Skip to content

v0.10.0

Choose a tag to compare

@github-actions github-actions released this 07 Nov 10:14
· 109 commits to main since this release
f8bc1c9

Basalt v0.10.0

brew tap goodmai/basalt https://github.com/goodmai/basalt
brew trust goodmai/basalt
brew install goodmai/basalt/basalt

Builds from source: Xcode 27.0+ (Swift 6.4) and macOS 27+, Apple Silicon.

Source tarball SHA256: 075368acd37a181321b6ad64a5ffd6ce532495be818e56b79433049b274a28ca

What's Changed

  • chore: point formula at v0.9.0 by @github-actions[bot] in #268
  • feat(#255): batched decode for recurrent hybrids; fix concurrent forwards racing compiled traces by @goodmai in #269
  • feat(#37): head-level KV split — retrieval heads keep the context, streaming heads keep sinks and a window by @goodmai in #271
  • feat(#33): /v1/models/preview, wired limit by default, mlx-swift 0.33, watchdog for long prefills by @goodmai in #276
  • fix(inference): every registerAliases caller waits until the aliases are registered by @goodmai in #279
  • fix(#277): TTFT counts the prefill the iterator runs when it is built by @goodmai in #278
  • feat(#33): experimental Quest block-sparse decode attention behind serve --quest-top-k by @goodmai in #281
  • chore(#236): remove the GemCLI, GemLogger and GemLogLevel typealiases by @goodmai in #283
  • docs(#284): size and complexity limits with a ratchet; Swift 6.4 adoption in CONTRIBUTING and the skills by @goodmai in #285
  • docs: register Xcode's MCP server and route API questions to the xcode-mcp skill by @goodmai in #288
  • test: rate-limit router test no longer depends on runner speed by @goodmai in #287
  • refactor(#284): JSONSchemaGate — split checkDroppedKeywords and visit along their seams by @goodmai in #289
  • ci(#284): size and complexity ratchet on changed Swift files by @goodmai in #286
  • refactor(#284): DiffParser.parseUnifiedDiff — one flush, body lines classified in diffLine by @goodmai in #293
  • refactor(#284): ServeCommand.run and ModelsCommand.run — one memory report, run split along its seams by @goodmai in #295
  • feat(#270): choose KV heads by ablation; measure head shares on harder tasks than passkey by @goodmai in #296
  • ci(#284): type ≤ 200 / extension ≤ 150 lines in the ratchet; value core, actor shell in CONTRIBUTING by @goodmai in #298
  • ci(#284): comment checker and xcode-mcp helper in Swift by @goodmai in #299
  • docs(#284, #300): legacy types ≤ 450, race-free by construction rules; size checker closes its pipes by @goodmai in #302
  • ci: Unit Tests skip docs-, skills- and checker-only changes on the single Mac runner by @goodmai in #306
  • test(#250): mark the Nemotron-H prompt-lookup equivalence as a known intermittent issue by @goodmai in #307
  • refactor(#284): MLXInferenceEngine.launch — FreshCache plan as a value core, one function per path by @goodmai in #303
  • types(#304): TokenCount and ByteCount units in KVCacheGeometry's public API by @goodmai in #305
  • refactor(#284): MLXInferenceEngine — load split along its seams, static helpers out of the actor body by @goodmai in #308
  • ci: Swift checker jobs run in the swift:6.1-noble container (Swift is leaving the Ubuntu runner images) by @goodmai in #312
  • refactor(#284): MCPServer — value core for routing, arguments and results; 677 → 40 lines by @goodmai in #310
  • refactor(#284): ServeCommand — option groups resolve into Sendable values (692 → 186 lines) by @goodmai in #309
  • refactor(#284): SessionLogWriter — value core, actor shell by @goodmai in #311
  • refactor(#284, #300): value cores for controllers, router, orchestrator, hub, disk cache, fit/inspector, schema gate, harness, table renderer; CacheBlendStore and MTPPrefix without @unchecked by @goodmai in #313
  • refactor(#284, #300, #33): function limits across CLI, server, core, harness and inference; prompt-cache leases stay on the engine actor; MTP drafter inputs as values; Quest review fixes by @goodmai in #314
  • concurrency(#300, #304): zero @unchecked Sendable — prompt caches, MTP drafter and batch state owned by the model context; typed prompt-cache bytes by @goodmai in #320
  • refactor(#300, #315–#318, #321): generation plan, engine state machine, one MLX owner, deterministic tests, property tests, CI ratchets, exhaustive error mapping by @goodmai in #330
