modelbeast df28940d81 docs: root-cause the sdpa finding — head_dim=56 misses the fast path ... Shape-spy on a real 2048 forward + exact-shape sweeps on M1/M3 Ultra: fused==unfused at hdim 56 (silent fallback) on both machines, so use_sdpa buys nothing and its transpose choreography costs ~4.5x on M1. Padding to hdim 64 engages the fast kernel: 2.2x (M3) / 2.7x (M1) on the dominant global-attention shape despite 14% extra FLOPs.		2026-07-16 23:05:13 +10:00
..
benchmarks	docs: wave2 ablation benchmarks + optimization plan + brainstorm	2026-03-09 19:32:04 -02:30
brainstorms	docs: wave2 ablation benchmarks + optimization plan + brainstorm	2026-03-09 19:32:04 -02:30
plans	docs: wave2 ablation benchmarks + optimization plan + brainstorm	2026-03-09 19:32:04 -02:30
2026-07-16-m-series-fleet-ablation-results.md	docs: root-cause the sdpa finding — head_dim=56 misses the fast path	2026-07-16 23:05:13 +10:00
MLX Vision Transformer Optimization Techniques.md	docs: wave2 ablation benchmarks + optimization plan + brainstorm	2026-03-09 19:32:04 -02:30