Run LLMs locally. A clojure wrapper for llama.cpp.

If you're just looking for a model to try things out, try the 0.5Gb qwen instruct model. Make sure to check the link for important info like license and use policy.

mkdir -p models
# Download 0.5Gb model to models/ directory
(cd models && curl -L -O 'https://huggingface.co/Qwen/Qwen2-0.5B-Instruct-GGUF/resolve/main/qwen2-0_5b-instruct-q4_0.gguf')
# mvn-llama alias pulls precompiled llama.cpp libs from maven
clojure -M:mvn-llama -m com.phronemophobic.llama "models/qwen2-0_5b-instruct-q4_0.gguf" "what is 2+2?"

Getting Started
Intro to Running LLMs Locally
API Reference Docs

For llama.clj with required native dependencies:

com.phronemophobic/llama-clj-combined {:mvn/version "0.8.5"}

For llama.clj only (see below for various alternatives for specifying native dependencies):

com.phronemophobic/llama-clj {:mvn/version "0.8.5"}

llama.clj relies on the excellent llama.cpp library.

The llama.cpp shared library can either be compiled locally or can be included as a standalone maven dependency.

Note: The support for ggml models has been deprecated. However, no breaking changes are planned.

The easiest method is to include the corresponding native dependency for your platform (including multiple is fine, but will increase the size of your dependencies). See the mvn-llama alias for an example.

;; gguf dependencies
com.phronemophobic.cljonda/llama-cpp-gguf-linux-x86-64 {:mvn/version "b3040"}
com.phronemophobic.cljonda/llama-cpp-gguf-darwin-aarch64 {:mvn/version "b3040"}
com.phronemophobic.cljonda/llama-cpp-gguf-darwin-x86-64 {:mvn/version "b3040"}

;; ggml dependencies
;; Note: deprecated
com.phronemophobic.cljonda/llama-cpp-darwin-aarch64 {:mvn/version "6e88a462d7d2d281e33f35c3c41df785ef633bc1"}
com.phronemophobic.cljonda/llama-cpp-darwin-x86-64 {:mvn/version "6e88a462d7d2d281e33f35c3c41df785ef633bc1"}
com.phronemophobic.cljonda/llama-cpp-linux-x86-64 {:mvn/version "6e88a462d7d2d281e33f35c3c41df785ef633bc1"}

Clone https://github.com/ggerganov/llama.cpp and follow the instructions for building. Make sure to include the shared library options.

Note: The llama.cpp ffi bindings are based on the b3040 release for gguf models and the 4329d1acb01c353803a54733b8eef9d93d0b84b2 git commit for ggml models. Future versions of llama.cpp might not be compatible if breaking changes are made. TODO: include instructions for updating ffi bindings.

For Example:

git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
git checkout b3040
mkdir build
cd build
cmake -DBUILD_SHARED_LIBS=ON ..
cmake --build . --config Release

Next, include an alias that includes the path to the directory where the shared library is located:

;; in aliases
;; add jvm opt for local llama build.
:local-llama {:jvm-opts ["-Djna.library.path=/path/to/llama.cpp/build/"]}

It's possible to use both ggml and gguf models in the same process (ie. "dual wielding"). The trick is to treat the older ggml llama.cpp version and the newer gguf llama.cpp versions as separate libraries. Each shared library must have a unique name. If using only one of the ggml or gguf formats is required, then using the libllama.(so,dylib) is sufficient. For dual wielding, the ggml version should be called libllama.(so,dylib) and the gguf version should be renamed to libllama-gguf.(so,dylib). Further, the soname of the gguf version must also be updated. For example:

Linux

mv libllama.so libllama-gguf.so
sudo apt-get install patchelf
patchelf --set-soname libllama-gguf.so libllama-gguf.so

Mac OSX

mv libllama.dylib libllama-gguf.dylib
install_name_tool -id libllama-gguf.dylib libllama-gguf.dylib

For more complete information about the models that llama.clj can work with, refer to the llama.cpp readme.

clojure -M -m com.phronemophobic.llama <path-to-model> <prompt>

Example:

mkdir -p models
# Download 0.5Gb model to models/ directory
(cd models && curl -L -O 'https://huggingface.co/Qwen/Qwen2-0.5B-Instruct-GGUF/resolve/main/qwen2-0_5b-instruct-q4_0.gguf')
clojure -M:mvn-llama -m com.phronemophobic.llama "models/qwen2-0_5b-instruct-q4_0.gguf" "what is 2+2?"

For gpu support on linux, cuda must be installed. The instructions for cuda installation can be found in nvidia's documentation.

Make sure to restart and follow the post installation instructions so that the cuda development tools like nvcc are available on the path.

Currently, pre-compiled binaries of llama.cpp with cuBLAS support are not available. The llama.cpp native dependencies must be compiled locally with -DLLAMA_CUBLAS=ON as argument. Something like:

git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
git checkout b3040
mkdir build
cd build
cmake -DBUILD_SHARED_LIBS=ON -DLLAMA_CUBLAS=ON ..
cmake --build . --config Release

• https://github.com/ggerganov/llama.cpp#cublas
• https://docs.nvidia.com/cuda/cuda-installation-guide-linux/contents.html

If you see a feature you're interested in or have a feature you want, please file an issue to help me prioritize future work.

• Add convenience namespace for obtaining and using models.
• Provide precompiled native dependencies for Windows
• API for sampling using a grammar
• Expose API for sessions
• Expose API for Sequence ids
• Expose API for more special tokens.
• Expose API for generating embeddings
• Expose API for chat templates
• Expose logging API
• Expose API for reading model metadata
• Windows support
• Update llama.cpp to support gguf format
• More docs!
  • Reference docs
  • Intro Guide to LLMs.

The MIT License (MIT)

Copyright © 2024 Adrian Smith

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.