The open‑source llama.cpp library released version b12364 on 2026-11-13. It now logs a warning if the --spec-draft-p-min flag is set while using probabilistic draft sampling, suggesting p_min be set to zero. The change skips the p_min cutoff for MTP and DFlash2 drafts, which previously could drop tokens the target model would accept. The warning is only shown for drafters that actually use p_min, avoiding noise for n‑gram and DSpark drafters.
Why it matters
Developers can avoid silent token loss when fine‑tuning draft models, leading to more reliable generation.
