sunmarc/local-gemma

Gemma 2 optimized for your local machine.

Run Gemma-2 locally in Python, fast!

This repository provides an easy way to run Gemma-2 locally directly from your CLI (or via a Python library) and fast. It is built on top of the 🤗 Transformers and bitsandbytes libraries.

It can be configured to give fully equivalent results to the original implementation, or reduce memory requirements down to just the largest layer in the model!

Installation

There are two installation flavors of local-gemma, which you can select depending on your use case:

pipx - Ideal for CLI

First, follow the installation steps here to install pipx on your environment.

Then, run one of the commands below, depending on your machine.

CUDA

pipx install local-gemma"[cuda]"

MPS

pipx install local-gemma"[mps]"

CPU

pipx install local-gemma"[cpu]"

pip - Ideal for Python (CLI + API)

Local Gemma-2 can be installed as a hardware-specific Python package through pip. The only requirement is a Python installation, details for which can be found here. You can check you have a Python installed locally by running:

python3 --version

(optional) Create a new Python environment

python3 -m venv gemma-venv
source gemma-venv/bin/activate

CUDA

pip install local-gemma"[cuda]"

MPS

pip install local-gemma"[mps]"

CPU

pip install local-gemma"[cpu]"

CLI Usage

You can chat with the Gemma-2 through an interactive session by calling:

local-gemma

Tip

Local Gemma will check for a Hugging Face "read" token to download the model. You can follow this guide to create a token, and pass it when prompted to log-in. If you're new to Hugging Face and never used a Gemma model, you'll also need to accept the terms at the top of this page.

Alternatively, you can request a single output by passing a prompt, such as:

local-gemma "What is the capital of France?"

By default, this loads the Gemma-2 9b it model. To load the Gemma-2 27b it model, you can set the --model argument accordingly:

local-gemma --model 27b

Local Gemma-2 will automatically find the most performant preset for your hardware, trading-off speed and memory. For more control over generation speed and memory usage, set the --preset argument to one of three available options:

1. exact: match the original results by maximizing accuracy
2. memory: reducing memory through 4-bit quantization
3. memory_extreme: minimizing memory through 4-bit quantization and CPU offload

You can also control the style of the generated text through the --mode flag, one of "chat", "factual" or "creative":

local-gemma --model 9b --preset memory --mode factual

Finally, you can also pipe in other commands, which will be appended to the prompt after a \n separator

ls -la | local-gemma "Describe my files"

To see all available decoding options, call local-gemma -h.

Python Usage

Note

The pipx installation method creates its own Python environment, so you will need to use the pip installation method to use this library in a Python script.

Local Gemma-2 can be run locally through a Python interpreter using the familiar Transformers API. To enable a preset, import the model class from local_gemma and pass the preset argument to from_pretrained. For example, the following code-snippet loads the Gemma-2 9b model with the "memory" preset:

from local_gemma import LocalGemma2ForCausalLM
from transformers import AutoTokenizer

model = LocalGemma2ForCausalLM.from_pretrained("google/gemma-2-9b", preset="memory")
tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-9b")

model_inputs = tokenizer("The cat sat on the mat", return_attention_mask=True, return_tensors="pt")
generated_ids = model.generate(**model_inputs.to(model.device))

decoded_text = tokenizer.batch_decode(generated_ids)

When using an instruction-tuned model (prefixed by -it) for conversational use, prepare the inputs using a chat-template. The following example loads Gemma-2 27b it model using the "auto" preset, which automatically determines the best preset for the device:

from local_gemma import LocalGemma2ForCausalLM
from transformers import AutoTokenizer

model = LocalGemma2ForCausalLM.from_pretrained("google/gemma-2-27b-it", preset="auto")
tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-27b-it")

messages = [
    {"role": "user", "content": "What is your favourite condiment?"},
    {"role": "assistant", "content": "Well, I'm quite partial to a good squeeze of fresh lemon juice. It adds just the right amount of zesty flavour to whatever I'm cooking up in the kitchen!"},
    {"role": "user", "content": "Do you have mayonnaise recipes?"}
]

encodeds = tokenizer.apply_chat_template(messages, return_tensors="pt")
model_inputs = encodeds.to(model.device)

generated_ids = model.generate(model_inputs, max_new_tokens=1000, do_sample=True)
decoded_text = tokenizer.batch_decode(generated_ids)

Presets

Local Gemma-2 provides three presets that trade-off accuracy, speed and memory. The following results highlight this trade-off using Gemma-2 9b with batch size 1 on an 80GB A100 GPU:

Mode	Performance*	Inference Speed (tok/s)	Memory (GB)
exact	73.0	17.2	18.3
memory	72.1	13.8	7.3
memory_extreme	72.1	13.8	7.3

While an 80GB A100 places the full model on the device, only 3.7GB is required with the memory_extreme preset. See the section Preset Details for details.

---

*Zero-shot results averaged over Wino, ARC Easy, Arc Challenge, PIQA, HellaSwag, MMLU, OpenBook QA.

Preset Details

Mode	9b Min Memory (GB)	27b Min Memory (GB)	Weights dtype	CPU Offload
exact	18.3	68.2	bf16	no
memory	7.3	17.0	int4	no
memory_extreme	3.7	4.7	int4	yes

memory_extreme implements CPU offloading through 🤗 Accelerate, reducing memory requirements down to the largest layer in the model (which in this case is the LM head).

Note: Due to logit soft-capping, SDPA and Flash Attention are not compatible with Gemma-2. We are aiming to bring a speed preset that uses torch.compile to improve the inference speed, stay tuned! Any contributions in Transformers are most welcome 🤗

Acknowledgements

Local Gemma-2 is a convenient wrapper around several open-source projects, which we thank explicitly below:

• Transformers for the PyTorch Gemma-2 implementation. Particularly Arthur Zucker for adding the model and the logit soft-capping fixes.
• bitsandbytes for the 4-bit optimization on CUDA.
• quanto for the 4-bit optimization on MPS + CPU.
• Accelerate for the large model loading utilities.

And last but not least, thank you to Google for the pre-trained Gemma-2 checkpoints, all of which you can find on the Hugging Face Hub.