Merge pull request #699 from SKaiNET-developers/release/0.28.1

Release/0.28.1

Author: michalharakal

CHANGELOG.md

@@ -2,6 +2,25 @@
 
 ## [Unreleased]
 
+## [0.28.1] - 2026-06-06
+
+### Fixed
+
+- **DAG-DSL StableHLO export now compiles end-to-end with IREE for the full conformance suite (7/7 models, 27/27 ops).** Shape-changing ops declared a result/return type inferred from operand-0 instead of the op's real output, so `iree-compile` rejected the modules (*"inferred shape … is incompatible with return type"*). `DagBuilder.inferDagOutputSpecs` now computes the correct output spec for:
+  - **`reshape`/`view`** — reads the target shape from the op's `newShape` parameter (a `Shape`, which the converter's `as? List<Int>` had missed). (#673, PR #674)
+  - **`matmul`/`dot`/`mm`/`bmm`** — `(…, M, K) @ (…, K, N) → (…, M, N)` instead of echoing operand-0. (#673, PR #674)
+  - **`concatenate`** — the corrected summed-axis extent propagates to the consumers and the `func.func` return type, not just the op line. (#673, PR #674)
+  - **`conv1d`** — windowed `(N, Cout, Lout)`; `conv2d` already inferred via `Conv2dOperation`, but `conv1d` was a `GenericOperation`. (#675, PR #676)
+  - **`gather`** — `table[:axis] ⊕ indices.shape ⊕ table[axis+1:]`. (#675, PR #676)
+  - **`maxpool2d`/`avgpool2d`** — windowed `(N, C, Hout, Wout)`. (#675, PR #676)
+  - **`flatten`** — collapses `[startDim..endDim]` while preserving the leading batch dim (it was collapsing everything to rank-1, breaking the dense layer in mnist-cnn). (#675, PR #676)
+- **`reduce_window` is emitted in IREE's parseable generic region form.** Pooling previously used the pretty `… applies <op> over window …` form, which IREE rejects (*"has no custom assembly form"*). Now emits `"stablehlo.reduce_window"(…) ({ ^bb0(…): … })` with full NCHW-rank window attributes; average pooling's divisor is splatted to the output type (was a scalar-vs-tensor mismatch). (#675, PR #676)
+- **`MlirValidator` understands region block arguments.** It now registers `^bb0(%a, %b)` block-argument SSA definitions and every `%x =` result on a line, so single-line region ops (e.g. `reduce_window`) validate. (PR #676)
+
+### Changed
+
+- Regenerated the JVM binary-compatibility baselines (`apiDump`) to match the public API exposed since 0.27.0 (`AttentionOperationsConverter`, multi-output port helpers, `KClass` dtype constructors).
+
 ## [0.28.0] - 2026-06-06
 
 ### Fixed

README.md

@@ -35,7 +35,7 @@ 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:0.28.0"))
+    implementation(platform("sk.ainet:skainet-bom:0.28.1"))
 
     implementation("sk.ainet.core:skainet-lang-core")
     implementation("sk.ainet.core:skainet-backend-cpu")
@@ -193,13 +193,15 @@ deployment, the StableHLO path for native and edge targets.
 
 ---
 
-## What's New in 0.28.0
+## What's New in 0.28.1
 
-- **Four StableHLO export bugs fixed — the Kotlin DSL → MLIR path lowers reshape, concatenate, constants, and reductions correctly.** `reshape` whose target shape lives only in an op parameter now lowers to a typed `stablehlo.reshape` (#666); multi-input `concatenate` sums the operands' extents on the concatenated axis instead of echoing operand-0's (`1×1 + 1×4 + 1×1` on dim 1 → `1×6`, not `1×1`) (#667); DAG constants are inlined into the module and reductions drop the reduced dimension (#663); and `HloGenerator`'s forward-pass tracing now binds the sample input and synthesizes external inputs so it emits real ops instead of a structure-only module (#668). (PRs #664, #670)
-- **Non-JVM image runtime support.** Image and data-transform modules are scoped to their supported KMP targets, with a non-JVM image runtime so the image/data-transform APIs build honestly across targets. (PR #671)
+- **The Kotlin DSL → StableHLO → IREE path is green end-to-end for the whole conformance suite — 7/7 models and 27/27 ops `iree-compile` to a `vmfb`** (grayscale, tiny-mlp, whisper, mnist-cnn, yolo, leaf-embed, gemma3-260m). Shape-changing ops had been declaring a result/return type from operand-0 instead of their real output, so IREE rejected the modules. `inferDagOutputSpecs` now computes the correct output spec for `reshape`, `matmul`, `concatenate` (#673), and `conv1d`, `gather`, `maxpool2d`/`avgpool2d`, `flatten` (#675), and the corrected shapes flow to both the op result type and the `func.func` return.
+- **`reduce_window` (pooling) is emitted in IREE's generic region form** instead of the pretty `applies … over window` form IREE rejects; the `MlirValidator` now understands region block arguments. (#675, PR #676)
+- Verified end-to-end by publishing each candidate to mavenLocal and running the `skainet-iree-conformance` harness before release.
 
