OhLLM

Pure C++ LLM inference engine. No third party dependencies.

• under construction, for education and learning purpose only (currently). Hope it will be eventually evolved to be useful in production.*

License: MIT

Build

./scripts/build.sh [--debug] [--test] [--nolog] [--clear]

# argments:
#   --debug: build debug version for debugging
#   --test: run all unit tests
#   --nolog: build without log
#   --clear: clear and rebuild

Usage

Benchmark

Features

This project will mainly focus on LLM inference on single machine with/without accelerator (Nvidia GPU, apple MPS, ...).

As the project proceeds:

• support cpu inference in FP32/FP16, no vectorization intrinsics, compatible with GGUF format
• support mainstream LLM models, like llama, phi, qwen, gemma and ...
• support cpu inference in FP32/FP16 plus vectorization intrinsics
• support cpu inference with quantization of Q8_0, Q4_0 and Q4_1
• support streaming output
• support nvidia gpu inference
• support nvidia gpu inference with cpu offloading
• support apple mps

Acknowledgements

This project is greatly inspired by ggml, fastllm and llama.cpp. Thanks a lot to all of the contributors.

Please buy me a ☕ if you find this project useful.