Remove ExperimentalMemoryApi opt-in gate (SKEEP-003)

CHANGELOG.md

@@ -31,6 +31,15 @@
   custom quantized payloads) instead of failing on the first unmapped one. Threaded through
   `withPolicy`.
 
+### Changed
+
+- **`ExperimentalMemoryApi` opt-in gate removed** (SKEEP-003): the `sk.ainet.lang.memory` API
+  (`Storage`, `Scope`, `Format`, `Layout`, `TensorView`, `WeightForm`, `WeightByteOrder`, …) no
+  longer requires `@OptIn(ExperimentalMemoryApi::class)` to use. The annotation's own message —
+  "usable, but may change until milestone M1 is complete" — was stale: M0/M1/M2 all shipped
+  complete in 0.49.0. `ExperimentalMemoryApi` is deleted along with every `@OptIn`/
+  `@ExperimentalMemoryApi` annotation referencing it.
+
 ## [0.53.0] - 2026-09-02
 
 Headline: **the export pipeline emits billion-parameter models.** Tracing a 4.5B-parameter

docs/modules/ROOT/examples/kotlin/sk/ainet/docs/samples/ScheduleDemo.kt

@@ -6,7 +6,6 @@ import sk.ainet.context.ExecutionContext
 import sk.ainet.context.schedule.Schedule
 import sk.ainet.context.withSchedule
 import sk.ainet.exec.schedule.CoroutineSchedule
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 import sk.ainet.lang.memory.trace.RecordingTraceSink
 import sk.ainet.lang.memory.trace.TraceEvent
 import sk.ainet.lang.tensor.Shape
@@ -19,7 +18,6 @@ import sk.ainet.lang.types.FP32
  * the trace that says which schedule ran. Every region is included verbatim into
  * `tutorials/schedule-getting-started.adoc`; `SamplesTest` executes it.
  */
-@OptIn(ExperimentalMemoryApi::class)
 object ScheduleDemo {
 
     class Result(

skainet-apps/skainet-plan/src/main/kotlin/sk/ainet/apps/plan/SkainetPlan.kt

@@ -1,4 +1,3 @@
-@file:OptIn(ExperimentalMemoryApi::class)
 
 package sk.ainet.apps.plan
 
@@ -8,7 +7,6 @@ import kotlinx.cli.default
 import sk.ainet.io.JvmRandomAccessSource
 import sk.ainet.io.gguf.StreamingGGUFReader
 import sk.ainet.io.gguf.planInput
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 import sk.ainet.lang.memory.plan.Budget
 import sk.ainet.lang.memory.plan.KvCacheMode
 import sk.ainet.lang.memory.plan.MemoryPlan

skainet-apps/skainet-plan/src/test/kotlin/sk/ainet/apps/plan/SkainetPlanTest.kt

@@ -1,8 +1,6 @@
-@file:OptIn(ExperimentalMemoryApi::class)
 
 package sk.ainet.apps.plan
 
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 import sk.ainet.lang.memory.Format
 import sk.ainet.lang.memory.plan.Budget
 import sk.ainet.lang.memory.plan.MemoryPlans

skainet-backends/benchmarks/jvm-cpu-publish/src/main/kotlin/sk/ainet/bench/publish/scenarios/TernaryF32GemvScenario.kt

@@ -5,7 +5,6 @@ import sk.ainet.backend.api.kernel.TernaryF32GemvKernel
 import sk.ainet.backend.api.kernel.TernaryKernelPacks
 import sk.ainet.bench.publish.runner.Scenario
 import sk.ainet.exec.kernel.NativeTernaryF32GemvKernel
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 import sk.ainet.lang.memory.Scope
 import sk.ainet.lang.memory.Storage
 import sk.ainet.lang.memory.TensorView
@@ -31,7 +30,6 @@ import kotlin.random.Random
  * Dims default to the BitNet-2B FFN projection (k=2560, n=6912); the n=6912 regime also crosses
  * the LUT kernel's internal pthread threshold. Primary metric: GOP/s over `2·k·n` ops.
  */
-@OptIn(ExperimentalMemoryApi::class)
 internal class TernaryF32GemvScenario(
     smoke: Boolean,
     private val providerName: String,

skainet-backends/benchmarks/jvm-cpu-publish/src/main/kotlin/sk/ainet/bench/publish/schema/GenerationMetricsRecord.kt

