Today
-
Survive a machine change
Yesterday
-
Run r3: recipe comparison, keep-the-better rule, extra measurements
-
Rebuild a half-present Python install; run r2 from scratch
-
Second training run: append data, start from the first run's head; upload missing draft weights
-
Record memory, disk and GPU use with every log upload; use the built-in sampler
-
Keep engine and compiler caches on the work disk
-
Engine-based data generation and an unattended GPU driver
3 days ago
-
Draft-model pipeline for the Speculative Decoding Championship