SKaiNET-developers · michalharakal · Sep 25, 2026 · Sep 25, 2026 · Sep 25, 2026 · Sep 25, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -9,8 +9,40 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.57.0] — 2026-09-25
+
+Lock-step with **SKaiNET engine 0.57.0**. Headline: **Qwen on the compiled IREE KV path**: the
+qwen-kv-v1 export, runtime and templates, with Qwen3-0.6B verified token-for-token against llama.cpp
+on host IREE; on-device validation is still to come (see the entry below for exact status).
+
+### Changed
+
+- **Engine 0.57.0, Kotlin 2.4.20** (#453). The engine fixes the StableHLO export of an explicit
+  attention mask under grouped-query attention and onto a dynamic key length (SKaiNET#1302), and
+  makes kotlinx-io part of `skainet-data-source`'s API. This repo moves to Kotlin 2.4.20 with it.
+- **Dependencies**: Ktor client 3.6.0 (#448), kctfork 0.14.0 (#449), jackson-databind 2.22.3 (#454).
+
+### Added
+
+- **FunctionGemma export contract carries the embedding geometry** (#452). `FunctionGemmaSpec` gains
+  `vocabSize` (default 262144), `nHeads`, `hiddenSize` and `slidingWindow`, and `manifest.json` emits
+  them, so `IreeKvSpec.fromManifest` no longer falls back to the stock constants. The export CLI reads
+  the vocabulary size from the checkpoint's `token_embd.weight` (`GEMMA_VOCAB` overrides): a fine-tune
+  that added special tokens has more rows, and the native KV session locates the embedding table by
+  `vocabSize × hiddenSize` bytes and clamps ids ≥ `vocabSize` to 0, so with the stock constant those
+  tokens would silently have become token 0. `IreeKvSpec.functionGemma270m` takes `vocabSize`.
+
+### Fixed
+
+- **Stale `functiongemma` API dump.** #452 changed the public API without re-dumping it, so
+  `apiCheck` failed on `develop`; refreshed in #453.
+
 ### Added — Qwen on the compiled IREE KV path (closes #411, #409)
 
+Status in this release: Qwen3-0.6B exports, compiles for Vulkan (valhall4) and arm32, and matches
+mainline llama.cpp's 32 greedy tokens on host IREE (`QwenVmfbParityTest`). Not yet run on a device.
+Qwen2.5-0.5B export is blocked by a SKaiNET core gap (attention bias does not externalize).
+
 - **`IreeKvSession` / `IreeKvSpec` generalized for GQA and single-RoPE-base models**
   (`llm-runtime:iree-android`): `nKvHeads` may now be any value with `nHeads % nKvHeads == 0`
   (grouped-query attention — Qwen2.5-0.5B nKvHeads=2, Qwen3-0.6B nKvHeads=8), not just
@@ -98,15 +130,6 @@ A transformers-only release against **SKaiNET engine 0.56.0** (unchanged).
 
 ### Added
 
-- **`IreeMoonshineStream`** (`llm-runtime:iree-android`, `libskainet_moonshine_stream.so`, arm64-v8a +
-  armeabi-v7a, Vulkan + local-task): the streaming Moonshine v2 speech-to-text runtime over the five
-  graphs of `MoonshineV2ExportCli` — PCM in, cumulative partial transcripts out, exact final on
-  `finish()`. The counterpart of `IreeKvSession` for ASR: the last piece a Moonshine cartridge needed
-  from a released artifact. `native/build-moonshine-stream.sh` builds it with the same image and
-  cache as the other two libraries.
-
-### Added
-
 - **`IreeMoonshineStream`** (`llm-runtime:iree-android`, `libskainet_moonshine_stream.so`, arm64-v8a +
   armeabi-v7a, Vulkan + local-task): the streaming Moonshine v2 speech-to-text runtime over the five
   graphs of `MoonshineV2ExportCli` — PCM in, cumulative partial transcripts out, exact final on

diff --git a/README.md b/README.md
@@ -109,18 +109,24 @@ Honest status — see the project-status note at the top of this README.
 
