Run PyTorch LLMs locally on servers, desktop and mobile

torchchat is a small codebase showcasing the ability to run large language models (LLMs) seamlessly. With torchchat, you can run LLMs using Python, within your own (C/C++) application (desktop or server) and on iOS and Android.

• Run models via PyTorch / Python
  • Chat
  • Generate
  • Run chat in the Browser
• Run models on desktop/server without python
  • Use AOT Inductor for faster execution
  • Running in c++ using the runner
• Run models on mobile
  • Deploy and run on iOS
  • Deploy and run on Android
• Evaluate a model

• Command line interaction with popular LLMs such as Llama 3, Llama 2, Stories, Mistral and more
• PyTorch-native execution with performance
• Supports popular hardware and OS
  • Linux (x86)
  • Mac OS (M1/M2/M3)
  • Android (Devices that support XNNPACK)
  • iOS 17+ (iPhone 13 Pro+)
• Multiple data types including: float32, float16, bfloat16
• Multiple quantization schemes
• Multiple execution modes including: Python (Eager, Compile) or Native (AOT Inductor (AOTI), ExecuTorch)

The following steps require that you have Python 3.10 installed.

# get the code
git clone https://github.com/pytorch/torchchat.git
cd torchchat

# set up a virtual environment
python3 -m venv .venv
source .venv/bin/activate

# install dependencies
./install_requirements.sh

The interfaces of torchchat are leveraged through Python Commands and Native Runners. While the Python Commands are enumerable in the --help menu, the latter are explored in their respective sections.

python3 torchchat.py --help

# Output
usage: torchchat [-h] {chat,browser,generate,export,eval,download,list,remove,where,server} ...

positional arguments:
  {chat,browser,generate,export,eval,download,list,remove,where,server}
                        The specific command to run
    chat                Chat interactively with a model via the CLI
    generate            Generate responses from a model given a prompt
    browser             Chat interactively with a model in a locally hosted browser
    export              Export a model artifact to AOT Inductor or ExecuTorch
    download            Download model artifacts
    list                List all supported models
    remove              Remove downloaded model artifacts
    where               Return directory containing downloaded model artifacts
    server              [WIP] Starts a locally hosted REST server for model interaction
    eval                Evaluate a model via lm-eval

options:
  -h, --help            show this help message and exit

Python Inference (chat, generate, browser, server)

• These commands represent different flavors of performing model inference in a Python enviroment.
• Models are constructed either from CLI args or from loading exported artifacts.

Exporting (export)

• This command generates model artifacts that are consumed by Python Inference or Native Runners.
• More information is provided in the AOT Inductor and ExecuTorch sections.

Inventory Management (download, list, remove, where)

• These commands are used to manage and download models.
• More information is provided in the Download Weights section.

Evaluation (eval)

• This command test model fidelity via EleutherAI's lm_evaluation_harness.
• More information is provided in the Evaluation section.

Most models use Hugging Face as the distribution channel, so you will need to create a Hugging Face account. Create a Hugging Face user access token as documented here with the write role.

Log into Hugging Face:

huggingface-cli login

Once this is done, torchchat will be able to download model artifacts from Hugging Face.

python3 torchchat.py download llama3.1

python3 torchchat.py list

python3 torchchat.py list

python3 torchchat.py remove llama3.1

The simplest way to run a model in PyTorch is via eager execution. This is the default execution mode for both PyTorch and torchchat. It performs inference without creating exporting artifacts or using a separate runner.

The model used for inference can also be configured and tailored to specific needs (compilation, quantization, etc.). See the customization guide for the options supported by torchchat.

