What's Changed
- Results: a chunked Gated DeltaNet prefill scan (WY form, fp32 simdgroup matrices) that did not pay on the M5 Max: 2.64 ms against gdn_step_parr4_l16's 1.21 ms on the 35B-A3B's shape, outputs at rounding level (results only) by @fearampgreed in #951
- README: the 35B-A3B's M5 Max cell re-measured on 0.2.28 (main 3879b44, after #941 and #946): 4K prefill 6,036 -> 6,137 tok/s (+1.7%), decode 346 -> 347, greedy 371 -> 376 (docs and data only) by @fearampgreed in #952
- serve: reclaim model-switch GPU buffers (4B: 4.47 to 2.76 GB) by @chainhelen in #933
- Weight loads stay out of the page cache: misaligned safetensors reads go through page-aligned staging, so F_NOCACHE applies (#939) by @alish-nace-23 in #945
- Records: re-measure of merged #922, #923 and #926 on the M5 Pro 48 GB, Flint against mlx-vlm (docs only) by @alish-nace-23 in #954
- Benchmark: MacBook Pro (16-inch, 2026) · M5 Max · 128 GB · Qwen3.8 27B 4-bit by @vongostev in #913
- bench/community/RESULTS.md: regenerate after #913 by @alish-nace-23 in #958
- metal: prepack 4B QMM on M4 Pro 48 GiB (paired decode about +1%) by @KeyingSong0123 in #929
- Dense prefill GEMMs on a 128-row tensor tile (NI_QMT128), the same bits: 4K prefill -5.2% on the 35B-A3B (662.5 -> 627.9 ms), -7.3% on the 27B, -5.2% on Clover on the M5 Max by @fearampgreed in #959
- The 27B's Neural Engine share retuned for the 128-row tile: 16 -> 14 of 34 units on the M5 Max, 4K prefill -12.7% (3,579.8 -> 3,126.9 ms) by @fearampgreed in #961
- Chat Completions: llama.cpp's "return_progress", prompt_progress chunks while the prompt is read, one per prefill pass (opt-in, no engine or numeric change) by @yonaskolb in #955
- Kapitsa 0.2.29: sets the version in CMakeLists.txt, the KAPITSA_VERSION fallbacks in src/, install.sh's example and packaging/homebrew/kapitsa.rb, as #948 did for 0.2.28. The tag v0.2.29 goes on the merge commit. by @fearampgreed in #962
New Contributors
- @vongostev made their first contribution in #913
- @yonaskolb made their first contribution in #955
Full Changelog: v0.2.28...v0.2.29