Dev Radar
Support
LiveUpdated 2026-09-19 19:24 UTC

your rtx 3060 was running bonsai 2 at 26 tok/s this morning and now does 40 tok/s, i…

your rtx 3060 was running bonsai 2 at 26 tok/s this morning and now does 40 tok/s, i spent the day inside the prismml…

This is a dev post classified by Jev as AI dev tools (a tool drop), kept by the Dev Radar because it carries real work, not commentary.

your rtx 3060 was running bonsai 2 at 26 tok/s this morning and now does 40 tok/s, i spent the day inside the prismml llama.cpp fork so your rtx 3060 can rejoice. the kernel that reads the weights was leaving two thirds of the card's threads idle on the biggest matrices. i fixed it, tested it against the cpu reference, the outputs are bit for bit the same, and both PRs are open on their repo. then i grafted the qwen 3.8 mtp head back onto the file, the one they dropped in the export, and with a batch invariance switch speculative decoding is lossless: 50 tok/s, twice what the stock fork does

Posted by Sudo su (36.2k followers) 2 h ago · 69 likes · 5.9k views · view the original post on X. Kept by the Dev Radar as AI dev tools.

More dev work like this

Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 13.5k posts from 4.7k X accounts over the last 21 days, 1.5k tools, 12 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-19 19:24 UTC. Full method.