One-command honest local-inference battery: /tokenize-true prompt sizing, ambient A/B noise control, spec-dec disclosure, machine-readable JSON
python benchmark decode zero-dependency measurement apu server-benchmark amd-gpu prefill local-inference llama-cpp speculative-decoding strix-halo honest-numbers
-
Updated
Sep 17, 2026 - Python