@@ -2,7 +2,6 @@ package sk.ainet.bench.publish.schema
 
 import kotlinx.serialization.SerialName
 import kotlinx.serialization.Serializable
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 import sk.ainet.lang.memory.trace.GenerationMetrics
 
 /**
@@ -49,7 +48,6 @@ public data class GenerationMetricsRecord(
 )
 
 /** This run's metrics as the record the benchmark JSON carries. */
-@OptIn(ExperimentalMemoryApi::class)
 public fun GenerationMetrics.toRecord(): GenerationMetricsRecord = GenerationMetricsRecord(
     prefillTokens = prefillTokens,
     prefillTokensPerSecond = prefillTokensPerSecond,

skainet-backends/benchmarks/jvm-cpu-publish/src/test/kotlin/sk/ainet/bench/publish/schema/GenerationMetricsRecordTest.kt

@@ -1,7 +1,6 @@
 package sk.ainet.bench.publish.schema
 
 import kotlinx.serialization.json.Json
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 import sk.ainet.lang.memory.trace.GenerationMetrics
 import sk.ainet.lang.memory.trace.ModuleCost
 import java.io.File
@@ -17,7 +16,6 @@ import kotlin.test.assertTrue
  * The written fixture is the script's own test input: `./scripts/check_engine_json.sh
  * skainet-backends/benchmarks/jvm-cpu-publish/build/engine-json-check` must pass on it.
  */
-@OptIn(ExperimentalMemoryApi::class)
 class GenerationMetricsRecordTest {
 
     private val json = Json { prettyPrint = true; encodeDefaults = true; explicitNulls = false }

skainet-backends/skainet-backend-api/src/androidMain/kotlin/sk/ainet/backend/api/kernel/DispatchMode.android.kt

@@ -1,6 +1,4 @@
 package sk.ainet.backend.api.kernel
 
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 
-@ExperimentalMemoryApi
 internal actual fun platformUseRegistry(): Boolean = System.getProperty(DispatchMode.PROPERTY) != "false"

skainet-backends/skainet-backend-api/src/androidMain/kotlin/sk/ainet/backend/api/kernel/ViewKernelPack.android.kt

@@ -1,14 +1,12 @@
 package sk.ainet.backend.api.kernel
 
 import java.util.ServiceLoader
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 
 /**
  * Android discovery for [ViewKernelPack]. `ServiceLoader` exists on Android, so the JNI packs a
  * consumer ships (e.g. the NEON row-major pack in `skainet-backend-jni-cpu`) are discovered the
  * same way as on the JVM, provided the packaging step keeps `META-INF/services` entries.
  */
-@ExperimentalMemoryApi
 internal actual fun installPlatformKernelPacks(): List<String> =
     runCatching {
         ServiceLoader.load(ViewKernelPack::class.java)
@@ -21,7 +19,6 @@ internal actual fun installPlatformKernelPacks(): List<String> =
  * `jvmMain`, which the Android source set does not see. Same two steps it performs — discover, then
  * register, letting [KernelRegistry] sort by priority on insertion.
  */
-@ExperimentalMemoryApi
 internal actual fun installPlatformKernelProviders(): List<String> =
     runCatching {
         ServiceLoader.load(KernelProvider::class.java)

skainet-backends/skainet-backend-api/src/commonMain/kotlin/sk/ainet/backend/api/kernel/BitNetGemvKernel.kt

@@ -1,6 +1,5 @@
 package sk.ainet.backend.api.kernel
 
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 import sk.ainet.lang.memory.Format
 import sk.ainet.lang.memory.I8Absmax
 import sk.ainet.lang.memory.Storage
@@ -25,7 +24,6 @@ import sk.ainet.lang.tensor.storage.TensorEncoding
  * The weight's codes are read once per call, not once per row: a decode step is one row against
  * the whole matrix, so hoisting it is the difference between O(rows·n·k) decodes and O(n·k).
  */
-@ExperimentalMemoryApi
 public class BitNetGemvKernel(override val key: KernelKey) : ViewKernel {
 
     override val name: String get() = "bitnet_gemv/reference"

skainet-backends/skainet-backend-api/src/commonMain/kotlin/sk/ainet/backend/api/kernel/DispatchMode.kt

