• Gemma
• • • Learn more about Gemma
  • Quick start
  • • Installation
    • Downloading the models
    • Running the unit tests
  • Examples
  • System Requirements
  • Contributing
  • License
  • Disclaimer

Gemma

Gemma is a family of open-weights Large Language
Model (LLM) by Google DeepMind, based on Gemini
research and technology.

This repository contains an inference implementation and examples, based on the
Flax and JAX.

Learn more about Gemma

• The Gemma technical report
  details the models' capabilities.
• For tutorials, reference implementations in other ML frameworks, and more,
  visit https://ai.google.dev/gemma.

Quick start

Installation

1. To install Gemma you need to use Python 3.10 or higher.

2. Install JAX for CPU, GPU or TPU. Follow instructions at
   the JAX website.

3. Run

python -m venv gemma-demo
. gemma-demo/bin/activate
pip install git+https://github.com/google-deepmind/gemma.git

Downloading the models

The model checkpoints are available through Kaggle at
http://kaggle.com/models/google/gemma. Select one of the Flax model
variations, click the ⤓ button to download the model archive, then extract the
contents to a local directory.

Alternatively, visit the gemma
models on the Hugging Face Hub. To download the model, you can run the following code if you have huggingface_hub installed:

from huggingface_hub import snapshot_download

local_dir = snapshot_download(repo_id="google/gemma-2b-flax")
snapshot_download(repo_id="google/gemma-2b-flax", local_dir=local_dir)

In both cases, the archive contains both the model weights and
the tokenizer, for example the 2b Flax variation contains:

2b/              # Directory containing model weights
tokenizer.model  # Tokenizer

Running the unit tests

To run the unit tests, install the optional [test] dependencies (e.g. using
pip install -e .[test] from the root of the source tree), then:

pytest .

Note that the tests in sampler_test.py are skipped by default since no
tokenizer is distributed with the Gemma sources. To run these tests, download a
tokenizer following the instructions above, and update the _VOCAB constant in
sampler_test.py with the path to tokenizer.model.

Examples

To run the example sampling script, pass the paths to the weights directory and
tokenizer:

python examples/sampling.py \
  --path_checkpoint=/path/to/archive/contents/2b/ \
  --path_tokenizer=/path/to/archive/contents/tokenizer.model

There are also several Colab notebook tutorials:

• colabs/sampling_tutorial.ipynb
  contains a Colab notebook with a sampling example.

• colabs/fine_tuning_tutorial.ipynb
  contains a Colab with a basic tutorial on how to fine
  tune Gemma for a task, such as English to French translation.

• colabs/gsm8k_eval.ipynb
  is a Colab with a reference GSM8K eval
  implementation.

To run these notebooks you will need to download a local copy of the weights and
tokenizer (see above), and update the ckpt_path and vocab_path variables
with the corresponding paths.

System Requirements

Gemma can run on a CPU, GPU and TPU. For GPU, we recommend a 8GB+ RAM on GPU for
the 2B checkpoint and 24GB+ RAM on GPU for the 7B checkpoint.

Contributing

We are open to bug reports, pull requests (PR), and other contributions. Please
see CONTRIBUTING.md for details on PRs.

License

Copyright 2024 DeepMind Technologies Limited

This code is licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License. You may obtain
a copy of the License at http://www.apache.org/licenses/LICENSE-2.0.

Unless required by applicable law or agreed to in writing, software distributed
under the License is distributed on an AS IS BASIS, WITHOUT WARRANTIES OR
CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.

Disclaimer

This is not an official Google product.