chore(kernelrace): SKaiNET 0.60.0 / transformers 0.61.0

KernelRace/gradle/libs.versions.toml

@@ -18,11 +18,10 @@ kotlinxIo = "0.9.1"
 # IosModelProvider: skainet-data-source is JVM-only (KtorRemoteDataSourceFetcher lives in
 # its jvmMain), so the iOS download path talks to Ktor's Darwin engine on its own.
 ktorClientCore = "3.5.2"
-# skainet-transformers has not released against the 0.40.x core line yet
-# (tops out at 0.39.1 on Maven Central), so this sample — like KllamaDemo —
-# stays one core version behind the transformers-free samples.
-skainet = "0.40.1"
-skainet-transformers = "0.40.2"
+# SKaiNET-transformers 0.61.0 is built against engine 0.60.0; keep the two pins
+# moving together (the transformers release notes name the engine version).
+skainet = "0.60.0"
+skainet-transformers = "0.61.0"
 
 [libraries]
 kotlin-test = { module = "org.jetbrains.kotlin:kotlin-test", version.ref = "kotlin" }

KernelRace/shared/build.gradle.kts

@@ -51,7 +51,7 @@ kotlin {
             api(libs.skainet.io.gguf)
 
             // SKaiNET-transformers: Llama inference + generation. agent and
-            // runtime-kllama both publish wasmJs targets as of 0.39.1, so the
+            // runtime-kllama both publish wasmJs targets (since 0.39.1), so the
             // whole engine (unlike KllamaDemo's 0.34.1-era jvmMain-only split)
             // lives in commonMain.
             api(project.dependencies.platform(libs.skainet.transformers.bom))

KernelRace/shared/src/androidMain/kotlin/sk/ainet/samples/kernelrace/engine/LlamaRuntimeBuilder.android.kt

@@ -5,23 +5,21 @@ import sk.ainet.apps.llm.OptimizedLLMRuntime
 import sk.ainet.apps.llm.tokenizer.TokenizerFactory
 import sk.ainet.context.ExecutionContext
 import sk.ainet.io.AndroidRandomAccessSource
-import sk.ainet.io.model.QuantPolicy
 import sk.ainet.lang.types.FP32
-import sk.ainet.models.llama.DecoderGgufWeightLoader
+import sk.ainet.lang.nn.dsl.decoder.DecoderGgufWeightLoader
 import sk.ainet.models.llama.LlamaNetworkLoader
 import sk.ainet.samples.kernelrace.model.ModelData
 
 /**
  * The "five lines" fast path: random-access reads straight off the file,
- * `QuantPolicy.NATIVE_OPTIMIZED` keeps Q8_0 packed for the hand-written ARM NEON kernels
+ * the loader's default `WeightForm` (packed, mapped) keeps Q8_0 packed for the hand-written ARM NEON kernels
  * (skainet-backend-jni-cpu), which register themselves through the kernel SPI just by being
  * on the classpath.
  */
 actual suspend fun buildLlamaComponents(ctx: ExecutionContext, model: ModelData): LlamaComponents {
     val path = (model as ModelData.FilePath).path
     val weights = DecoderGgufWeightLoader(
         randomAccessProvider = { AndroidRandomAccessSource.open(path) },
-        quantPolicy = QuantPolicy.NATIVE_OPTIMIZED,
         acceptedArchitectures = setOf("llama", "mistral"), // SmolLM2 is llama-family
     ).loadToMapStreaming<FP32, Float>(ctx)
     val runtime = OptimizedLLMRuntime(

KernelRace/shared/src/commonMain/kotlin/sk/ainet/samples/kernelrace/engine/LlamaRuntimeBuilder.kt

@@ -6,7 +6,6 @@ import sk.ainet.apps.llm.Tokenizer
 import sk.ainet.apps.llm.tokenizer.TokenizerFactory
 import sk.ainet.context.ExecutionContext
 import sk.ainet.io.RandomAccessSource
-import sk.ainet.io.model.QuantPolicy
 import sk.ainet.lang.types.FP32
 import sk.ainet.models.llama.LlamaNetworkLoader
 import sk.ainet.samples.kernelrace.model.ModelData
@@ -47,7 +46,7 @@ private class InMemoryRandomAccessSource(private val bytes: ByteArray) : RandomA
 /**
  * Platform-split loader:
  * - Android takes the "five lines" fast path: random-access reads straight off the file,
- *   `QuantPolicy.NATIVE_OPTIMIZED` keeps Q8_0 packed for the hand-written ARM NEON kernels.
+ *   the loader's default `WeightForm` (packed, mapped) keeps Q8_0 packed for the hand-written ARM NEON kernels.
  * - Desktop JVM takes the same file-based path (no NEON kernels to dispatch to on non-Android
  *   hardware, but still avoids materializing the whole file on-heap).
  * - wasmJs has no filesystem, so it falls back to [buildLlamaComponentsFallback]: the whole GGUF
@@ -74,7 +73,7 @@ suspend fun buildLlamaComponentsFallback(ctx: ExecutionContext, bytes: ByteArray
     val tokenizer = TokenizerFactory.fromGgufSource(InMemoryRandomAccessSource(bytes))
     val randomAccessProvider = { InMemoryRandomAccessSource(bytes) }
     val model = LlamaNetworkLoader
-        .fromGguf(randomAccessProvider, QuantPolicy.NATIVE_OPTIMIZED, false)
+        .fromGguf(randomAccessProvider)
         .load<FP32, Float>(ctx)
     val runtime = OptimizedLLMRuntime(
         model = model,

