Release 0.54.0

CHANGELOG.md

@@ -9,6 +9,53 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.54.0] — 2026-09-07
+
+Version lock-step with the engine continues: this release ships against **SKaiNET 0.54.0**
+(SKEEP-005's `Schedule` API, the `CoroutineSchedule` pool-deadlock fix, `SafeTensorsParametersLoader`
+`tensorFilter` parity, and the `ExperimentalMemoryApi` opt-in gate removed). It also closes out the
+FunctionGemma/IREE-Android chunked-KV work (#410) and fixes two bugs found on real hardware: a
+broken `runtime-kgemma` Maven Central POM (#408) and an Android crash in FunctionGemma tool-call
+parsing (#407).
+
+### Added — chunked prefill and a stateful Android KV session for FunctionGemma (#406, #410)
+
+- **Position-selected graphs `gemma_at` / `gemma_prefill_at`** (#415): the LM head runs on one
+  one-hot-selected position instead of every position in `SEQ`, cutting the redecode step's wasted
+  work (#406).
+- **Chunk prefill-with-past graph `gemma_prefill_with_past`** (#417): a fixed 64-token chunk against
+  the dynamic cache in one call, with per-head chunk masks (a broadcast over heads to a dynamic
+  shape isn't expressible in static StableHLO).
+- **`IreeKvSession` / `IreeKvDecoder`** (#416, #418): the Android-native stateful KV session —
+  three IREE sessions, device-resident K/V, zero-copy 512-position tail views for the sliding
+  layers, native RoPE tables + chunk masks, embedding rows read from the archive, snapshot/restore
+  without copies. Measured on a MagentaTV One (Mali via Vulkan, bf16 archives): a once-per-process
+  843-token catalog prefill, then **p50 5.87 s / max 6.25 s per utterance** (restore + one chunk +
+  16 decode tokens), down from minutes on the stateless redecode contract.
+- **`iree-android` failure reporting** (#404): native failures surface as reported errors instead
+  of a silent `null`; bare function names are qualified with `module.` automatically.
+
+### Fixed
+
+- **FunctionGemma export `ClassCastException` on `BufferHandle.Floats`** (#405, #420):
+  `FunctionGemmaExportHarness`, `SmolLm2ExportHarness`, and the bake-irpa tests hard-cast every
+  external constant to `BufferHandle.Owned`; since engine 0.53.0 (SKaiNET#1247) a constant can
+  arrive as the aliased `BufferHandle.Floats` instead. Both export harnesses now read every handle
+  through `DefaultBufferResolver`.
+- **`runtime-kgemma`'s Maven Central POM depended on an unpublished coordinate** (#408): pulling
+  in `:llm-runtime:kgemma3n` (never published) leaked
+  `SKaiNET-transformers.llm-runtime:kgemma3n-jvm:unspecified` into the POM, breaking resolution for
+  any external consumer. Rather than just changing the dependency's scope, Gemma 3n itself stops
+  being published (SKaiNET-transformers#377: maturity gate 0/5, hand-rolled runtime that
+  force-dequantizes the whole model, postponed by decision) — source-only until #377's maturity
+  gate is met. `runtime-kgemma`'s CLI loses its `--arch gemma3n` variant accordingly;
+  `skainet-cli` (never published) is unaffected.
+- **Android crash in the official FunctionGemma tool-call parser** (#407):
+  `FunctionGemmaOfficialToolCallParserStrategy.CALL_RE` had an unescaped closing `}` — tolerated by
+  `java.util.regex` on the JVM, rejected by Android's ICU-backed engine with a
+  `PatternSyntaxException` at class-init, so every tool-call parse crashed on ART before the first
+  match. One-character fix; no-op on the JVM.
+
 ## [0.53.0] — 2026-09-02
 
 Version lock-step with the engine is restored: this release ships against **SKaiNET 0.53.0**

README.md

@@ -108,32 +108,30 @@ Honest status — see the project-status note at the top of this README.
 