This mode allows you to chat with an LLM in an interactive fashion.

python3 torchchat.py chat llama3.1

This mode generates text based on an input prompt.

python3 torchchat.py generate llama3.1 --prompt "write me a story about a boy and his bear"

This mode allows you to chat with the model using a UI in your browser Running the command automatically open a tab in your browser.

streamlit run torchchat.py -- browser llama3.1

Note: This feature is still a work in progress and not all endpoints are working

python3 torchchat.py server llama3.1

curl http://127.0.0.1:5000/chat \
  -H "Content-Type: application/json" \
  -d '{
    "model": "llama3.1",
    "stream": "true",
    "messages": [
      {
        "role": "system",
        "content": "You are a helpful assistant."
      },
      {
        "role": "user",
        "content": "Hello!"
      }
    ]
  }'

{"response":" I'm a software developer with a passion for building innovative and user-friendly applications. I have experience in developing web and mobile applications using various technologies such as Java, Python, and JavaScript. I'm always looking for new challenges and opportunities to learn and grow as a developer.\n\nIn my free time, I enjoy reading books on computer science and programming, as well as experimenting with new technologies and techniques. I'm also interested in machine learning and artificial intelligence, and I'm always looking for ways to apply these concepts to real-world problems.\n\nI'm excited to be a part of the developer community and to have the opportunity to share my knowledge and experience with others. I'm always happy to help with any questions or problems you may have, and I'm looking forward to learning from you as well.\n\nThank you for visiting my profile! I hope you find my information helpful and interesting. If you have any questions or would like to discuss any topics, please feel free to reach out to me. I"}

AOTI compiles models before execution for faster inference. The process creates a DSO model (represented by a file with extension .so) that is then loaded for inference. This can be done with both Python and C++ enviroments.

The following example exports and executes the Llama3.1 8B Instruct model. The first command compiles and performs the actual export.

python3 torchchat.py export llama3.1 --output-dso-path exportedModels/llama3.1.so

For more details on quantization and what settings to use for your use case visit our customization guide.

To run in a python enviroment, use the generate subcommand like before, but include the dso file.

python3 torchchat.py generate llama3.1 --dso-path exportedModels/llama3.1.so --prompt "Hello my name is"

Note: Depending on which accelerator is used to generate the .dso file, the command may need the device specified: --device (cuda | cpu).

To run in a C++ enviroment, we need to build the runner binary.

scripts/build_native.sh aoti

Then run the compiled executable, with the exported DSO from earlier.

cmake-out/aoti_run exportedModels/llama3.1.so -z `python3 torchchat.py where llama3.1`/tokenizer.model -l 3 -i "Once upon a time"

Note: Depending on which accelerator is used to generate the .dso file, the runner may need the device specified: -d (CUDA | CPU).

ExecuTorch enables you to optimize your model for execution on a mobile or embedded device.

Before running any commands in torchchat that require ExecuTorch, you must first install ExecuTorch.

To install ExecuTorch, run the following commands. This will download the ExecuTorch repo to ./et-build/src and install various ExecuTorch libraries to ./et-build/install.

export TORCHCHAT_ROOT=${PWD}
./scripts/install_et.sh

Similar to AOTI, to deploy onto device, we first export the PTE artifact, then we load the artifact for inference.

The following example uses the Llama3.1 8B Instruct model.

# Export
python3 torchchat.py export llama3.1 --quantize config/data/mobile.json --output-pte-path llama3.1.pte

For more details on quantization and what settings to use for your use case visit our customization guide.

While ExecuTorch does not focus on desktop inference, it is capable of doing so. This is handy for testing out PTE models without sending them to a physical device.

Specifically there are 2 ways of doing so: Pure Python and via a Runner

# Execute
python3 torchchat.py generate llama3.1 --device cpu --pte-path llama3.1.pte --prompt "Hello my name is"

scripts/build_native.sh et

cmake-out/et_run llama3.1.pte -z `python3 torchchat.py where llama3.1`/tokenizer.model -l 3 -i "Once upon a time"

The following assumes you've completed the steps for Setting up ExecuTorch.

The following assumes you've completed the steps for Setting up ExecuTorch.

export TORCHCHAT_ROOT=$(pwd)
export USE_TIKTOKEN=ON # Set this only for tiktoken tokenizer
sh scripts/android_example.sh

Note: This feature is still a work in progress and not all features are working

Uses the lm_eval library to evaluate model accuracy on a variety of tasks. Defaults to wikitext and can be manually controlled using the tasks and limit args. See Evaluation

Examples

Eager mode:

python3 torchchat.py eval llama3.1 --dtype fp32 --limit 5

To test the perplexity for a lowered or quantized model, pass it in the same way you would to generate:

python3 torchchat.py eval llama3.1 --pte-path llama3.1.pte --limit 5

The following models are supported by torchchat and have associated aliases.