@@ -1,6 +1,5 @@
 package sk.ainet.backend.api.kernel
 
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 
 /**
  * Whether the generic (non-fast-path) matmul goes through [KernelDispatch] or the legacy
@@ -11,7 +10,6 @@ import sk.ainet.lang.memory.ExperimentalMemoryApi
  * construction* instead of a `ClassCastException` (#993). The legacy path stays one flag away
  * while the migration settles; it is deleted once the golden parity and benchmark evidence is in.
  */
-@ExperimentalMemoryApi
 public object DispatchMode {
     /** Set to `false` (`skainet.dispatch.registry=false`) to force the legacy generic fallback. */
     public const val PROPERTY: String = "skainet.dispatch.registry"
@@ -24,5 +22,4 @@ public object DispatchMode {
 }
 
 /** Platform reading of [DispatchMode.PROPERTY]; defaults to `true` where there is no property store. */
-@ExperimentalMemoryApi
 internal expect fun platformUseRegistry(): Boolean

skainet-backends/skainet-backend-api/src/commonMain/kotlin/sk/ainet/backend/api/kernel/KernelDispatch.kt

@@ -1,7 +1,6 @@
 package sk.ainet.backend.api.kernel
 
 import sk.ainet.lang.memory.BlockOrder
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 import sk.ainet.lang.memory.Format
 import sk.ainet.lang.memory.Scope
 import sk.ainet.lang.memory.TensorView
@@ -23,7 +22,6 @@ import sk.ainet.lang.tensor.storage.TensorEncoding
  * Adapters allocate in the caller's [Scope] (a `Forward` scope in a generation loop) and are
  * emitted as [TraceEvent.AdapterInserted] — the "hidden 12 GB" of #782 becomes a visible event.
  */
-@ExperimentalMemoryApi
 public object KernelDispatch {
 
     /**
@@ -255,7 +253,6 @@ public object KernelDispatch {
 }
 
 /** A view of the same contiguous bytes under a different shape (rule 5: reshape is a view). */
-@ExperimentalMemoryApi
 public fun TensorView.reshapeContiguous(newShape: Shape): TensorView {
     require(isContiguous) { "reshape needs a contiguous view" }
     require(newShape.volume.toLong() == elementCount) { "reshape must keep the element count ($elementCount), got ${newShape.volume}" }

skainet-backends/skainet-backend-api/src/commonMain/kotlin/sk/ainet/backend/api/kernel/KernelKey.kt

@@ -1,6 +1,5 @@
 package sk.ainet.backend.api.kernel
 
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 import sk.ainet.lang.memory.Format
 import sk.ainet.lang.memory.TensorView
 import sk.ainet.lang.tensor.storage.TensorEncoding
@@ -14,7 +13,6 @@ import sk.ainet.lang.tensor.storage.TensorEncoding
  * Keys are values: equal keys select the same kernel, and a key prints as something a log or an
  * `UnsupportedKernel` message can show: `matmul(F32/Dense(4B) contiguous × F32/Q4_K blocked) @host`.
  */
-@ExperimentalMemoryApi
 public data class KernelKey(
     val op: String,
     val operands: List<OperandKey>,
@@ -43,7 +41,6 @@ public data class KernelKey(
 }
 
 /** One operand of a [KernelKey]: its [Format] plus the layout class the kernel must cope with. */
-@ExperimentalMemoryApi
 public data class OperandKey(val format: Format, val layout: LayoutClass) {
     override fun toString(): String = "$format ${layout.name.lowercase()}"
 
@@ -77,7 +74,6 @@ public data class OperandKey(val format: Format, val layout: LayoutClass) {
  * orders, and #973 is what happens when that is left implicit. A kernel declares which one it
  * takes, and the dispatcher relayouts when the operand disagrees.
  */
-@ExperimentalMemoryApi
 public enum class LayoutClass {
     CONTIGUOUS,
     STRIDED,
@@ -94,15 +90,13 @@ public enum class LayoutClass {
 }
 
