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.