Picchio is a single Python file designed to measure the performance of local Large Language Models (LLMs) using `llama.cpp` and `ollama`. It addresses common issues with GPU speed claims by analyzing three key "tok/s lanes": prefill (prompt processing), decode (answer generation), and wallclock (overall time). The tool detects silent CPU fallback and mislabeled quantization types, providing a more accurate picture of LLM performance.

Picchio works by splitting prefill, decode, and wallclock times, and comparing the LLM engine's logs against the operating system's GPU utilization metrics. It then reports whether the GPU was actually utilized and explains why. Installation is straightforward via curl, and the script automatically detects model locations for `ollama`, the current directory, and common cache locations.

Key features include diagnosing `llama.cpp` behavior, measuring `ollama` models, and analyzing running LLM servers. It can also wrap custom commands to warn about GPU offload issues. Picchio provides detailed breakdowns of quantization, revealing discrepancies between advertised and actual bits per weight. The tool offers commands like `id` to inspect quantization details and `watch` to monitor GPU activity in real-time. Exit codes provide clear indications of performance issues, such as silent CPU fallback (exit 4) or conflicting evidence (exit 5).