 /** Thrown when no registered kernel and no adapter chain can serve a key; lists what is registered. */
-@ExperimentalMemoryApi
 public class UnsupportedKernelException(
     public val key: KernelKey,
     public val candidates: List<String>,
     message: String = "No kernel for $key" + if (candidates.isEmpty()) "" else "; registered: ${candidates.joinToString(", ")}",
 ) : IllegalArgumentException(message)
 
 /** The encoding name a [KernelKey] uses for a format, matching `KernelProvider.supports`' dtype keys. */
-@ExperimentalMemoryApi
 public val Format.kernelEncodingName: String
     get() = when (val e = encoding) {
         is TensorEncoding.Dense -> dtype.name

skainet-backends/skainet-backend-api/src/commonMain/kotlin/sk/ainet/backend/api/kernel/KernelPacks.kt

@@ -1,6 +1,5 @@
 package sk.ainet.backend.api.kernel
 
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 import sk.ainet.lang.memory.Format
 import sk.ainet.lang.memory.Storage
 import sk.ainet.lang.memory.TensorView
@@ -21,7 +20,6 @@ import sk.ainet.lang.types.FP32
  * the dispatcher relayouts. That distinction is what #1029 was missing and what made mixing the two
  * a silent-wrong-numbers bug rather than a crash (#968, #971).
  */
-@ExperimentalMemoryApi
 public object KernelPacks {
 
     /** Capability marker for a provider that needs an explicit vector unit (Panama, NEON, …). */
@@ -81,7 +79,6 @@ public object KernelPacks {
  * output-major (`[n, k]`, the shape SKaiNET's dispatch normalises to). Unwraps each view once —
  * per the Phase-2 spike (#1016) — and calls the pack's strided GEMM.
  */
-@ExperimentalMemoryApi
 public class Fp32ViewMatmulKernel(
     providerName: String,
     private val kernel: Fp32MatmulKernel,

skainet-backends/skainet-backend-api/src/commonMain/kotlin/sk/ainet/backend/api/kernel/MappedCapableKernel.kt

@@ -1,6 +1,5 @@
 package sk.ainet.backend.api.kernel
 
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 
 /**
  * Marks a [ViewKernel] that serves its `BLOCKED_ROW_MAJOR` weight operand straight from off-heap
@@ -17,5 +16,4 @@ import sk.ainet.lang.memory.ExperimentalMemoryApi
  * so a kernel gaining or losing this marker without updating them fails CI instead of drifting
  * silently.
  */
-@ExperimentalMemoryApi
 public interface MappedCapableKernel

skainet-backends/skainet-backend-api/src/commonMain/kotlin/sk/ainet/backend/api/kernel/MatmulKernels.kt

@@ -1,6 +1,5 @@
 package sk.ainet.backend.api.kernel
 
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 import sk.ainet.lang.memory.Format
 import sk.ainet.lang.memory.TensorView
 
@@ -9,7 +8,6 @@ import sk.ainet.lang.memory.TensorView
  * *Kernel*, §5.2). Custom kernels are written against views and registered — an author never
  * touches a `TensorData` subclass.
  */
-@ExperimentalMemoryApi
 public interface ViewKernel {
     /** What this kernel serves. */
     public val key: KernelKey
@@ -39,7 +37,6 @@ public interface ViewKernel {
  * `out = a × bᵀ` in the shapes SKaiNET's dispatch normalises to: `a` is `[m, k]`, `b` is `[n, k]`
  * (a weight stored output-major, as GGUF does), `out` is `[m, n]`.
  */
-@ExperimentalMemoryApi
 public class ReferenceMatmulKernel(override val key: KernelKey) : ViewKernel {
     override val name: String get() = "reference"
 
@@ -61,7 +58,6 @@ public class ReferenceMatmulKernel(override val key: KernelKey) : ViewKernel {
 
     public companion object {
         /** The reference kernel for the formats of [a] and [b]. */
-        @ExperimentalMemoryApi
         public fun forOperands(a: Format, b: Format): ReferenceMatmulKernel =
             ReferenceMatmulKernel(KernelKey("matmul", listOf(OperandKey.contiguous(a), OperandKey.contiguous(b))))
     }

skainet-backends/skainet-backend-api/src/commonMain/kotlin/sk/ainet/backend/api/kernel/PackedLayoutFixtures.kt

