Release 0.56.0

CHANGELOG.md

@@ -2,6 +2,17 @@
 
 ## [Unreleased]
 
+## [0.56.0] - 2026-09-20
+
+Headline: **grouped-query attention is native to the engine, and the compiled leg of SKEEP-005 lands —
+structure at compile time, cores at run time.** `scaledDotProductAttention` takes K/V with their own head
+count, so model code no longer tiles K/V up to the query heads before attention — not eagerly, not on
+the tape, and not in the exported StableHLO. Graph execution contexts now run under the schedule of the
+ops they wrap.
+
+Version note: 0.55.0 is skipped so the engine and SKaiNET-transformers share a version line again
+(transformers 0.55.0 was a transformers-only release against engine 0.54.0).
+
 ### Added
 
 - **SKEEP-005 phase 2 — the compiled leg: structure at compile time, cores at run time.**
@@ -15,6 +26,18 @@
   `DefaultGraphExecutionContext` answers `schedule`/`withSchedule` from the ops it wraps, so the
   JVM `ComputeGraphExecutor` runs under the caller's schedule. Key decision and diagram in SKEEP-005.
 
+### Changed
+
+- **SDPA shape validation follows the grouped-query contract**: Q heads must be a positive multiple of
+  the K/V head count (K and V must still agree). Callers that tiled K/V upstream keep working unchanged.
+- **Security pins**: `fast-uri` 3.1.6, `qs` 6.16.0 and `js-yaml` 4.3.2 (GHSA-2883-xcg3-v3hh) in the
+  JS/Wasm build-tooling graph.
+- **Docs toolchain installed with `npm ci`**: the docs image declares its Antora inputs in
+  `docs/.docker/package.json` with a committed `package-lock.json` (every transitive package pinned by
+  integrity hash, `js-yaml` overridden to 4.3.2); the docs preview job no longer downloads an unpinned
+  npm package.
+- Dependency bumps: kotest 6.2.5, kotlinx-benchmark 0.5.0, binary-compatibility-validator.
+
 ## [0.54.0] - 2026-09-06
 
 Headline: **every `ExecutionContext` gets a `Schedule` — and the CI run that exercised it found a

