A lightweight SPEED-Bench client for benchmarking an already-running llama-server through its OpenAI-compatible API. It is primarily meant to evaluate speculative decoding (draft model, n-gram, MTP, ...
Some results have been hidden because they may be inaccessible to you
