Frank
8d88b3a0a9
frank-small: int8 KV (per-token, per-head) — doubles the pool
...
fp8 KV is hardware-dead on the 3090 (sm_86, no cvt.e4m3x2), so int8 is
the cheapest quant the TRITON_ATTN backend can run: pure ALU dequant,
same backend, no fp8 tensor cores needed. 289,647 -> ~579k-token pool,
two -> four full-context sessions. Quality verified against the bf16
baseline (needle-in-haystack at 213k + short battery) before the swap.
2026-09-18 23:24:37 -07:00
..
2026-09-16 14:59:23 -07:00
2026-09-16 07:33:38 -07:00
2026-08-28 11:18:41 -07:00
2026-09-16 07:46:23 -07:00
2026-08-28 02:20:06 -07:00
2026-09-16 07:46:23 -07:00
2026-09-16 14:59:23 -07:00
2026-09-16 07:46:23 -07:00
2026-08-13 20:19:41 -07:00
2026-09-01 22:46:00 -07:00
2026-09-17 23:32:02 -07:00
2026-09-18 21:06:46 -07:00
2026-09-16 07:43:33 -07:00
2026-09-16 07:33:38 -07:00
2026-08-13 20:19:41 -07:00
2026-09-18 21:06:46 -07:00
2026-09-18 15:25:55 -07:00
2026-09-18 23:24:37 -07:00
2026-09-04 12:54:28 -07:00
2026-08-22 15:49:44 -07:00
2026-09-16 19:28:34 -07:00
2026-08-29 15:11:18 -07:00
2026-08-23 16:49:33 -07:00
2026-09-16 13:59:49 -07:00
2026-08-25 20:13:33 -07:00
2026-09-17 23:25:50 -07:00
2026-08-22 15:49:44 -07:00
2026-09-03 16:15:43 -07:00
2026-09-12 10:09:12 -07:00
2026-08-25 15:38:57 -07:00
2026-09-17 18:28:20 -07:00
2026-09-03 16:15:43 -07:00
2026-08-26 20:08:30 -07:00
2026-08-13 20:19:41 -07:00
2026-09-16 06:59:44 -07:00
2026-08-25 20:13:33 -07:00
2026-08-19 00:00:42 -07:00
2026-09-01 14:40:10 -07:00