KERNEL RESEARCH REPORTBACK TO MASTER SHEETCAMPAIGN

CATEGORYKERNEL
DATES2026-03-18 TO 2026-05-01
WINDOWS2 ACTIVE / 4 TOTAL
ACTIVE DAYS9
DATAPOINTS9
PRIMARY METRICSKERNEL 100% (9)

SECTION 1.0 — BI-WEEKLY TRAJECTORY

WINDOW 03 · 2026-04-15 TO 2026-04-28

KERNEL window 03 kernel_fb_ratio
#DATEMETRICBESTROWSTITLEOPEN
12026-04-16CUSTOM/FA3 FB RATIO0.948×81sweep_fused_bwd_20260416_221808
provenance
category datapoint
category: kernel
window: 03 · 2026-04-15 → 2026-04-28
date: 2026-04-16 open daily report
metric: CUSTOM/FA3 FB RATIO (kernel_fb_ratio)
best valid value: 0.948×
category rows on date: 81 (best value shown above)
row title: sweep_fused_bwd_20260416_221808
source: SOTA_FINAL
public evidence: download file
commit: 5267399 - dim-576 attention-kernel body fwd: BM=192 autotune + bwd autotune-expand + early-exit + bench harness
file path: legs/2026-04-16_whale_bwd_persistent_atomic/evidence/sweep_fused_bwd_20260416_221808.json
22026-04-17CUSTOM/FA3 FB RATIO0.945×74dqmr_128_20260417_194025
provenance
category datapoint
category: kernel
window: 03 · 2026-04-15 → 2026-04-28
date: 2026-04-17 open daily report
metric: CUSTOM/FA3 FB RATIO (kernel_fb_ratio)
best valid value: 0.945×
category rows on date: 74 (best value shown above)
row title: dqmr_128_20260417_194025
source: SOTA_FINAL
commit: 16cf4c5 - Preserve dim-576 attention-kernel body T2048 winner sweep evidence
32026-04-18CUSTOM/FA3 FB RATIO0.748×1282qmr7168_probe_dq_128_64_8_2_dkdvmr_256_20260418_031216
provenance
category datapoint
category: kernel
window: 03 · 2026-04-15 → 2026-04-28
date: 2026-04-18 open daily report
metric: CUSTOM/FA3 FB RATIO (kernel_fb_ratio)
best valid value: 0.748×
category rows on date: 1282 (best value shown above)
row title: qmr7168_probe_dq_128_64_8_2_dkdvmr_256_20260418_031216
source: SOTA_FINAL
commit: 16bdc9e - Preserve dim-576 attention-kernel body T2048 2gpu breakthrough
42026-04-19CUSTOM/FA3 FB RATIO0.794×1289p175_overnight_fallback_vs_hybridB_04_g0_dq128_64_8_2_dkdv128_128_8_3_qmr1920_20260419_100310
provenance
category datapoint
category: kernel
window: 03 · 2026-04-15 → 2026-04-28
date: 2026-04-19 open daily report
metric: CUSTOM/FA3 FB RATIO (kernel_fb_ratio)
best valid value: 0.794×
category rows on date: 1289 (best value shown above)
row title: p175_overnight_fallback_vs_hybridB_04_g0_dq128_64_8_2_dkdv128_128_8_3_qmr1920_20260419_100310
source: SOTA_FINAL
commit: 9317e30 - Prepare anchor-driven dim-576 attention-kernel body downstream scaffold
52026-04-20CUSTOM/FA3 FB RATIO0.813×133p311_staged_fallback_vs_hybridA_11_g1_dq128_64_8_2_dkdv128_64_8_3_qmr928_20260420_015925
provenance
category datapoint
category: kernel
window: 03 · 2026-04-15 → 2026-04-28
date: 2026-04-20 open daily report
metric: CUSTOM/FA3 FB RATIO (kernel_fb_ratio)
best valid value: 0.813×
category rows on date: 133 (best value shown above)
row title: p311_staged_fallback_vs_hybridA_11_g1_dq128_64_8_2_dkdv128_64_8_3_qmr928_20260420_015925
source: SOTA_FINAL
commit: 9317e30 - Prepare anchor-driven dim-576 attention-kernel body downstream scaffold
62026-04-21CUSTOM/FA3 FB RATIO0.834×7blessed_dim-576 attention-kernel body
provenance
category datapoint
category: kernel
window: 03 · 2026-04-15 → 2026-04-28
date: 2026-04-21 open daily report
metric: CUSTOM/FA3 FB RATIO (kernel_fb_ratio)
best valid value: 0.834×
category rows on date: 7 (best value shown above)
row title: blessed_dim-576 attention-kernel body
source: SOTA_FINAL
commit: ff9d686 - Add blessed dim-576 attention-kernel body integration bundle
72026-04-22CUSTOM/FA3 FB RATIO0.834×4blessed_dim-576 attention-kernel body
provenance
category datapoint
category: kernel
window: 03 · 2026-04-15 → 2026-04-28
date: 2026-04-22 open daily report
metric: CUSTOM/FA3 FB RATIO (kernel_fb_ratio)
best valid value: 0.834×
category rows on date: 4 (best value shown above)
row title: blessed_dim-576 attention-kernel body
source: SOTA_FINAL
commit: ff9d686 - Add blessed dim-576 attention-kernel body integration bundle
82026-04-23CUSTOM/FA3 FB RATIO0.834×3blessed_dim-576 attention-kernel body
provenance
category datapoint
category: kernel
window: 03 · 2026-04-15 → 2026-04-28
date: 2026-04-23 open daily report
metric: CUSTOM/FA3 FB RATIO (kernel_fb_ratio)
best valid value: 0.834×
category rows on date: 3 (best value shown above)
row title: blessed_dim-576 attention-kernel body
source: SOTA_FINAL
commit: bdf8741 - blessed_dim-576 attention-kernel body: stabilize hybrid compile dispatch for flat 2x head-to-head

