SKaiNET-embedded-vendors

Vendor-specific SKaiNET plugins that live outside the agnostic SKaiNET core and the model definitions. Each plugin registers a hardware target's optimizations with the SKaiNET TargetOptimizers registry; an application selects a target and installs the matching plugin at its composition root.

synaptics-torq — Synaptics Torq NPU (SL2610)

The vendor optimization handler for the "torq" target.

• DAG passes (registered via TorqPlugin.install()): attention head/query-sequence tiling and FFN hidden-dim tiling, so the batched matmuls fit the NPU's on-chip layout.
• Compile-time flags (TorqPlugin.compileFlags): the torq-compile knobs the build tool applies — notably --torq-disable-slices, required to avoid a buffer-aliasing blow-up in large graphs.
• Optional bf16-trace workarounds (present, not registered by default): TorqF32ReducePass, TorqMatmulBf16Pass.

Note: the SL2610 v2.0.0 compiler (with --torq-disable-slices) compiles the Moonshine encoder without the attention/FFN tiling — the tiling once worked around a MatMulPattern.cpp:57 assertion that the newer compiler no longer hits. The passes are kept here for other models / older toolchains, but Moonshine's Torq path no longer requires them.

Use (from the application / build tool)

import sk.ainet.vendors.torq.TorqPlugin

// once, at the composition root — the app has chosen the Torq target:
TorqPlugin.install()                 // registers the "torq" DAG passes
// ... trace model -> ComputeGraph ...
val tiled = dagPipelineFor(TorqPlugin.TARGET).optimize(graph).graph
val mlir  = toStableHlo(tiled, "encoder").content
// then invoke torq-compile with TorqPlugin.compileFlags (+ -o out.vmfb)

The SKaiNET core never names "torq"; it only sees the target string and asks the registry which passes to run. This keeps all Synaptics specifics out of core and out of the model.

Dependency layering

app (e.g. SKaiNET-embedded/sl2610-function-calling)
 ├── moonshine-dsl model  → transitive: skainet core          [HW-agnostic]
 └── synaptics-torq plugin (this repo) → skainet-compile-opt   [the TargetOptimizer API only]
       app calls TorqPlugin.install() once

Consume

Published to Maven Central under sk.ainet.vendors:

dependencies {
    implementation("sk.ainet.vendors:synaptics-torq:0.1.0")
}

Build

Self-contained — depends on the published sk.ainet.core:* artifacts (Maven Central, version pinned in gradle/libs.versions.toml), which include the TargetOptimizer / pluggable-optimization mechanism.

./gradlew :synaptics-torq:compileKotlinJvm

To develop against a local SKaiNET checkout instead, uncomment includeBuild("../SKaiNET") in settings.gradle.kts (Gradle substitutes the sk.ainet.core:* deps with the local projects).