quantal-ternary: four nights of training, three dead GPUs, and the format that finally earned its third bit

This site does not allow itself to be embedded. You can still read it on the original site — the toolbar below keeps your place in the directory.

The full engineering history of the quantal nightly runs — from 11.34 masked val through a thresholded-ternary rewrite, a GPU that died mid-epoch, a stranger who counted zeros in the published bytes, and the 1.3e-5 parity gate that finally made training equal deployment. Every number, every crash, every fix.

Read on pocoo.vaked.dev ↗