alleyxd/quantpicker v0.1.0 is a command-line tool that directly measures the speed and quality loss of GGUF quantization candidates on a GPU so you can pick which one to run. The latest release is v0.1.0, and the repository README presents it as a selection tool based on measurements on your own GPU rather than on estimates.
The tool uses llama-bench to measure the generation and input processing speed of each candidate, and compares output distribution differences through Kullback–Leibler (KL) divergence against a reference model. KL divergence is a value that indicates how different two probability distributions are. It picks the fastest quantization while also considering whether it fits in memory and whether the allowed quality loss is respected. Selection criteria include not only speed but also how much the output differs from the reference model. It is designed so you can adjust the quality tolerance to decide how much output difference to accept. The project is written in Rust and uses the MIT license.
However, the measurement figures in the documentation are relative to ordinary English test sentences. The README advises specifying your own sentences for tasks where code or other languages matter. The published measurement figures and examples are results presented by the repository author, not independent verification. So the selection results are better regarded as a reference value for narrowing down candidates under those sentences and GPU conditions, rather than a quality guarantee that applies to all tasks.