tope1129/cas-smo-v33

cas-smo-v32

Fork of garuda-labs/gen-04f5171856b6 aimed at more training series per wall-clock hour.

The parent always emits length 4096. Under stream_cpu the trainer stops at a fixed token budget, so mean length sets how many series fit in one hour. Pool 2026-08-09-block-8805600 has mean length ≈1552 (p25=512, p90≈4057).

Knob	Parent	cas-smo-v32
p00 (min length)	4096	512
p01 (max length)	4096	4096
p02 (prefetch)	2	4
Expected mean L	4096	~2304
Series / token-hour	1×	~1.8×

Also: each generate batch builds at the batch max length (not always p01), so short draws cost less CPU and starve the GPU less.

Family mix / recipes unchanged from the parent. Confirm with:

cascade score ./cas-smo-v32 --pool-dir ./dataset/snapshots/2026-08-09-block-8805600 --train-hours 1 --device cuda