pi-prefill

pi-prefill - live prefill progress for local models Hey, I wanted to share my new pi extension: pi-prefill It puts a live prefill progress bar in pi's loading line when you run a local model. A long prefill looks exactly like a hang. Now you can watch it: i.redd.it 52%. → share of uncached prompt tokens processed 14.1k/27.3k → processed / total prompt tokens gen in ~14s. → estimated time until the first token 3.4k tok/s. → instantaneous prefill rate (optional view) Example: prefill [██████░░░░░░] 52% 14.1k/27.3k gen in ~14s It reads the progress chunks your own server sends. No host scanning and no guessing: if the server reports progress, the bar shows. If it does not, pi's normal loading line stays untouched. Works today with llama.cpp, ExLlamaV3, unsloth and Strata. Install: pi install npm:pi-prefill Repo: eliaskg/pi-prefill npmjs.com Would love feedback from anyone running a local server.