llm is an ecosystem of Rust libraries for working with large language models -
it's built on top of the fast, efficient GGML library for
machine learning.
Image by @darthdeus, using Stable Diffusion
The primary entrypoint for developers is
the llm crate, which wraps llm-base and
the supported model crates.
Documentation for released version is available on
Docs.rs.
For end-users, there is a CLI application,
llm-cli, which provides a convenient interface for
interacting with supported models. Text generation can be done as a
one-off based on a prompt, or interactively, through
REPL or chat modes. The CLI can also be
used to serialize (print) decoded models,
quantize GGML files, or compute the
perplexity of a model. It
can be downloaded from
the latest GitHub release or by
installing it from crates.io.
llm is powered by the ggml tensor
library, and aims to bring the robustness and ease of use of Rust to the world
of large language models. At present, inference is only on the CPU, but we hope
to support GPU inference in the future through alternate backends.
Currently, the following models are supported:
- BLOOM
- GPT-2
- GPT-J
- GPT-NeoX (includes StableLM, RedPajama, and Dolly 2.0)
- LLaMA (includes Alpaca, Vicuna, Koala, GPT4All, and Wizard)
- MPT
See getting models for more information on how to download supported models.
This project depends on Rust v1.65.0 or above and a modern C toolchain.
The llm crate exports llm-base and the model crates (e.g. bloom, gpt2
llama).
Add llm to your project by listing it as a dependency in Cargo.toml. To use
the version of llm you see in the main branch of this repository, add it
from GitHub (although keep in mind this is pre-release software):
[dependencies]
llm = { git = "https://github.com/rustformers/llm" , branch = "main" }To use a released version, add it from crates.io by specifying the desired version:
[dependencies]
llm = "0.1"By default, llm builds with support for remotely fetching the tokenizer from Hugging Face's model hub.
To disable this, disable the default features for the crate, and turn on the models feature to get llm
without the tokenizer:
[dependencies]
llm = { version = "0.1", default-features = false, features = ["models"] }NOTE: To improve debug performance, exclude the transitive ggml-sys
dependency from being built in debug mode:
[profile.dev.package.ggml-sys]
opt-level = 3The llm library is engineered to take advantage of hardware accelerators such as cuda and metal for optimized performance.
To enable llm to harness these accelerators, some preliminary configuration steps are necessary, which vary based on your operating system. For comprehensive guidance, please refer to Acceleration Support in our documentation.
Bindings for this library are available in the following languages:
- Python: LLukas22/llm-rs-python
- Node: Atome-FE/llama-node
The easiest way to get started with llm-cli is to download a pre-built
executable from a released
version of llm, but the releases are currently out of date and we recommend
you install from source instead.
To install the main branch of llm with the most recent features to your Cargo bin
directory, which rustup is likely to have added to your PATH, run:
cargo install --git https://github.com/rustformers/llm llm-cliThe CLI application can then be run through llm. See also features and
acceleration support to turn features on as required.
Note that GPU support (CUDA, OpenCL, Metal) will not work unless you build with the relevant feature.
Note that the currently published version is out of date and does not include support for the most recent models. We currently recommend that you install from source.
To install the most recently released version of llm to your Cargo bin
directory, which rustup is likely to have added to your PATH, run:
cargo install llm-cliThe CLI application can then be run through llm. See also features
to turn features on as required.
By default, llm builds with support for remotely fetching the tokenizer from Hugging Face's model hub.
This adds a dependency on your system's native SSL stack, which may not be available on all systems.
To disable this, disable the default features for the build:
cargo build --release --no-default-featuresTo enable hardware acceleration, see Acceleration Support for Building section, which is also applicable to the CLI.
GGML models are easy to acquire. They are primarily located on Hugging Face (see From Hugging Face), but can be obtained from elsewhere.
Models are distributed as single files, and do not need any additional files to be downloaded. However, they are quantized with different levels of precision, so you will need to choose a quantization level that is appropriate for your application.
Additionally, we support Hugging Face tokenizers to improve the quality of
tokenization. These are separate files (tokenizer.json) that can be used
with the CLI using the -v or -r flags, or with the llm crate by
using the appropriate TokenizerSource enum variant.
For a list of models that have been tested, see the known-good models.
Certain older GGML formats are not supported by this project, but the goal is to maintain feature parity with the upstream GGML project. For problems relating to loading models, or requesting support for supported GGML model types, please open an Issue.
Hugging Face π€ is a leader in open-source machine learning and hosts hundreds of GGML models. Search for GGML models on Hugging Face π€.
This Reddit community maintains a wiki related to GGML models, including well organized lists of links for acquiring GGML models (mostly from Hugging Face π€).
Once the llm executable has been built or is in a $PATH directory, try
running it. Here's an example that uses the open-source
RedPajama
language model:
llm infer -a gptneox -m RedPajama-INCITE-Base-3B-v1-q4_0.bin -p "Rust is a cool programming language because" -r togethercomputer/RedPajama-INCITE-Base-3B-v1In the example above, the first two arguments specify the model architecture and
command, respectively. The required -m argument specifies the local path to
the model, and the required -p argument specifies the evaluation prompt. The
optional -r argument is used to load the model's tokenizer from a remote
Hugging Face π€ repository, which will typically improve results when compared
to loading the tokenizer from the model file itself; there is also an optional
-v argument that can be used to specify the path to a local tokenizer file.
For more information about the llm CLI, use the --help parameter.
There is also a simple inference example that is helpful for debugging:
cargo run --release --example inference gptneox RedPajama-INCITE-Base-3B-v1-q4_0.bin -r $OPTIONAL_VOCAB_REPO -p $OPTIONAL_PROMPTYes, but certain fine-tuned models (e.g.
Alpaca,
Vicuna,
Pygmalion) are more suited to chat use-cases than
so-called "base models". Here's an example of using the llm CLI in REPL
(Read-Evaluate-Print Loop) mode with an Alpaca model - note that the
provided prompt format is tailored to the model
that is being used:
llm repl -a llama -m ggml-alpaca-7b-q4.bin -f utils/prompts/alpaca.txtThere is also a Vicuna chat example that demonstrates how to create a custom chatbot:
cargo run --release --example vicuna-chat llama ggml-vicuna-7b-q4.binSessions can be loaded (--load-session) or saved (--save-session) to file.
To automatically load and save the same session, use --persist-session. This
can be used to cache prompts to reduce load time, too.
llm can produce a q4_0- or
q4_1-quantized model from an
f16-quantized GGML model
cargo run --release quantize -a $MODEL_ARCHITECTURE $MODEL_IN $MODEL_OUT {q4_0,q4_1}The llm Dockerfile is in the utils directory; the
NixOS flake manifest and lockfile are in the project root.
GitHub Issues and Discussions are welcome, or come chat on Discord!
Absolutely! Please see the contributing guide.
- llmcord: Discord bot for generating
messages using
llm. - local.ai: Desktop app for hosting an
inference API on your local machine using
llm. - secondbrain: Desktop app to download and run LLMs locally in your computer using
llm. - floneum: A graph editor for local AI workflows.
- llm-chain: Build chains in large language models for text summarization and completion of more complex tasks