 ## Current release
 
-The current release is **0.53.0** (against **SKaiNET 0.53.0**) — version lock-step with the engine
-is restored, and the release ships everything accumulated since 0.40.2: BitNet b1.58, the
-engine-loader migration (every GGUF weight loader a thin engine wrapper, MAPPED residency by
-default), the Gemma 3n DSL path with its StableHLO/IREE export harness, and the Qwen / Apertus
-decode and tool-calling fixes.
-
-**The Gemma 3n export actually emits.** Engine 0.53.0 closes SKaiNET#1247 — allocation-free void
-tracing, aliased constant extraction, strict StableHLO conversion, and an array-free path for the
-≥2 GiB tied embedding — and the harness streams weights through the `BufferResolver` (#396). The
-full 30-layer E2B export produces a 15k-line module with zero failure comments and a 4.6 GB
-safetensors in under a minute, where it previously OOMed a 46 GB heap.
-
-**SafeTensors loading rides the engine.** The per-family hand-rolled SafeTensors materialization
-(Gemma, the shared decoder loader, Apertus, Gemma 3n) collapses onto the engine's sharded /
-single-file `ParametersLoader`s (#398, #400, #401): each family keeps only its slot table and
-allowlist, every dtype decision is the engine's, and `Require(BF16)`/`Require(FP16)` keep-native is
-accepted on the SafeTensors lane.
-
-**skainet-decode on Android.** `llm-apps:skainet-decode-core` (the shared `DecodeSession`) and
-`llm-apps:skainet-decode-android`, the repository's first Android application: load a pushed GGUF
-memory-mapped, refuse before allocating when it doesn't fit, decode on one thread, and report
-`GenerationMetrics` with the page-fault and RSS rows only a device can show (#395).
-
-**Tool calling re-verified against the new engine.** `Gemma4E2BToolCallSmokeTest` is re-enabled —
-the real E2B checkpoint now emits parseable `<|tool_call>` markup (#399) — and FunctionGemma and
-Qwen3 tool calling were re-run green in the same pass.
+The current release is **0.54.0** (against **SKaiNET 0.54.0**) — version lock-step with the engine
+continues, plus two bugs fixed on real hardware and the FunctionGemma/IREE-Android chunked-KV work
+closed out.
+
+**A stateful Android KV session for FunctionGemma.** `IreeKvSession` / `IreeKvDecoder`
+(`llm-runtime/iree-android`) prefill the tool catalog once per process, snapshot the KV state, and
+per-turn prefill only the new chunk — device-resident K/V, zero-copy tail views for the sliding
+layers, embedding rows read straight from the archive. Measured on a MagentaTV One (Mali via
+Vulkan): **p50 5.87 s / max 6.25 s per utterance**, down from minutes on the stateless redecode
+contract. Position-selected graphs (`gemma_at`/`gemma_prefill_at`) and a chunk prefill-with-past
+graph (`gemma_prefill_with_past`, #415, #417) get there without every step re-running the LM head
+over the whole sequence.
+
+**Two bugs found on real hardware, fixed.** `runtime-kgemma`'s published POM depended on an
+unpublished `:llm-runtime:kgemma3n` coordinate and failed to resolve for any external consumer —
+fixed by unpublishing Gemma 3n itself rather than papering over the dependency (it's maturity-gate
+0/5 and postponed per its own tracking issue). The official FunctionGemma tool-call parser crashed
+with `PatternSyntaxException` on every Android device (ICU rejects an unescaped `}` that the JVM's
+regex engine tolerates) — a one-character fix.
+
+It builds on **0.53.0**, which restored version lock-step with the engine (SKaiNET 0.53.0's
+billion-parameter export fixes and sharded SafeTensors loader), collapsed the per-family
+hand-rolled SafeTensors materialization onto the engine's loaders, shipped `skainet-decode` as the
+repository's first Android application, and re-verified tool calling against the new engine.
 
 It builds on **0.40.2**, which rounded out the **compiled on-device path** (standalone
 DSL → StableHLO → IREE export modules for FunctionGemma and SmolLM2, the generic Android JNI
@@ -210,7 +208,7 @@ The recommended way to consume is via the BOM. It pins every published `skainet-
 
 ```kotlin
 dependencies {
-    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.53.0"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.54.0"))
 
     // Versions resolved from the BOM:
     implementation("sk.ainet.transformers:skainet-transformers-core")

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:0.53.0"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.54.0"))
     implementation("sk.ainet.transformers:skainet-transformers-inference-moonshine")
 }
 ----

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:0.53.0"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.54.0"))
     implementation("sk.ainet.transformers:skainet-transformers-core")
     implementation("sk.ainet.transformers:skainet-transformers-runtime-kllama")
     implementation("sk.ainet.transformers:skainet-transformers-inference-llama")

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:0.53.0"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.54.0"))
 
     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>0.53.0</version>
+      <version>0.54.0</version>
       <type>pom</type>
       <scope>import</scope>
     </dependency>

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:0.53.0"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.54.0"))
     implementation("sk.ainet.transformers:skainet-transformers-providers")
 }
 ----

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:0.53.0"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.54.0"))
 
     implementation("sk.ainet.transformers:skainet-transformers-runtime-kllama")
     implementation("sk.ainet.transformers:skainet-transformers-agent")

gradle.properties

@@ -1,5 +1,5 @@
 GROUP=sk.ainet.transformers
-VERSION_NAME=0.53.0
+VERSION_NAME=0.54.0
 
 POM_DESCRIPTION=SKaiNET-transformers
 

llm-inference/gemma/README.md

@@ -4,7 +4,7 @@ Reusable **Gemma** model (incl. the FunctionGemma tool-calling fine-tune) author
 a portable graph producer with **no runtime/board/Torq code**. Pair it with the runtime module below to
 decode on-device.
 
-- **Coordinate:** `sk.ainet.transformers:skainet-transformers-inference-gemma:0.53.0`
+- **Coordinate:** `sk.ainet.transformers:skainet-transformers-inference-gemma:0.54.0`
 - **Targets:** `android`, `iosArm64`, `iosSimulatorArm64`, `macosArm64`, `linuxX64`, `linuxArm64` (broadly
   portable — mobile through server).
 - **Entry point:** `gemmaNetwork()` / `GemmaNetworkLoader` (loads a GGUF, builds the DSL graph, incl. the
@@ -24,8 +24,8 @@ FunctionGemma has a one-liner facade in `…:skainet-transformers-runtime-kgemma
 
 ```kotlin
 dependencies {
-    implementation("sk.ainet.transformers:skainet-transformers-inference-gemma:0.53.0")
-    implementation("sk.ainet.transformers:skainet-transformers-runtime-gemma-iree:0.53.0") // on-device decode
+    implementation("sk.ainet.transformers:skainet-transformers-inference-gemma:0.54.0")
+    implementation("sk.ainet.transformers:skainet-transformers-runtime-gemma-iree:0.54.0") // on-device decode
 }
 ```