release: prepare 0.38.0

CHANGELOG.md

@@ -9,11 +9,32 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
-Requires **SKaiNET engine 0.38.0** (narrow-float codec, `Fp16DenseTensorData`, FP16 matmul
-kernels, codec-driven dispatch — engine PR #886).
+## [0.38.0] — 2026-07-31
+
+Ships against **SKaiNET engine 0.38.0**, which adds first-class dynamic tensor shapes (`Dim`) plus
+the narrow-float codec (`Fp16DenseTensorData`, FP16 matmul kernels, codec-driven dispatch — engine
+PR #886). Two headlines: **Moonshine v2 streaming ASR authored end-to-end in the SKaiNET NN DSL**
+(the last vendor-ONNX graph is gone) and **narrow-float `KEEP_NATIVE` weights** across the LLM loaders.
 
 ### Added
 
+- **Moonshine v2 — the complete streaming pipeline in the NN DSL**, self-compiled DSL → StableHLO →
+  IREE with no vendor neural binaries (`skainet-transformers-inference-moonshine`):
+  - **Audio frontend** (`MoonshineV2Frontend`): CMVN → `asinh` compression → filterbank matmul →
+    SiLU → two causal `Conv1d(k5,s2)` — the last vendor-ONNX graph, now DSL-authored (bit-exact vs
+    `frontend.onnx`, cos > 0.999).
+  - **Encoder** (position-free sliding-window local attention) and **adapter** (learned absolute
+    positional embedding, pos-embed add only) bridging the position-free memory to the decoder.
+  - **Decoder** authored in the DSL, reusing the shared KV-cache decoder.
+- **True-dynamic KV-cache decode graphs.** `MOONSHINE_V2_TRUE_DYNAMIC` / `GEMMA_TRUE_DYNAMIC` trace
+  the cache seq dim as a real dynamic extent (`Dim.DYNAMIC`), so one compiled vmfb serves every
+  autoregressive position instead of a fixed-shape re-decode. Requires engine 0.38.0's `Dim`.
+- **Fixed-max-pad cross-attention mask** for streaming decode: pad the encoder memory to a fixed MAX
+  and mask the padding, so one prefill + one `with_past` pair serve any encoder length ≤ MAX while the
+  self-cache stays dynamic (growing). `transformer-core`'s `MultiHeadAttention` gains an optional
+  trailing `crossMask` (default `null` → byte-identical for existing callers).
+- **Gemma row-dequant of the packed `token_embd`** in the shared `Embedding`, cutting host memory at
+  load.
 - **FP16 KEEP_NATIVE on the SafeTensors path.** `DecoderSafeTensorsLoader` gains the F16 arm
   that BF16 has had since 0.25.0: with a `DTypePolicy` admitting FP16 (`Require(FP16)`,
   `Prefer(FP16)`, or `OneOf` containing FP16) it stops widening F16 tensors and wraps the
@@ -68,6 +89,9 @@ kernels, codec-driven dispatch — engine PR #886).
   prevent. They now declare `keepNative = emptySet()` and reject it. **Callers relying on the
   old acceptance must switch to `Prefer(BF16)`** (a soft constraint, which still passes) until
   those chains grow a KEEP_NATIVE path.
+- Moonshine v2 encoder sliding-window off-by-one (was cos 0.991 vs ONNX); the v2 config is set to the
+  real tiny-streaming dims; the adapter is pos-embed add only (no LayerNorm).
+- kgemma heavy-trace test heap raised to 12 g, with honest skips.
 
 ## [0.36.1] — 2026-07-17
 
@@ -869,6 +893,7 @@ Version-aligned with **SKaiNET 0.21.0**.
 Last published transformers release before the engine-aligned version line.
 See `git log v0.16.0..0.18.0` for details.
 
+[0.38.0]: https://github.com/SKaiNET-developers/SKaiNET-transformers/releases/tag/0.38.0
 [0.36.1]: https://github.com/SKaiNET-developers/SKaiNET-transformers/releases/tag/0.36.1
 [0.36.0]: https://github.com/SKaiNET-developers/SKaiNET-transformers/releases/tag/0.36.0
 [0.31.0]: https://github.com/SKaiNET-developers/SKaiNET-transformers/releases/tag/0.31.0

README.md

@@ -106,30 +106,36 @@ Honest status — see the project-status note at the top of this README.
 
