What's Changed
- README: the 27B's and 35B-A3B's M5 Max cells on 0.2.29's prefill changes (446165e, #959 and #961): 4K prefill 904 -> 1,104 tok/s (+22%) and 6,137 -> 6,558 (+6.9%), ahead of Splash at 4K on both; decode unchanged (docs and data only) by @fearampgreed in #963
- Qwen35: fuse small-row RMSNorm and X8 prep (1K native MTP +1.2% on M4 Pro) by @chainhelen in #965
- sampling: native SIMD argmax cuts K20/K32 GPU time 8-25% on M3 Ultra by @chainhelen in #934
- The 128-row tile's outputs through the cooperative tensors' store, the same bits: 4K prefill -1.8% on the 35B-A3B (627.8 -> 616.4 ms), the 27B level, M5 Max by @fearampgreed in #964
- The 27B's prefill in 4096-row chunks on the 40-core M5 Max, the Neural Engine split's wavefront halves at 2048 rows: 4K prefill -4.2% (3,085 -> 2,954 ms), 8K -4.5% by @fearampgreed in #969
- The 27B's GEMMs of 17-63 rows on the 64-row tensor tile on the M5 Max (NI_QMT_MIN 17): a 33-row prompt 171 -> 88 ms, 4K prefill -2.5% (3,265 -> 3,182 ms) by @fearampgreed in #970
- Kapitsa 0.2.30: sets the version in CMakeLists.txt, the KAPITSA_VERSION fallbacks in src/, install.sh's example and packaging/homebrew/kapitsa.rb, as #962 did for 0.2.29. The tag v0.2.30 goes on the merge commit. by @fearampgreed in #972
Full Changelog: v0.2.29...v0.2.30