KernelRace/shared/src/commonMain/kotlin/sk/ainet/samples/kernelrace/engine/LlmEngine.kt

@@ -47,9 +47,9 @@ class LlmEngine private constructor(private val components: LlamaComponents) : G
             onToken = { tokenId ->
                 if (tokenCount == 0) firstTokenAt = TimeSource.Monotonic.markNow()
                 tokenCount++
-                onToken(tokenizer.decode(tokenId))
+                onToken(tokenizer.decodeToken(tokenId))
             },
-            decode = { tokenizer.decode(it) },
+            decode = { tokenizer.decodeToken(it) },
         )
         val totalMs = genStart.elapsedNow().inWholeMilliseconds
         val decodeS = firstTokenAt?.elapsedNow()?.inWholeMilliseconds?.let { it / 1000.0 } ?: 0.0

KernelRace/shared/src/iosMain/kotlin/sk/ainet/samples/kernelrace/engine/LlamaRuntimeBuilder.ios.kt

@@ -5,23 +5,21 @@ import sk.ainet.apps.llm.OptimizedLLMRuntime
 import sk.ainet.apps.llm.tokenizer.TokenizerFactory
 import sk.ainet.context.ExecutionContext
 import sk.ainet.io.PosixPreadRandomAccessSource
-import sk.ainet.io.model.QuantPolicy
 import sk.ainet.lang.types.FP32
-import sk.ainet.models.llama.DecoderGgufWeightLoader
+import sk.ainet.lang.nn.dsl.decoder.DecoderGgufWeightLoader
 import sk.ainet.models.llama.LlamaNetworkLoader
 import sk.ainet.samples.kernelrace.model.ModelData
 
 /** Same file-based random-access shape as the JVM/Android actuals, backed by POSIX `pread(2)`
  *  (shared by macOS/iOS/Linux native in skainet-io-core's native64Main) instead of a JDK/ART
- *  file API. `QuantPolicy.NATIVE_OPTIMIZED` keeps Q8_0 packed for the Apple native-cinterop
+ *  file API. the loader's default `WeightForm` (packed, mapped) keeps Q8_0 packed for the Apple native-cinterop
  *  kernels — see kllama's registerPlatformBackends()/installNativeKernels() wiring, which
  *  installs the same provider this app relies on. */
 actual suspend fun buildLlamaComponents(ctx: ExecutionContext, model: ModelData): LlamaComponents {
     val path = (model as ModelData.FilePath).path
     fun openSource() = checkNotNull(PosixPreadRandomAccessSource.open(path)) { "Cannot open GGUF at $path" }
     val weights = DecoderGgufWeightLoader(
         randomAccessProvider = { openSource() },
-        quantPolicy = QuantPolicy.NATIVE_OPTIMIZED,
         acceptedArchitectures = setOf("llama", "mistral"), // SmolLM2 is llama-family
     ).loadToMapStreaming<FP32, Float>(ctx)
     val runtime = OptimizedLLMRuntime(

KernelRace/shared/src/jvmMain/kotlin/sk/ainet/samples/kernelrace/engine/LlamaRuntimeBuilder.jvm.kt

@@ -5,9 +5,8 @@ import sk.ainet.apps.llm.OptimizedLLMRuntime
 import sk.ainet.apps.llm.tokenizer.TokenizerFactory
 import sk.ainet.context.ExecutionContext
 import sk.ainet.io.JvmRandomAccessSource
-import sk.ainet.io.model.QuantPolicy
 import sk.ainet.lang.types.FP32
-import sk.ainet.models.llama.DecoderGgufWeightLoader
+import sk.ainet.lang.nn.dsl.decoder.DecoderGgufWeightLoader
 import sk.ainet.models.llama.LlamaNetworkLoader
 import sk.ainet.samples.kernelrace.model.ModelData
 
@@ -17,7 +16,6 @@ actual suspend fun buildLlamaComponents(ctx: ExecutionContext, model: ModelData)
     val path = (model as ModelData.FilePath).path
     val weights = DecoderGgufWeightLoader(
         randomAccessProvider = { JvmRandomAccessSource.open(path) },
-        quantPolicy = QuantPolicy.NATIVE_OPTIMIZED,
         acceptedArchitectures = setOf("llama", "mistral"),
     ).loadToMapStreaming<FP32, Float>(ctx)
     val runtime = OptimizedLLMRuntime(