What's Changed
- Benchmark: Mac Studio (2025) · M3 Ultra · 512 GB · Qwen3.8 27B Nace Q4 by @Halkins in #1314
- Benchmark: Mac mini (2026) · M5 Pro · 64 GB · Ornith-1.5-35B-A3B 4-bit by @carven-ji in #1315
- Nemotron-H keeps its drafter at top_k 0: KFlash's tree verify samples the whole vocabulary, Lightning + DSpark 94.8 -> 152.2 tok/s on an M5 Pro (#1286) by @alish-nace-23 in #1318
- Benchmark: Mac mini (2026) · M6 · 32 GB · Qwen3.6 35B-A3B 4-bit by @supervised-nl in #1320
- Benchmark: Mac mini (2026) · M6 · 32 GB · Qwen3.8 27B Nace Q4 by @supervised-nl in #1319
- Qwen3.5/3.6/3.8 (Qwen35): skip last-layer tail on non-sampling chunks without MTP by @supervised-nl in #1317
- Grammar: unmask duplicate-byte token IDs in TokenTable so fill_mask matches advance (Not verified on a GPU) by @supervised-nl in #1325
- API: prevent integer wrap-around on uint64 token limits/budgets and top_k narrowing (Not verified on a GPU) by @supervised-nl in #1326
- GGUF import: align Qwen3.5/3.6 shape limits with Qwen35::check_shapes and retry EINTR (Not verified on a GPU) by @supervised-nl in #1327
- DFlash2: omit 41.9 MB debug cdbg_ buffer and size logits_ to head_rows() (-51.7 MB on 27B) by @supervised-nl in #1323
- SSD prefix cache: align file header to 16 KiB page size so F_NOCACHE evicts full payload pages by @supervised-nl in #1321
- SSD prefix cache: save and restore system-prompt prefix checkpoints with DFlash2 by @supervised-nl in #1322
- Tokenizer: match added_tokens with any leading byte and skip empty content (Not verified on a GPU) by @supervised-nl in #1324
- server: cancel queued Responses on client disconnect before prefill by @alish-nace-23 in #1330
- test: keep slot holders alive through queued-stream heartbeats by @alish-nace-23 in #1333
- Tokenizer: replace u2b unordered_map with stack lookup and reserve vocab/merges (-25 ms load) (Not verified on a GPU) by @supervised-nl in #1328
- Expert store: slots of any plane layout, up to nine planes and one layout per layer, for Hy4-flash; Flash-Next's Q2_0 output bitwise unchanged by @alish-nace-23 in #1334
- MLX affine routed experts (3-/4-bit, group 32/64/128), resident or streamed from the SSD: 0.24 ms a layer per decode row at Hy4's shape on an M5 Pro by @alish-nace-23 in #1335
- Hy4-flash GPU blocks: hyper-connections, sigmoid router, Engram mix, latent MLA with sinks and windows, compressed indexer; existing models' outputs bitwise unchanged by @alish-nace-23 in #1336
- Expert store refuses slot layouts whose planes overlap; fn_store_test covers all 11 refused layouts (Flash-Next Q2_0 outputs bitwise identical) by @alish-nace-23 in #1342
- Hy4-flash Engram lookup on the CPU, its rows bitwise equal to mx.dequantize with fp32 scales by @alish-nace-23 in #1337
- Run Hy4-flash on M5 Pro 48 GB: 3.1 to 4.1 tok/s with streamed experts by @alish-nace-23 in #1338
- Logit floor: reject cached dumps labeled for another model (Not verified on a GPU) by @chainhelen in #1343
- Hub pull: encode slashes in model revisions (Not verified on a GPU) by @chainhelen in #1344
- Hy4-flash reference check: hy4_dump writes logits, hy4_ref.py scores them against BF16 (re-scores Kapitsa vs MLX at KL 0.0225) by @alish-nace-23 in #1339
- Hub: keep slash and dash revisions in separate caches (Not verified on a GPU) by @chainhelen in #1345
- Hy4-flash against MLX on the same pack: hy4_mlx_layers.py and NI_HY4_DUMP_SEL compare logits and the indexer's selections, not verified on a GPU by @alish-nace-23 in #1340
- Hy4 score: emit parseable JSON with --json - (Not verified on a GPU) by @chainhelen in #1346
- Hy4 score: select the matching recorded prompt when token prefixes overlap (Not verified on a GPU) by @chainhelen in #1347
- Responses: preserve namespaced custom tool calls (Not verified on a GPU) by @chainhelen in #1348
- Reuse a spare Qwen checkpoint for prompts with a trimmed ending by @chainhelen in #1350
- Hy4 score: fail when no requested prompts can be compared (Not verified on a GPU) by @chainhelen in #1351
- GGUF import: record the NVFP4 source format and reuse earlier imports by @chainhelen in #1355
- feat(qwen35): add general output token suppression (--suppress-tokens) by @audreyt in #1356
- Hy4-flash through the API: thinking off renders no_think, tool calls parse; kapitsa bench's decode runs valid (24.2 tok/s on the M5 Max) by @fearampgreed in #1359
- Hy4-flash keeps every expert in memory when the working set holds them: decode 25.3 -> 36.7 tok/s on a 128 GB M5 Max, logits bitwise the same by @fearampgreed in #1360
- Hy4-flash's expert prefill GEMMs on the M5's matrix units, a fast profile member: 4K prefill 596 -> 930 tok/s, 16K 625 -> 989 on the M5 Max (86 flips / KL 0.137 against MLX's bar of 130 / 0.259) by @fearampgreed in #1361
- README: Hy4-flash's M5 Max cell, 596 / 36 tok/s (prefill 4K / decode T 0.6, plain decoding), and its results folder (docs and data only) by @fearampgreed in #1362
- Kapitsa 0.2.38: sets the version in CMakeLists.txt, the KAPITSA_VERSION fallbacks in src/, install.sh's example and packaging/homebrew/kapitsa.rb, as #1310 did for 0.2.37. The tag v0.2.38 goes on the merge commit by @fearampgreed in #1364
New Contributors
Full Changelog: v0.2.37...v0.2.38