Kernel work: DIM-576 ATTENTION-KERNEL BODY (x2214): A 12-layer dim-576 model body plus the custom attention-kernel benchmark shape used against FA3. Specs: num_layers=12; model_dim=576; embedding_dim=576; head_dim=64; matrix_bits=5. STAGED HYBRID BENCH (x646): Attention-kernel sweep comparing staged primary/fallback kernels against hybridA/hybridB variants. Specs: head_dim=64; seq_len=2048; q_heads=8; kv_heads=4; custom/FA3 fb ratio target. CRAWLER ABLATION CONTROL SET (x5): architecture family detected in measured rows; exact flags are preserved in the row details.
Experimental work: CRAWLER-SAFE TRAIN/EVAL SURFACE (x1): Crawler-safe training/eval surface around a widened shared-crawler stack with final int6 and int8 checks. Specs: model_dim=512; vocab_size=1024; num_heads=8; num_kv_heads=4; BigramHash.
Neural work: 12L GQA MIXED-BIT TOWER (x29): 12-layer GQA tower: 8/4 GQA, RoPE-16, Bigram 2048, XSA on the last 11 layers, and mixed-bit packaging. Specs: num_layers=12; num_heads=8; num_kv_heads=4; rope_dims=16; eval_seq_len=2048. LEADERBOARD TRANSFORMER VARIANTS (x4): architecture family detected in measured rows; exact flags are preserved in the row details.
Week 05 covers 7 daily reports from 2026-04-15 through 2026-04-21. Measured surfaces: 7 daily datapoints across CUSTOM/FA3 FB RATIO (×6) · VAL_BPB (×1).
Techniques to copy/reference this week: custom attention path compared directly to FA3; custom/FA3 fb ratio target; attention int5; mlp int6; aux int6; embed int8; other int8; Brotli artifact; 16 MB artifact cap; BigramHash.