  • feat(#322–#327, #300): capability registry, metamorphic matrix, reference logits, parity runner, load-time probe, health monitors, fault injection by @goodmai in #331
  • feat(#328): model X-ray — MCP tools for structure, tensors, KV cache, traces (logit lens, attention, experts) and sandboxed experiments (steering, ablation, weight deltas, compare) by @goodmai in #333
  • refactor(#284): tail — engine stages, relay, DTO inits, Task.detached gone; #323 KV quantization gate, #326 probe ceilings, #321, #301 by @goodmai in #334
  • fix(#329): Nemotron-H group routing takes the two largest scores per group; lfm2_moe reference uses transformers' router by @goodmai in #335
  • feat(#332): research modes over MCP — gate, live adapters, pack/unpack, introspection, contrastive steering, self-training, weight conversion by @goodmai in #336
  • test(#319): batched decode contract — quality parity with solo, exact invariants pinned on tiny and real hybrids by @goodmai in #337
  • fix(#325): health monitors checked on real models — a reply cut off at max_tokens is not degenerate; tiny-model tests; mid-stream error message by @goodmai in #338
  • fix(#250): the Mamba-2 decode kernel wrote no state for state sizes below 32 — the lookup flake's root cause; pin cd871d3 by @goodmai in #339
  • refactor(#304): KV and memory budgets are ByteCount end to end; decode tok/s unchanged by @goodmai in #340
  • test(#322): LFM2.5-24B parity goldens — the benchmark prints the reasoning when the answer is empty by @goodmai in #342
  • refactor(#284): exhaustive switches over basalt's own enums — the default: audit by @goodmai in #341
  • refactor(#300): stage 3 — reducer transitions in the orchestrator, router, hub and session-log actors; three reentrancy races fixed by @goodmai in #343
  • test(#291): real-model regression gate — golden top-64 logit signatures (KL 1e-4) on Qwen3.8-27B, LFM2.5-24B and Gemma 4 26B-A4B, plus a footprint check by @goodmai in #347
  • feat(#346): regex, EBNF and structural-tag response formats; fix constrained answers reported as reasoning on LFM2.5 by @goodmai in #348
  • feat(#345): image input end to end on REST and MCP — images were dropped; Gemma 4 26B-A4B answers the same through all three paths by @goodmai in #349
  • feat(#346): constrained requests join the batched decoder, each row with its own grammar mask by @goodmai in #350
  • Batch: Gemma 4 joins the batched decoder (#255), embeddings yield to chat (#233), roofline with prefill (#272), seeded MTP tests (#290) by @goodmai in #354
  • Batch: download digests and a symlink-blob fix (#353), rerank yields to chat (#355), --kv-bits refusal (#165), jump-forward with prefill (#205) by @goodmai in #356
  • Batch: prompt lookup with quantized KV, 2.1–2.4× on code (#163), thermal pacing (#45), chat history digest (#23) by @goodmai in #368
  • Batch 8: YAML config, LoRA adapters, inspect --verify, basalt review, release CI, experimental zero-copy body (#344 #179 #367 #35 #206 #345) by @goodmai in #369
  • Batch 9: acceptance gate, verify overhead, small-M matmul, spectral and key-trajectory spikes, sessions LTL check (#246 #210 #208 #358 #364 #359) by @goodmai in #375
  • ci: bump github/codeql-action/analyze from 4.39.1 to 4.39.2 by @dependabot[bot] in #371
  • deps: bump github.com/hummingbird-project/hummingbird from 2.28.0 to 2.28.1 by @dependabot[bot] in #373
  • ci: bump github/codeql-action/init from 4.39.1 to 4.39.2 by @dependabot[bot] in #370
  • deps: swift-transformers 1.4.0 with compatibility for its breaking changes (tokenizer errors, BPE decoding) by @goodmai in #378
  • NF4 phases (a) and (b) behind experimental opt-ins: codec, tensor_pack, convert, fork loader fix, error/KL/speed runners (#357) by @goodmai in #377
  • Batch 10: image CPU cuts, kv-evict on Gemma 4, head split with k8v4 (refused on hybrids), adaptive thinking budget, mixed-bit convert, MoE streaming fit tier (#374 #171 #294 #41 #292 #242) by @goodmai in #379
  • Batch 11: small-M 4-bit kernel, MoE expert streaming, fidelity-gated head split, engine configuration refactor (#208 #225 #282 #237) by @goodmai in #380
  • release: 0.10.0 by @goodmai in #381

Full Changelog: v0.9.0...v0.10.0