 ## Current release
 
-The current release is **0.36.1** (against **SKaiNET 0.36.0**) — a patch on 0.36.0 with two
-additions, both purely additive for existing consumers.
-
-**BGE embedding models** (`BAAI/bge-small-en-v1.5` and siblings) now run on the BERT DSL path:
-
-- **CLS pooling** — `BertPooling { MEAN, CLS }`, auto-detected from the sentence-transformers
-  `1_Pooling/config.json`. Pooling stays outside the traced graph, so OPTIMIZED mode and
-  StableHLO export are unaffected.
-- **Query/document asymmetry** — `EmbeddingModel` gains `embedQuery` / `embedDocument` /
-  `embedDocuments`, and `PrefixedEmbeddingModel` applies the per-model retrieval instruction
-  prefixes (E5 `query: `/`passage: `, BGE query instruction) that these models need to score
-  correctly. `fromHuggingFace` wires them automatically.
-- Design notes: [embedding-model-coverage](docs/specs/embedding-model-coverage.md).
-
-**Beam search** for the T5 decoder and the vec2text inversion loop — vec2text's main quality lever:
-
-- `T5Runtime.generateBeam(...)` does token-level beam over the decoder, returning candidates
-  best-first by length-normalized log-probability.
-- `Vec2TextInverter.invert(..., sequenceBeamWidth, tokenBeams)` adds a sequence-level beam that
-  keeps several hypotheses across correction rounds, ranked by cosine similarity to the target
-  embedding.
-- Both are **off by default** — width 1 keeps the existing greedy behaviour, so this is a
-  drop-in upgrade. On the round-trip test's example sentence, one correction step with beam
-  (sequence width 3, token beams 3) improves cosine **0.765 → 0.818** over greedy.
+The current release is **0.38.0** (against **SKaiNET 0.38.0**) — the release that completes the
+**Moonshine v2 streaming ASR entirely in the SKaiNET NN DSL** and adds **narrow-float weights**.
+
+**Moonshine v2 — the whole pipeline is now DSL-authored and self-compiled** (DSL → StableHLO → IREE,
+no vendor neural binaries) in `skainet-transformers-inference-moonshine`: the **audio frontend**
+(CMVN → asinh → filterbank → SiLU → two causal `Conv1d`), the position-free sliding-window
+**encoder**, the learned-positional-embedding **adapter**, and the KV-cache **decoder**. The last
+vendor-ONNX graph — the frontend — is gone.
+
+- **Streaming decode over a growing cache.** The decoder exports **true-dynamic** KV-cache graphs
+  (`Dim.DYNAMIC`), so one compiled artifact serves every autoregressive position instead of a fixed
+  re-decode. A **fixed-max-pad cross-attention mask** lets one prefill + one `with_past` pair serve
+  any encoder-memory length ≤ MAX. `transformer-core`'s `MultiHeadAttention` gains an optional
+  trailing `crossMask` (default null — byte-identical for existing callers).
+
+**Narrow-float `KEEP_NATIVE` weights** — keep FP16/BF16 packed on the SafeTensors and GGUF paths
+instead of widening to FP32 at load (half the weight bytes at rest), and relay matmul weights
+input-major so the per-forward `.t()` becomes a zero-copy view: **BF16 1.8–1.9×, FP16 1.5–1.7×** over
+FP32, and a 4.4 s-per-projection widening on `ffn_up` 8B removed. Wired across
+llama/qwen/gemma/apertus/voxtral through the `DTypePolicyValidation` policy surface.
+
+**Gemma** now **row-dequants the packed `token_embd`** in the shared `Embedding`, cutting host memory
+at load.
+
+This is the first release on **SKaiNET engine 0.38.0**, which ships first-class dynamic tensor shapes
+(`Dim`) — the capability the true-dynamic decode export is built on.
+
+It builds on **0.36.1**, which added **BGE embedding models** on the BERT DSL path (CLS pooling +
+retrieval prefixes) and **beam search** for the T5 decoder and the vec2text inversion loop — both
+additive, drop-in for existing consumers.
 
 It builds on **0.36.0**, in which **BERT became completely defined in the SKaiNET NN DSL** and the
 deprecated hand-coded eager BERT stack was **removed (BREAKING)**:
@@ -188,7 +194,7 @@ The recommended way to consume is via the BOM. It pins every published `skainet-
 
 ```kotlin
 dependencies {
-    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.36.1"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.38.0"))
 
     // Versions resolved from the BOM:
     implementation("sk.ainet.transformers:skainet-transformers-core")

gradle.properties

@@ -1,5 +1,5 @@
 GROUP=sk.ainet.transformers
-VERSION_NAME=0.36.1
+VERSION_NAME=0.38.0
 
 POM_DESCRIPTION=SKaiNET-transformers