README.md

@@ -51,7 +51,7 @@ Add the core dependencies (Gradle Kotlin DSL):
 ```kotlin
 dependencies {
     // Recommended: import the umbrella BOM and drop versions on the engine modules.
-    implementation(platform("sk.ainet:skainet-bom:0.54.0"))
+    implementation(platform("sk.ainet:skainet-bom:0.56.0"))
 
     implementation("sk.ainet.core:skainet-lang-core")
     implementation("sk.ainet.core:skainet-backend-cpu")
@@ -308,26 +308,24 @@ val withoutLabel = dataPipeline<RawDataset>()
 
 ---
 
-## What's New in 0.54.0
-
-Structured concurrency lands as a first-class citizen — and the CI run that exercised it found a
-real deadlock:
-
-- **`Schedule` on every `ExecutionContext`** (SKEEP-005) — `sk.ainet.context.schedule.Schedule`
-  splits *what* an op computes from *how its independent chunks spread across cores*.
-  `scaledDotProductAttention` is the first scheduled op, `parallelChunks` no longer hides a
-  `runBlocking(Dispatchers.Default)` island, and the JVM `CoroutineSchedule.hardware()` default
-  spreads chunks across cores while `Schedule.Sequential` keeps every kernel single-threaded.
-- **A real deadlock, found by turning scheduling on** — `CoroutineSchedule.forRange`'s region
-  waited on children the pool had no thread left to run, once every worker was itself inside a
-  region; it looked like the OOM hang a first CI fix assumed. A region is now a shared chunk
-  queue, so a caller can always finish its own region alone, whatever the pool is doing.
-- **`tensorFilter` on the single-file `SafeTensorsParametersLoader`** — parity with the sharded
-  loader; lets a family load selectively from a checkpoint that carries tensors the requested
-  dtype can't accept.
-- **`ExperimentalMemoryApi` opt-in gate removed** — SKEEP-003's M0–M2 shipped complete back in
-  0.49.0, so `Storage`, `Scope`, `Format`, `TensorView`, `WeightForm`, and the rest of
-  `sk.ainet.lang.memory` no longer need `@OptIn`.
+## What's New in 0.56.0
+
+Grouped-query attention becomes native to the engine, and the compiled leg of SKEEP-005 lands:
+
+- **Grouped-query `scaledDotProductAttention`** — K/V arrive with their own head count
+  (`[b, nKV, Sk, hd]`); query head `h` reads K/V head `h / (H / nKV)`. Bit-identical to before when
+  `nKV == H`, and to the old tiled form otherwise. Model code no longer expands K/V heads before
+  attention — not eagerly, not on the tape.
+- **Leaner StableHLO for GQA models** — attention lowers with the head groups as a batching
+  dimension, so exports carry no broadcast or concatenate of K/V.
+- **Structure at compile time, cores at run time** — `ScheduleAnnotationPass` stamps structural
+  defaults (`parallel_dims = [batch, heads]` on every attention); an explicit `parallelism` is
+  advisory and a default can never carry a core count.
+- **Graph contexts honour the schedule of their ops** — the new `ScheduledOps` seam lets
+  `DefaultGraphExecutionContext` answer `schedule`/`withSchedule` from the ops it wraps, so the JVM
+  `ComputeGraphExecutor` runs under the caller's schedule.
+
+0.55.0 is skipped so the engine and SKaiNET-transformers share a version line again.
 
 See [CHANGELOG.md](CHANGELOG.md) for full release notes, including every prior release.
 
@@ -353,12 +351,11 @@ We love contributions! Whether it's a new operator, documentation, or a bug fix:
 
 Browse the full codebase documentation on [DeepWiki](https://deepwiki.com/SKaiNET-developers/SKaiNET).
 
-### Contributors (0.54.0)
+### Contributors (0.56.0)
 
-- **Michal Harakal** ([@michalharakal](https://github.com/michalharakal)) — the SKEEP-005
-  structured-concurrency Schedule API, the coroutine-pool deadlock it uncovered on CI and its
-  shared-chunk-queue fix, `SafeTensorsParametersLoader` `tensorFilter` parity, and retiring the
-  `ExperimentalMemoryApi` opt-in gate now that SKEEP-003 has shipped
+- **Michal Harakal** ([@michalharakal](https://github.com/michalharakal)) — grouped-query-native
+  `scaledDotProductAttention` and its StableHLO lowering, structural schedule defaults, and
+  schedule-aware graph execution contexts (SKEEP-005 phase 2)
 
 ### Contributors (0.53.0)
 

docs/antora.yml

@@ -15,7 +15,7 @@ asciidoc:
     framework_name: SKaiNET
     # Current SKaiNET release — bump once per release; referenced as
     # {skainet_version} in dependency snippets (blocks need subs="attributes+").
-    skainet_version: 0.54.0
+    skainet_version: 0.56.0
     ksp_version: 2.2.21-2.0.5
     dokka_version: 2.1.0
     asciidoctorj_version: 3.0.0

docs/modules/ROOT/pages/reference/kernel-support-matrix.adoc

@@ -1,7 +1,7 @@
 = Kernel × platform support matrix
 :description: Which compute-kernel provider serves each weight format on each KMP target.
 
-Generated from `kernel-support.json` (version `0.54.0`) by `KernelSupportMatrixTest` — registry introspection of the registered `KernelProvider` implementations. Do not edit by hand; run `./gradlew generateKernelMatrix` to refresh.
+Generated from `kernel-support.json` (version `0.56.0`) by `KernelSupportMatrixTest` — registry introspection of the registered `KernelProvider` implementations. Do not edit by hand; run `./gradlew generateKernelMatrix` to refresh.
 
 Each cell is the best (highest-priority) provider that serves `Float32 × format` `matmul` on that platform: *native-ffm* (100) → *panama-vector* (50) → *scalar* (0). An empty cell (`—`) means no provider carries a kernel there (the format is dequant-to-FP32 only).
 

docs/modules/ROOT/pages/reference/operators/generated/index.adoc

@@ -1,6 +1,6 @@
 = AI-NET Operators Reference
 
-Generated from version `0.54.0` on 2026-09-06
+Generated from version `0.56.0` on 2026-09-20
 
 == Operators by Modality
 

docs/modules/ROOT/pages/reference/ops-status-matrix.adoc

@@ -1,7 +1,7 @@
 = Operator Coverage Matrix
 :description: Cross-backend status for every operator function in SKaiNET.
 
-Generated from `operators.json` version `0.54.0` on 2026-09-06.
+Generated from `operators.json` version `0.56.0` on 2026-09-20.
 
 Rows are `Operator.function` pairs. The `Validated` column shows whether the function's documentation has been DARC-validated by a reviewer (see xref:contributing/darc-workflow.adoc[DARC workflow]). Remaining columns are backends that appear in any function's `statusByBackend` map — a missing entry means the backend makes no claim about the function (treat it as "unknown", not "not supported").
 

docs/modules/ROOT/pages/tutorials/android-classifier-getting-started.adoc

@@ -19,12 +19,12 @@ An Android project with Kotlin. Add the SKaiNET modules:
 [source,kotlin]
 ----
 dependencies {
-    implementation("sk.ainet.core:skainet-lang-core:0.54.0")     // tensors, DSL, training
-    implementation("sk.ainet.core:skainet-backend-cpu:0.54.0")   // CPU ops
-    implementation("sk.ainet.core:skainet-compile-dag:0.54.0")   // autograd (training context)
-    implementation("sk.ainet.core:skainet-data-api:0.54.0")      // Dataset / DataBatch
-    implementation("sk.ainet.core:skainet-data-simple:0.54.0")   // embedded Iris
-    runtimeOnly("sk.ainet.core:skainet-backend-jni-cpu:0.54.0")  // NEON kernels (see below)
+    implementation("sk.ainet.core:skainet-lang-core:0.56.0")     // tensors, DSL, training
+    implementation("sk.ainet.core:skainet-backend-cpu:0.56.0")   // CPU ops
+    implementation("sk.ainet.core:skainet-compile-dag:0.56.0")   // autograd (training context)
+    implementation("sk.ainet.core:skainet-data-api:0.56.0")      // Dataset / DataBatch
+    implementation("sk.ainet.core:skainet-data-simple:0.56.0")   // embedded Iris
+    runtimeOnly("sk.ainet.core:skainet-backend-jni-cpu:0.56.0")  // NEON kernels (see below)
 }
 ----
 

gradle.properties

@@ -1,5 +1,5 @@
 GROUP=sk.ainet.core
-VERSION_NAME=0.54.0
+VERSION_NAME=0.56.0
 POM_DESCRIPTION=SKaiNET
 
 POM_URL=https://github.com/SKaiNET-developers/skainet/