@@ -1,6 +1,5 @@
 package sk.ainet.backend.api.kernel
 
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 import sk.ainet.lang.memory.blockSpec
 import sk.ainet.lang.tensor.storage.TensorEncoding
 
@@ -21,7 +20,6 @@ import sk.ainet.lang.tensor.storage.TensorEncoding
  * Every fixture is **three blocks wide** on purpose. At one block per row the two orders coincide,
  * which is exactly the shape that hid #968.
  */
-@ExperimentalMemoryApi
 public object PackedLayoutFixtures {
 
     /** The formats the contract covers — everything with a block geometry and a matmul kernel. */

skainet-backends/skainet-backend-api/src/commonMain/kotlin/sk/ainet/backend/api/kernel/PackedViewMatmulKernel.kt

@@ -1,7 +1,6 @@
 package sk.ainet.backend.api.kernel
 
 import sk.ainet.lang.memory.BlockOrder
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 import sk.ainet.lang.memory.Format
 import sk.ainet.lang.memory.Storage
 import sk.ainet.lang.memory.TensorView
@@ -20,7 +19,6 @@ import sk.ainet.lang.types.FP32
  * One row at a time: the SPI is a matrix-vector kernel, so an activation of `m` rows is `m` calls,
  * which is what the decode path does anyway (`m == 1`).
  */
-@ExperimentalMemoryApi
 public class PackedViewMatmulKernel(
     providerName: String,
     private val encodingName: String,

skainet-backends/skainet-backend-api/src/commonMain/kotlin/sk/ainet/backend/api/kernel/PackedWeights.kt

@@ -1,7 +1,6 @@
 package sk.ainet.backend.api.kernel
 
 import sk.ainet.lang.memory.BlockOrder
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 import sk.ainet.lang.memory.Scope
 import sk.ainet.lang.memory.blockSpec
 import sk.ainet.lang.memory.TensorView
@@ -23,7 +22,6 @@ import sk.ainet.lang.tensor.storage.TensorEncoding
  * The relayout is O(bytes). Call it **once, at load**; a weight prepacked at load hits the packed
  * kernel's key directly and the dispatcher copies nothing per call (#1095).
  */
-@ExperimentalMemoryApi
 public object PackedWeights {
 
     /**

skainet-backends/skainet-backend-api/src/commonMain/kotlin/sk/ainet/backend/api/kernel/RegistryKernelCapabilities.kt

@@ -1,6 +1,5 @@
 package sk.ainet.backend.api.kernel
 
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 import sk.ainet.lang.memory.plan.KernelCapabilities
 import sk.ainet.lang.tensor.storage.TensorEncoding
 
@@ -26,7 +25,6 @@ import sk.ainet.lang.tensor.storage.TensorEncoding
  * CPU cannot feed anything, whatever it declares. Dispatch kernels carry their requirements in
  * their key's capability set instead, and are filtered when they are selected.
  */
-@ExperimentalMemoryApi
 public object RegistryKernelCapabilities : KernelCapabilities {
 
     /** The activation dtype every packed matmul kernel in the tree takes. */

skainet-backends/skainet-backend-api/src/commonMain/kotlin/sk/ainet/backend/api/kernel/TernaryF32GemvKernel.kt

@@ -1,6 +1,5 @@
 package sk.ainet.backend.api.kernel
 
-import sk.ainet.lang.memory.ExperimentalMemoryApi
 import sk.ainet.lang.memory.Format
 import sk.ainet.lang.memory.Storage
 import sk.ainet.lang.memory.TensorView
@@ -30,7 +29,6 @@ import sk.ainet.lang.types.FP32
  * row-major order. Output `[rows, n]` FP32. The per-tensor scale is applied to the output —
  * native implementations ([TernaryF32GemvNative]) do not see it.
  */
-@ExperimentalMemoryApi
 public class TernaryF32GemvKernel(override val key: KernelKey) : ViewKernel {
 
     override val name: String get() = "ternary_f32_gemv/reference"