 ### Recent releases
 
+- **0.28.0** — Four StableHLO export bugs fixed (reshape #666, concatenate #667, constants/reductions #663, `HloGenerator` tracing #668) plus non-JVM image runtime support (#671). (PRs #664, #670, #671)
 - **0.27.0** — A full gemma3 network lowers to StableHLO and compiles to an IREE `vmfb` (zero op gaps, verified by `GemmaTraceTest`): new `scaledDotProductAttention` (with causal + explicit additive mask), `permute`, `narrow`, and multi-output `split` converters, plus boxing-free `FloatArray` weight externalization for `.irpa` baking. (PRs #661 et al.)
 - **0.26.0** — Q4_0 promoted to a first-class quantized format across the provider stack, `tanh` as a first-class activation primitive, and a CPU tensor `convert` op, plus test/build/CI hygiene. (PRs #648–#651, #631, #636)
 - **0.25.0** — BF16 and Q8_0 matmul kernels end-to-end across the provider stack, autograd completeness for `pow`/`log` and the conv/pool/upsample/split family, the hybrid adaptive dtype-constraint DSL, the `@DarcValidated` operator-doc flag, and the SentencePiece special-token splitter. (PRs #595, #605–#628)

gradle.properties

@@ -1,5 +1,5 @@
 GROUP=sk.ainet.core
-VERSION_NAME=0.28.0
+VERSION_NAME=0.28.1
 POM_DESCRIPTION=SKaiNET
 
 POM_URL=https://github.com/SKaiNET-developers/skainet/

skainet-lang/skainet-lang-core/api/jvm/skainet-lang-core.api

@@ -2019,8 +2019,8 @@ public final class sk/ainet/lang/nn/metrics/MetricKt {
 }
 
 public final class sk/ainet/lang/nn/normalization/BatchNormalization : sk/ainet/lang/nn/Module, sk/ainet/lang/nn/topology/ModuleParameters {
-	public fun <init> (IDDZLjava/lang/String;Lsk/ainet/lang/tensor/Tensor;Lsk/ainet/lang/tensor/Tensor;)V
-	public synthetic fun <init> (IDDZLjava/lang/String;Lsk/ainet/lang/tensor/Tensor;Lsk/ainet/lang/tensor/Tensor;ILkotlin/jvm/internal/DefaultConstructorMarker;)V
+	public fun <init> (IDDZLjava/lang/String;Lsk/ainet/lang/tensor/Tensor;Lsk/ainet/lang/tensor/Tensor;Lkotlin/reflect/KClass;)V
+	public synthetic fun <init> (IDDZLjava/lang/String;Lsk/ainet/lang/tensor/Tensor;Lsk/ainet/lang/tensor/Tensor;Lkotlin/reflect/KClass;ILkotlin/jvm/internal/DefaultConstructorMarker;)V
 	public final fun eval ()V
 	public fun forward (Lsk/ainet/lang/tensor/Tensor;Lsk/ainet/context/ExecutionContext;)Lsk/ainet/lang/tensor/Tensor;
 	public fun getModules ()Ljava/util/List;
@@ -2030,17 +2030,17 @@ public final class sk/ainet/lang/nn/normalization/BatchNormalization : sk/ainet/
 }
 
 public final class sk/ainet/lang/nn/normalization/GroupNormalization : sk/ainet/lang/nn/Module, sk/ainet/lang/nn/topology/ModuleParameters {
-	public fun <init> (IIDZLjava/lang/String;Lsk/ainet/lang/tensor/Tensor;Lsk/ainet/lang/tensor/Tensor;)V
-	public synthetic fun <init> (IIDZLjava/lang/String;Lsk/ainet/lang/tensor/Tensor;Lsk/ainet/lang/tensor/Tensor;ILkotlin/jvm/internal/DefaultConstructorMarker;)V
+	public fun <init> (IIDZLjava/lang/String;Lsk/ainet/lang/tensor/Tensor;Lsk/ainet/lang/tensor/Tensor;Lkotlin/reflect/KClass;)V
+	public synthetic fun <init> (IIDZLjava/lang/String;Lsk/ainet/lang/tensor/Tensor;Lsk/ainet/lang/tensor/Tensor;Lkotlin/reflect/KClass;ILkotlin/jvm/internal/DefaultConstructorMarker;)V
 	public fun forward (Lsk/ainet/lang/tensor/Tensor;Lsk/ainet/context/ExecutionContext;)Lsk/ainet/lang/tensor/Tensor;
 	public fun getModules ()Ljava/util/List;
 	public fun getName ()Ljava/lang/String;
 	public fun getParams ()Ljava/util/List;
 }
 
 public final class sk/ainet/lang/nn/normalization/LayerNormalization : sk/ainet/lang/nn/Module, sk/ainet/lang/nn/topology/ModuleParameters {
-	public fun <init> ([IDZLjava/lang/String;Lsk/ainet/lang/tensor/Tensor;Lsk/ainet/lang/tensor/Tensor;)V
-	public synthetic fun <init> ([IDZLjava/lang/String;Lsk/ainet/lang/tensor/Tensor;Lsk/ainet/lang/tensor/Tensor;ILkotlin/jvm/internal/DefaultConstructorMarker;)V
+	public fun <init> ([IDZLjava/lang/String;Lsk/ainet/lang/tensor/Tensor;Lsk/ainet/lang/tensor/Tensor;Lkotlin/reflect/KClass;)V
+	public synthetic fun <init> ([IDZLjava/lang/String;Lsk/ainet/lang/tensor/Tensor;Lsk/ainet/lang/tensor/Tensor;Lkotlin/reflect/KClass;ILkotlin/jvm/internal/DefaultConstructorMarker;)V
 	public fun forward (Lsk/ainet/lang/tensor/Tensor;Lsk/ainet/context/ExecutionContext;)Lsk/ainet/lang/tensor/Tensor;
 	public fun getModules ()Ljava/util/List;
 	public fun getName ()Ljava/lang/String;