You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Apple half of the mobile-kernels story (engine SKaiNET#920 / #959; the Android half landed via #285). Engine develop now publishes skainet-backend-native-cpu for iosArm64 / iosSimulatorArm64 / macosArm64 (engine PR #962) with the kernel archives embedded in the klibs and runtime FEAT_DotProd dispatch (engine #958: A13+/M-series fast path, A12 scalar-int fallback — one archive, no SIGILL).
Registration on Kotlin/Native is manual (no ServiceLoader): a consumer calls installNativeKernels() once at startup. Today no transformers module does, so Apple K/N consumers of the runtime facades run packed-quant matmul on the scalar floor — the field report's 0.11 tok/s iOS simulator number.
README target-matrix note (Apple gets native packed-quant kernels through the facades out of the box; direct inference-llama consumers call installNativeKernels() themselves).
Numbers: run the existing cross-target SmolLm2InferenceSpike on iosSimulatorArm64Test (SIMCTL_CHILD_ env forwarding is already wired) on a macOS host, and physical-device numbers via the field-report author's standing test offer — A13+ fast arm; A12 no-SIGILL check if an XS/XR is available.
Blocked on: the next engine release publishing the Apple klibs to Maven Central (engine #964 adds the release plumbing). Until then the wiring can only be compile-verified on a macOS host via the composite build — this box (Linux) cannot build the engine's Apple cinterop.
Apple half of the mobile-kernels story (engine SKaiNET#920 / #959; the Android half landed via #285). Engine develop now publishes
skainet-backend-native-cpuforiosArm64/iosSimulatorArm64/macosArm64(engine PR #962) with the kernel archives embedded in the klibs and runtime FEAT_DotProd dispatch (engine #958: A13+/M-series fast path, A12 scalar-int fallback — one archive, no SIGILL).Registration on Kotlin/Native is manual (no ServiceLoader): a consumer calls
installNativeKernels()once at startup. Today no transformers module does, so Apple K/N consumers of the runtime facades run packed-quant matmul on the scalar floor — the field report's 0.11 tok/s iOS simulator number.Plan (mirrors the Android JNI wiring from #285):
llm-runtime/kllama+llm-runtime/kgemma: addsk.ainet.core:skainet-backend-native-cputo the Apple source sets and callinstallNativeKernels()from the existingregisterPlatformBackendsactual (kllama'ssrc/iosMainhas exactly this seam — it was dead code before feat(kllama,kgemma): add iosArm64/iosSimulatorArm64 targets; document the target matrix (#271) #274 added the targets).inference-llamaconsumers callinstallNativeKernels()themselves).SmolLm2InferenceSpikeoniosSimulatorArm64Test(SIMCTL_CHILD_ env forwarding is already wired) on a macOS host, and physical-device numbers via the field-report author's standing test offer — A13+ fast arm; A12 no-SIGILL check if an XS/XR is available.Blocked on: the next engine release publishing the Apple klibs to Maven Central (engine #964 adds the release plumbing). Until then the wiring can only be compile-verified on a macOS host via the composite build — this box (Linux) cannot build the engine's Apple cinterop.