Skip to content

release: 0.32.0 — GroupNorm StableHLO converter #753

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
michalharakal wants to merge 1 commit into from
Closed

release: 0.32.0 — GroupNorm StableHLO converter #753

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 22 additions & 0 deletions CHANGELOG.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,28 @@

## [Unreleased]

## [0.32.0] - 2026-06-22

### Added

- **GroupNorm StableHLO converter.** `NeuralNetOperationsConverter` now lowers `groupNorm`
(and the `groupNormalization` / `GroupNormalization` / `group_norm` aliases) to real
`stablehlo.*` ops instead of falling through to the "operation not supported" path. The
lowering mirrors the existing LayerNorm/RMSNorm decomposition: reshape `(N, C, *spatial)`
to `(N, G, M)` so each group's channels and spatial positions collapse into one trailing
axis, take per-group `mean` / `variance` over that axis (reusing the `@reduce_mean` /
`@reduce_variance` custom_calls), normalize `(x - mean) / sqrt(var + eps)`, reshape back,
and apply the optional per-channel `scale` / `offset` (shape `C`, broadcast over the channel
dim). Adds `GroupNormConverterTest` (commonTest). (PR #752)
- **SKEEP proposals docs module.** (PR #750)
- **Quantization-process explanation doc** (weights, activations, calibration). (PR #747)

### Changed

- **Dependency bumps:** `com.vanniktech.maven.publish` → 0.37.0 (PR #748),
`com.networknt:json-schema-validator` → 3.0.5 (PR #749), kotest → 6.2.1 (PR #744),
Gradle wrapper → 9.6.0 (PR #745), `actions/checkout` → 7 (PR #743).

## [0.31.2] - 2026-06-18

### Added
Expand Down
23 changes: 17 additions & 6 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,7 +36,7 @@ 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:0.31.2"))
implementation(platform("sk.ainet:skainet-bom:0.32.0"))

implementation("sk.ainet.core:skainet-lang-core")
implementation("sk.ainet.core:skainet-backend-cpu")
Expand Down Expand Up @@ -241,6 +241,20 @@ Runnable examples:

---

## What's New in 0.32.0

- **GroupNorm StableHLO converter.** `groupNorm` now lowers to real `stablehlo.*` ops —
reshape `(N, C, *spatial)` into `(N, G, M)` groups → per-group `mean`/`variance` via the
`@reduce_mean` / `@reduce_variance` custom_calls → normalize → reshape back → optional
per-channel `scale`/`offset` — mirroring the LayerNorm/RMSNorm decomposition. Previously a
`groupNorm` node fell through to the "operation not supported" path. Adds
`GroupNormConverterTest`. (PR #752)
- **Docs:** a SKEEP proposals module (PR #750) and a quantization-process explanation —
weights, activations, calibration (PR #747).
- **Dependencies:** `com.vanniktech.maven.publish` → 0.37.0 (#748),
`com.networknt:json-schema-validator` → 3.0.5 (#749), kotest → 6.2.1 (#744),
Gradle wrapper → 9.6.0 (#745), `actions/checkout` → 7 (#743).

## What's New in 0.31.2

- **`RowDequantSource` + `ops.gather` row-dequant.** A `TensorData` can now mark itself `RowDequantSource`
Expand All @@ -250,13 +264,10 @@ Runnable examples:
packed and be looked up via `ops.gather` directly — moving the per-row-dequant trick out of model code
into the engine. (PR #741)

## What's New in 0.31.0

- **`ops.transpose` lazily handles every packed matmul dtype.** The CPU backend rewraps packed bytes with a flipped shape (metadata-only "lazy transpose") so a packed weight survives `linearProject`'s `matmul(x, transpose(W))` instead of inflating to FP32 — but **Q8_0 and Q4_0** were missing and threw `Byte → Float ClassCastException`. Now the full dispatch set (Q4_K/Q5_K/Q6_K/Q5_0/Q5_1/Q8_0/Q4_0) transposes lazily, so a packed Q8_0/Q4_0 matmul weight (e.g. a tied Q8_0 `lm_head`) stays packed end-to-end on its NEON/SIMD kernel. Regression-tested across all seven packed types. (PRs #736, #737)
- **Dependency:** `com.networknt:json-schema-validator` → 3.0.4. (PR #733)

### Recent releases

- **0.31.0** — `ops.transpose` lazily handles every packed matmul dtype: **Q8_0 and Q4_0** added (they threw `Byte → Float ClassCastException`), completing the Q4_K/Q5_K/Q6_K/Q5_0/Q5_1/Q8_0/Q4_0 set so packed weights stay packed end-to-end on their NEON/SIMD kernel; plus `com.networknt:json-schema-validator` → 3.0.4. (PRs #736, #737, #733)

- **0.30.0** — First-class **Q5_K packed in-kernel dequant-matmul** across the CPU backends (`Q5_KBlockTensorData` + `Q5KMatmulKernel` SPI: scalar / Panama Vector / native-C), **hand-written ARM NEON kernels** (fp32/q8_0/q4k/q5k, `-march=armv8.2-a+fp16+dotprod`), and **Kotlin/Native consumption of the C kernels via cinterop** (`skainet-backend-native-cpu` static archive + `linuxX64`/`linuxArm64` `KernelProvider`). (PR #734)

- **0.29.1** — `sk.ainet.core:skainet-compile-minerva` now publishes to Maven Central (packaging fix for the Minerva export module shipped in 0.29.0).
Expand Down
2 changes: 1 addition & 1 deletion gradle.properties
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
GROUP=sk.ainet.core
VERSION_NAME=0.31.2
VERSION_NAME=0.32.0
POM_DESCRIPTION=SKaiNET

POM_URL=https://github.com/SKaiNET-developers/skainet/
Expand Down
Loading