WINDOW 04 · 2026-04-29 TO 2026-05-01

KERNEL window 04 kernel_fb_ratio
#DATEMETRICBESTROWSTITLEOPEN
12026-04-29CUSTOM/FA3 FB RATIO0.933×2fallback_anchor
provenance
category datapoint
category: kernel
window: 04 · 2026-04-29 → 2026-05-01
date: 2026-04-29 open daily report
metric: CUSTOM/FA3 FB RATIO (kernel_fb_ratio)
best valid value: 0.933×
category rows on date: 2 (best value shown above)
row title: fallback_anchor
source: sota_rascal
public evidence: download file
commit: bff195b - publish 11L GQA U-Net body lab snapshot
file path: legs/2026-04-29_mikey_kernel_smoke_gpu0/fallback_anchor.json

SECTION 1.5 — ARCHITECTURE / TEST SURFACES

architecture / test surfacerowsdaysbest CUSTOM/FA3 FB RATIOmedianrangefirstlast
DIM-576 ATTENTION-KERNEL BODY222180.748×0.956×0.748× - 3.722×2026-04-162026-04-23
STAGED ATTENTION-KERNEL BENCH64630.776×0.956×0.776× - 1.369×2026-04-182026-04-20
CRAWLER ABLATION CONTROL SET521.106×1.298×1.106× - 1.580×2026-04-162026-04-17
11L GQA U-NET BODY210.933×0.934×0.933× - 0.936×2026-04-292026-04-29
OTHER110.933×0.933×0.933× - 0.933×2026-04-202026-04-20
ARCHITECTURE RECIPE
THE CRAWLER — DEFINITIONFrugendorff → hybrid

A recursive weight-sharing transformer module. A small set of K unique blocks is reused N times in sequence to produce effective depth K·N from a smaller stored parameter set; the saved budget is reinvested into wider hidden states. Inherited from the Frugendorff line, hardened into a hybrid form for int6 quantization survival.

Origin · 2026-03-19First introduced as NUM_LOOPS in commit “feat: add fractal weight sharing via NUM_LOOPS env var.” When NUM_LOOPS > 1 the model reuses num_layers blocks in a loop instead of the U-Net skip pattern. A learned per-loop position embedding p_loop disambiguates which pass the shared weights are on.
Forward · Frugendorffhi+1 = TransformerBlockθ[i mod K](hi + p_loop⌊i / K⌋)
i ∈ {0, 1, …, K·N − 1}
Origin · ResultProof of concept at 3 × 3, 864-dim hit −7.1% BPB vs the unique-layer baseline at fewer parameters. Gravity losses on early loops converged to ≈ [0.13, 0.13, 0.70] — the model learned to suppress early loop contributions.
Failure modesAt competition budgets, pure recursion surfaced two compounding issues. (1) Shared-weight gradient conflict — the same θ must simultaneously map early → mid and mid → late representations, and short budgets cannot find the compromise. (2) Quantization catastrophe — int6 rounding error compounds multiplicatively across loop iterations (post-quant BPB > 5× pre-quant at 3+ loops). Frugendorff was declared dead 2026-03-27, then rebuilt as a hybrid 2026-03-29.
Final hybrid formA flat stack frames a small shared core. NUM_FLAT_LAYERS standard transformer blocks process the embedded input; NUM_CRAWLER_LAYERS shared-weight blocks then loop CRAWLER_LOOPS times. Each pass k uses a distinct RoPE scale from CRAWLER_LOOP_ROPE_SCALES and is conditioned by a content-derived instruction vector π(h) of dimension INST_DIM. Loop-aware GPTQ, Hessian collection on the shared block, and int8 / int6 sliding-window export keep the model under the 16 MB cap.
Forward · Hybridh0 = StackFlatθf(embed(x))
for k = 1 … K:  hk = CrawlerBlockθc(hk−1, π(hk−1), RoPEk)
y = Head(hK)
Goals(1) effective depth from weight reuse, not from new parameters.  (2) reinvest the saved parameter budget into wider hidden states.  (3) differentiate each pass via loop-position RoPE and a content-derived instruction, so the same θ does different jobs each loop.  (4) survive int6 quantization at the competition cap.  (5) act as the recurrent core of a flat-around-crawler topology, so the model has both standard transformer expressivity and recurrent depth.
DIM-576 ATTENTION-KERNEL BODYx2221

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
STAGED HYBRID BENCHx646

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
11L GQA U-NET BODYx2

11-layer dim-512 GQA U-Net transformer base with XSA, BigramHash, value embedding, and SmearGate.

BODYvocab_size=1024 · num_layers=11 · model_dim=512 · num_heads=8 · num_kv_heads=4 · head_dim=64 · mlp_mult=3.0 · tie_embeddings=1
STRUCTUREencoder_layers=5 · decoder_layers=6 · skip_weights=5 · GQA attention · RMSNorm · relu_sq MLP · Muon matrix optimizer path
FEATURESBigramHash vocab=2048 dim=128 · XSA last_n=11 · rope_dims=16 · VE dim=128 layers=9,10 · SmearGate · logit_softcap=30

SECTION 2.0 — CATEGORY DAILY DATAPOINTS