# LlamaToaster

A benchmarking tool that sweeps llama.cpp settings on your own GPU or CPU to find the fastest configuration for a given model, and compares the result against a shared database of other people's hardware and settings.

LlamaToaster points at a model and a machine and sweeps the llama.cpp parameter space -- GPU layer offload, context size, batch size, CPU MoE offload, speculative decoding, and more -- using `llama-bench` and `llama-server`, recording every measurement and producing the configuration to paste into a real `llama-server` or `llama-cli` command line for that specific hardware and build.

An assistant reads a user's own test results and answers what is actually limiting performance and what to try next, instead of leaving the user to interpret a results table. A GGUF index keyed by SHA-256 content hash covers roughly 4 million Hugging Face files, so a renamed or mirrored model file is still recognized as the same file rather than becoming a duplicate entry.

The hosted instance at llamatoaster.com lets a shared results database grow across contributors' hardware, so the assistant can eventually predict good settings for hardware it has already seen; a worker installed with one command only makes outbound HTTPS calls and can also be self-hosted.

Website: https://llamatoaster.com