While we describe how to use torchchat using the popular llama3 model, you can perform the example commands with any of these models.

torchchat embodies PyTorch’s design philosophy details, especially "usability over everything else".

torchchat is a native-PyTorch library. While we provide integrations with the surrounding ecosystem (eg: Hugging Face models, etc), all of the core functionality is written in PyTorch.

torchchat is designed to be easy to understand, use and extend.

• Composition over implementation inheritance - layers of inheritance for code re-use makes the code hard to read and extend
• No training frameworks - explicitly outlining the training logic makes it easy to extend for custom use cases
• Code duplication is preferred over unnecessary abstractions
• Modular building blocks over monolithic components

torchchat provides well-tested components with a high-bar on correctness. We provide

• Extensive unit-tests to ensure things operate as they should

We really value our community and the contributions made by our wonderful users. We'll use this section to call out some of these contributions! If you'd like to help out as well, please see the CONTRIBUTING guide.

CERTIFICATE_VERIFY_FAILED Run pip install --upgrade certifi.

Access to model is restricted and you are not in the authorized list Some models require an additional step to access. Follow the link provided in the error to get access.

Installing ET Fails If ./scripts/install_et.sh fails with an error like Building wheel for executorch (pyproject.toml) did not run successfully It's possible that it's linking to an older version of pytorch installed some other way like via homebrew. You can break the link by uninstalling other versions such as brew uninstall pytorch Note: You may break something that depends on this, so be aware.

Please include the exact command you ran and the output of that command. Also, run this script and include the output saved to system_info.txt so that we can better debug your issue.

(echo "Operating System Information"; uname -a; echo ""; cat /etc/os-release; echo ""; echo "Python Version"; python --version || python3 --version; echo ""; echo "PIP Version"; pip --version || pip3 --version; echo ""; echo "Installed Packages"; pip freeze || pip3 freeze; echo ""; echo "PyTorch Version"; python -c "import torch; print(torch.__version__)" || python3 -c "import torch; print(torch.__version__)"; echo ""; echo "Collection Complete") > system_info.txt

The torchchat Repository Content is provided without any guarantees about performance or compatibility. In particular, torchchat makes available model architectures written in Python for PyTorch that may not perform in the same manner or meet the same standards as the original versions of those models. When using the torchchat Repository Content, including any model architectures, you are solely responsible for determining the appropriateness of using or redistributing the torchchat Repository Content and assume any risks associated with your use of the torchchat Repository Content or any models, outputs, or results, both alone and in combination with any other technologies. Additionally, you may have other legal obligations that govern your use of other content, such as the terms of service for third-party models, weights, data, or other technologies, and you are solely responsible for complying with all such obligations.

Thank you to the community for all the awesome libraries and tools you've built around local LLM inference.

• Georgi Gerganov and his GGML project shining a spotlight on community-based enablement and inspiring so many other projects.

• Andrej Karpathy and his llama2.c project. So many great (and simple!) ideas in llama2.c that we have directly adopted (both ideas and code) from his repo. You can never go wrong by following Andrej's work.

• Michael Gschwind, Bert Maher, Scott Wolchok, Bin Bao, Chen Yang, Huamin Li and Mu-Chu Li who built the first version of nanogpt (DSOGPT) with AOT Inductor proving that AOTI can be used to build efficient LLMs, and DSOs are a viable distribution format for models. nanoGPT.

• Bert Maher and his llama2.so, which built on Andrej's llama2.c and on DSOGPT to close the loop on Llama models with AOTInductor.

• Christian Puhrsch, Horace He, Joe Isaacson and many more for their many contributions in Accelerating GenAI models in the "Anything, Fast!" pytorch.org blogs, and, in particular, Horace He for GPT, Fast!, which we have directly adopted (both ideas and code) from his repo.

torchchat is released under the BSD 3 license. (Additional code in this distribution is covered by the MIT and Apache Open Source licenses.) However you may have other legal obligations that govern your use of content, such as the terms of service for third-party models.