DAILY RESEARCH REPORTBACK TO MASTER SHEET2026-04-20

kernel_fb_ratio · 167 of 692 · 0 warnings
{
  "date": "2026-04-20",
  "fatal": 0,
  "headline_metric": "kernel_fb_ratio",
  "raw_records": 692,
  "render_records": 167,
  "test_rows": 133,
  "warnings": 0
}
16 commits on 2026-04-20
  1. 9317e30c 05:04:26 sota_crawler Prepare anchor-driven dim-576 attention-kernel body downstream scaffold
  2. 9317e30c 05:04:26 SOTA_FINAL Prepare anchor-driven dim-576 attention-kernel body downstream scaffold
  3. c8605bc0 22:53:41 sota_crawler Bless Hybrid B single-GPU winner for 2x
  4. c8605bc0 22:53:41 SOTA_FINAL Bless Hybrid B single-GPU winner for 2x
  5. 186461e1 02:23:26 sota_crawler Record final 2x dim-576 attention-kernel body kernel winner and pod evidence
  6. 186461e1 02:23:26 SOTA_FINAL Record final 2x dim-576 attention-kernel body kernel winner and pod evidence
  7. b8f42d2f 02:42:54 sota_crawler Scaffold donor integration lane for blessed dim-576 attention-kernel body kernel
  8. b8f42d2f 02:42:54 SOTA_FINAL Scaffold donor integration lane for blessed dim-576 attention-kernel body kernel
  9. ef316b98 02:50:04 sota_crawler Add wh_kern_int integration plan and matrix
  10. ef316b98 02:50:04 SOTA_FINAL Add wh_kern_int integration plan and matrix
  11. b4781e19 03:00:46 sota_crawler Define wh_kern_int benchmark contract and fallback ladder
  12. b4781e19 03:00:46 SOTA_FINAL Define wh_kern_int benchmark contract and fallback ladder
  13. b20c73db 03:39:08 sota_crawler Make wh_kern_int benchmark ladder executable
  14. b20c73db 03:39:08 SOTA_FINAL Make wh_kern_int benchmark ladder executable
  15. ff9d6866 03:57:43 sota_crawler Add blessed dim-576 attention-kernel body integration bundle
  16. ff9d6866 03:57:43 SOTA_FINAL Add blessed dim-576 attention-kernel body integration bundle
DATE2026-04-20
DAY034
PRIMARY METRICCUSTOM/FA3 FB RATIO
TESTS RUN133 kernel benchmarks
ARCHITECTURE FAMILYSTAGED ATTENTION-KERNEL BENCH (×130) · DIM-576 ATTENTION-KERNEL BODY (×2) · OTHER (×1)
CLASSIFICATIONKERNEL 100% (133)
ARCHITECTURE RECIPE
STAGED HYBRID BENCHx130

Attention-kernel sweep comparing staged primary/fallback kernels against hybridA/hybridB variants.

KERNEL SHAPEseq_len=2048 · q_heads=8 · kv_heads=4 · head_dim=64 · custom/FA3 fb ratio target
TILE SWEEPDQ=128,64,8,2 · DQ=128,64,8,5 · DKDV=128,64,8,3 · DKDV=128,128,8,3 · QMR sweep=768..2304
MODESstaged_primary_vs_hybridB · staged_fallback_vs_hybridA · g0/g1 GPU lane split · kernel watchdog benchmark
DIM-576 ATTENTION-KERNEL BODYx2

A 12-layer dim-576 model body plus the custom attention-kernel benchmark shape used against FA3.

MODEL BODYnum_layers=12 · model_dim=576 · embedding_dim=576 · matrix_bits=5 · xsa_last_n=11 · parallel_residual_start=8 · model_params=48551136
KERNEL SHAPEbatch=2 · seq_len=2048 · q_heads=8 · kv_heads=4 · head_dim=64 · custom attention path compared directly to FA3
PROMOTED KERNELDQ=128,64,8,5 · DKDV=128,64,8,3 · DKDV_MAXNREG=256 · QMR=1984 for 2x primary · QMR=1664 single-GPU anchor

SECTION 1.0 — TRAJECTORY

CUSTOM/FA3 FB RATIO for 2026-04-20
CUSTOM/FA3 FB RATIO SPREADnminmedianmaxdownload
STAGED ATTENTION-KERNEL BENCH1300.813×0.958×0.995×download
DIM-576 ATTENTION-KERNEL BODY20.933×0.934×0.936×download
OTHER10.933×0.933×0.933×
ALL VALID1330.813×0.957×0.995×

SECTION 2.0 — TEST RUNS

KERNEL TEST RUNS — 133

KERNEL TEST RUNS

133 kernel rows133 measuredmin 0.813× / median 0.957× / max 0.995×H100SOTA_FINAL