 ## Current release
 
-The current release is **0.56.2** (against **SKaiNET 0.56.0** — a transformers-only release, same
-pattern as 0.54.1: no new engine version needed).
-
-**`IreeMoonshineStream`: the streaming Moonshine v2 speech-to-text runtime.** `libskainet_moonshine_stream.so`
-(`llm-runtime:iree-android`, arm64-v8a + armeabi-v7a, Vulkan or CPU) drives the five graphs of
-`MoonshineV2ExportCli` — frontend, encoder, adapter, masked prefill, dynamic with-past step — plus the
-shared decoder parameter archive as a real streaming loop on the device: PCM in, cumulative partial
-transcripts out, one exact full re-decode on `finish()`. The Kotlin binding is shaped like `IreeKvSession`:
-files by absolute path, no `Context`. With the exporter (0.56.1) this is the last piece a Moonshine
-cartridge needed from a released artifact rather than from C source of its own.
-
-It builds on **0.56.1**, which released `MoonshineV2ExportCli` (a Hugging Face snapshot in, the five
+The current release is **0.57.0**, in lock-step with **SKaiNET 0.57.0**.
+
+**Qwen on the compiled IREE KV path.** `QwenExportHarness` / `QwenExportCli` trace Qwen2 and Qwen3
+checkpoints to the three graphs of the `qwen-kv-v1` contract (catalog-prefix prefill, one-call chunk,
+dynamic-cache decode step), `IreeKvSession` runs grouped-query models, and `Qwen25ChatTemplate` renders
+Qwen2.5's official tool-calling template. Qwen3-0.6B compiles for Vulkan and arm32 and reproduces
+llama.cpp's greedy tokens exactly on host IREE; it has not been run on a device yet, and Qwen2.5 export
+waits on an engine fix for attention bias. **The Moonshine v2 stream** caps its final re-decode at the
+model card's length budget, so short commands finish sooner.
+
+**Engine 0.57.0 and Kotlin 2.4.20.** The engine fixes the StableHLO export of an explicit attention mask
+under grouped-query attention and onto a dynamic key length, the shape of a chunked-prefill graph over a
+KV cache. **The FunctionGemma export contract now carries the embedding geometry** (`vocabSize`,
+`nHeads`, `hiddenSize`, `slidingWindow`) into `manifest.json`, read from the checkpoint, so a fine-tune
+with added special tokens exports and runs without its new tokens silently collapsing to token 0.
+
+It builds on **0.56.2**, which released `IreeMoonshineStream`, the streaming Moonshine v2 runtime on the
+device; **0.56.1**, which released `MoonshineV2ExportCli` (a Hugging Face snapshot in, the five
 StableHLO graphs and two host-side tables out, no Python, nothing to configure per language or size), and
 **0.56.0**, which put the two version lines back in lock-step with **SKaiNET 0.56.0** (the engine skipped
 0.55.0 to realign them).
@@ -256,7 +262,7 @@ The recommended way to consume is via the BOM. It pins every published `skainet-
 
 ```kotlin
 dependencies {
-    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.56.2"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.57.0"))
 
     // Versions resolved from the BOM:
     implementation("sk.ainet.transformers:skainet-transformers-core")

diff --git a/docs/modules/ROOT/pages/how-to/export-moonshine-v2-streaming.adoc b/docs/modules/ROOT/pages/how-to/export-moonshine-v2-streaming.adoc
@@ -32,7 +32,7 @@ From your own build, resolve the published module and run its main class:
 ----
 val exportTool by configurations.creating
 dependencies {
-    exportTool(platform("sk.ainet.transformers:skainet-transformers-bom:0.56.2"))
+    exportTool(platform("sk.ainet.transformers:skainet-transformers-bom:0.57.0"))
     exportTool("sk.ainet.transformers:skainet-transformers-inference-moonshine")
 }
 tasks.register<JavaExec>("exportMoonshine") {

diff --git a/docs/modules/ROOT/pages/reference/moonshine-encoder.adoc b/docs/modules/ROOT/pages/reference/moonshine-encoder.adoc
@@ -24,7 +24,7 @@ a DSL decoder is future work.
 [source,kotlin]
 ----
 dependencies {
-    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.56.2"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.57.0"))
     implementation("sk.ainet.transformers:skainet-transformers-inference-moonshine")
 }
 ----

diff --git a/docs/modules/ROOT/pages/tutorials/android-getting-started.adoc b/docs/modules/ROOT/pages/tutorials/android-getting-started.adoc
@@ -55,7 +55,7 @@ dependencies {
     // self-registers it on ART at process start — nothing to call.
     runtimeOnly("sk.ainet.core:skainet-backend-jni-cpu")
 
-    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.56.2"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.57.0"))
     implementation("sk.ainet.transformers:skainet-transformers-core")
     implementation("sk.ainet.transformers:skainet-transformers-runtime-kllama")
     implementation("sk.ainet.transformers:skainet-transformers-inference-llama")

diff --git a/docs/modules/ROOT/pages/tutorials/getting-started-java.adoc b/docs/modules/ROOT/pages/tutorials/getting-started-java.adoc
@@ -25,7 +25,7 @@ In your `build.gradle.kts`:
 [source,kotlin]
 ----
 dependencies {
-    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.56.2"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.57.0"))
 
     implementation("sk.ainet.transformers:skainet-transformers-runtime-kllama")
     implementation("sk.ainet.transformers:skainet-transformers-agent")
@@ -41,7 +41,7 @@ Or in Maven (Maven needs the `-jvm` classifier suffix on platform artifacts):
     <dependency>
       <groupId>sk.ainet.transformers</groupId>
       <artifactId>skainet-transformers-bom</artifactId>
-      <version>0.56.2</version>
+      <version>0.57.0</version>
       <type>pom</type>
       <scope>import</scope>
     </dependency>

diff --git a/docs/modules/ROOT/pages/tutorials/getting-started-leaf.adoc b/docs/modules/ROOT/pages/tutorials/getting-started-leaf.adoc
@@ -34,7 +34,7 @@ encoder output to the advertised dimensionality. The runtime applies it automati
 [source,kotlin]
 ----
 dependencies {
-    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.56.2"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.57.0"))
     implementation("sk.ainet.transformers:skainet-transformers-providers")
 }
 ----

diff --git a/docs/modules/ROOT/pages/tutorials/llama3-tool-calling.adoc b/docs/modules/ROOT/pages/tutorials/llama3-tool-calling.adoc
@@ -52,7 +52,7 @@ The pieces you need live in three modules:
 [source,kotlin]
 ----
 dependencies {
-    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.56.2"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.57.0"))
 
     implementation("sk.ainet.transformers:skainet-transformers-runtime-kllama")
     implementation("sk.ainet.transformers:skainet-transformers-agent")

diff --git a/gradle.properties b/gradle.properties
@@ -1,5 +1,5 @@
 GROUP=sk.ainet.transformers
-VERSION_NAME=0.56.2
+VERSION_NAME=0.57.0
 
 POM_DESCRIPTION=SKaiNET-transformers