Skip to content

release: prepare 0.38.0 - #268

Merged
michalharakal merged 1 commit into
developfrom
release/0.38.0
Jul 31, 2026
Merged

release: prepare 0.38.0#268
michalharakal merged 1 commit into
developfrom
release/0.38.0

Conversation

@michalharakal

Copy link
Copy Markdown
Contributor

Prepares the 0.38.0 release (transformers tracks the engine version — ships against SKaiNET engine 0.38.0, which adds first-class dynamic tensor shapes, Dim).

Contents (since 0.36.1)

  • Moonshine v2 — the whole streaming pipeline authored in the SKaiNET NN DSL and self-compiled (DSL → StableHLO → IREE, no vendor neural binaries): audio frontend (MoonshineV2Frontend — the last vendor-ONNX graph, now gone), position-free sliding-window encoder, learned-positional-embedding adapter, KV-cache decoder.
  • True-dynamic KV-cache decode graphs (Dim.DYNAMIC) + fixed-max-pad cross-attention mask — one compiled artifact serves every decode position / any encoder-memory length ≤ MAX. transformer-core MultiHeadAttention gains an optional trailing crossMask (default null → byte-identical for existing callers).
  • Narrow-float KEEP_NATIVE weights (fp16/bf16 packed on SafeTensors + GGUF, input-major matmul relayout) — BF16 1.8–1.9×, FP16 1.5–1.7× over FP32.
  • Gemma row-dequant of the packed token_embd.

Changes in this PR

  • VERSION_NAME 0.36.1 → 0.38.0
  • CHANGELOG.md — new [0.38.0] section (merged with the narrow-float entry already in [Unreleased])
  • README.md — "Current release" rewrite + BOM snippet bumped to 0.38.0

Merging lands the version bump on develop. Publishing to Maven Central is a separate, manual step — pushing the 0.38.0 tag triggers the release workflow (./gradlew publish), then Sonatype is promoted manually (SONATYPE_AUTOMATIC_RELEASE=false).

Ships against SKaiNET engine 0.38.0 (first-class dynamic tensor shapes, Dim).
Headlines: Moonshine v2 streaming ASR authored end-to-end in the SKaiNET NN DSL
(the last vendor-ONNX graph — the audio frontend — is gone; the whole frontend/
encoder/adapter/decoder pipeline self-compiles DSL -> StableHLO -> IREE), and
narrow-float KEEP_NATIVE fp16/bf16 weights across the LLM loaders. Also: true-dynamic
KV-cache decode graphs, fixed-max-pad cross-attention mask, Gemma token_embd row-dequant.

Bumps VERSION_NAME 0.36.1 -> 0.38.0 (transformers tracks the engine version), updates
README + CHANGELOG.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@michalharakal
michalharakal merged commit 0bb37ce into develop Jul 31, 2026
4 checks passed
@michalharakal
michalharakal deleted the release/0.38.0 branch July 31, 2026 16:58
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant