker model run hf.co/google/paligemma-3b-pt-224
```

A from-scratch PyTorch implementation of this model, bit-exact with transformers, plus an interactive explainer

#32

by melihgenel - opened 8 days ago

Discussion

melihgenel

8 days ago

Hi! I rebuilt PaliGemma (SigLIP + linear projector + Gemma) from scratch in plain PyTorch and loaded these weights into it, to
understand the architecture end to end.

With paligemma-3b-pt-224 in bf16 it matches transformers exactly: logits are bit-identical, layer-by-layer error is 0, and greedy
generation is token-for-token the same, with and without a KV cache.

Two details that might help others comparing implementations:

HF only makes the prefix (image + prompt) bidirectional when token_type_ids is passed. Without it the mask is fully causal, which
differs from the paper and gives very different logits.
Only text embeddings are scaled by √d. Dividing image features by √d and letting Gemma multiply everything is the same math, but
in bf16 it adds two roundings and shows up as a ~0.2 mean logit difference.

Code: https://github.com/mlihgenel/paligemma-from-scratch
Interactive handbook (English/Turkish): https://mlihgenel.github.io/paligemma-from-scratch/

Feedback welcome!

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