/benchmark

Racing Llama Benchmark (rlb)

Primary LanguagePythonMIT LicenseMIT

rlb

The Racing Llama Benchmark (rlb) is designed to provide consistent LLM benchmarking across Linux, MacOS, and Windows.

NOTE: This project is in Alpha state and is currently being developed for MacOS. Linux and Windows support will be added next.

TODO

  • Support Running llama.cpp against multiple parameter and quantization types
  • Graph output
  • Comparing thread counts
  • Improved benchmark prompts
  • Memory benchmarking
  • CPU benchmarking
  • Possible HPL-MxP support
  • Possible jitter measurements
  • JSON output
  • Versions in output, such as llama.cpp, python-llama-cpp
  • Linux support
  • Windows Support