What's Changed
- GGUF: remove per-row four-bit permutation copies by @chainhelen in #1082
- GGUF: cut Q6_K CPU row decode time 76–81% by @chainhelen in #1084
- GGUF: cut four-bit CPU row packing time 89–92% by @chainhelen in #1086
- GGUF: cut Q4_0 CPU code decode time 69–78% on M4 Pro by @chainhelen in #1087
- GGUF: stop descriptor leaks on split-count errors (Not verified on a GPU) by @chainhelen in #1089
- GGUF: join row workers on launch failure (Not verified on a GPU) by @chainhelen in #1090
- Fold 4 narrow M4 chip gates into simd_gpu; keep m1_max_32 guard (#393) by @alish-nace-23 in #1091
- GGUF: close output descriptors on import failure (Not verified on a GPU) by @chainhelen in #1092
- Keep Clover out of the KFlash history lane: measured 4% slower at temperature 0.6 by @alish-nace-23 in #1094
- NI_QMT_MIN 17 for dense 4B/9B-class models on the M5 (27B keeps 64 below 40 cores): 40-token prefill ~2x faster (#393) by @alish-nace-23 in #1096
- NI_PF_WAVE on the M5 class, Clover included: 16K prefill ~4% faster on the M5 Pro (#393) by @alish-nace-23 in #1097
- Qwen model types in one table, so a new Qwen under a new name is one row (no behavior change) by @alish-nace-23 in #1095
- Kapitsa 0.2.33: sets the version in CMakeLists.txt, the KAPITSA_VERSION fallbacks in src/, install.sh's example and packaging/homebrew/kapitsa.rb, as #1085 did for 0.2.32. The tag v0.2.33 goes on the merge commit. by @fearampgreed in #1099
Full Changelog: v0.2.32...v0.2.33