@omote/core 0.6.4 → 0.6.6

package/dist/index.mjs

@@ -2194,7 +2194,7 @@ async function getOnnxRuntimeForPreference(preference = "auto") {
   const ort = await getOnnxRuntime(backend);
   return { ort, backend };
 }
-function getSessionOptions(backend) {
+function getSessionOptions(backend, config) {
   if (backend === "webgpu") {
     return {
       executionProviders: [
@@ -2210,7 +2210,7 @@ function getSessionOptions(backend) {
   if (isIOS()) {
     return {
       executionProviders: ["wasm"],
-      graphOptimizationLevel: "basic",
+      graphOptimizationLevel: config?.iosDisableOptimization ? "disabled" : "basic",
       enableCpuMemArena: false,
       enableMemPattern: false
     };
@@ -2477,7 +2477,7 @@ var _Wav2Vec2Inference = class _Wav2Vec2Inference {
       logger3.info("ONNX Runtime loaded", { backend: this._backend });
       const modelUrl = this.config.modelUrl;
       const dataUrl = this.config.externalDataUrl !== false ? typeof this.config.externalDataUrl === "string" ? this.config.externalDataUrl : `${modelUrl}.data` : null;
-      const sessionOptions = getSessionOptions(this._backend);
+      const sessionOptions = getSessionOptions(this._backend, { iosDisableOptimization: true });
       let isCached = false;
       if (isIOS()) {
         logger3.info("iOS: passing model URLs directly to ORT (low-memory path)", {
@@ -7320,10 +7320,11 @@ function createA2E(config = {}) {
     useCpu = false;
     logger12.info("Forcing GPU A2E model (Wav2Vec2)");
   } else {
-    useCpu = shouldUseCpuA2E();
-    logger12.info("Auto-detected A2E model", {
-      useCpu,
-      isSafari: isSafari()
+    useCpu = false;
+    logger12.info("Auto-detected A2E model: trying GPU first (fp16 external data)", {
+      isSafari: isSafari(),
+      isIOS: isIOS(),
+      fallbackOnError
     });
   }
   if (useCpu) {
@@ -7385,6 +7386,7 @@ var A2EWithFallback = class {
     }
   }
   async fallbackToCpu(reason) {
+    console.error("[A2EWithFallback] GPU\u2192CPU FALLBACK TRIGGERED. Reason:", reason);
     logger12.warn("GPU model load failed, falling back to CPU model", { reason });
     try {
       await this.implementation.dispose();
@@ -10253,10 +10255,7 @@ var EmotionResolver = class {
       if (upper[i] > 1) upper[i] = 1;
       if (lower[i] > 1) lower[i] = 1;
     }
-    return {
-      upper: new Float32Array(upper),
-      lower: new Float32Array(lower)
-    };
+    return { upper, lower };
   }
 };
 
@@ -10279,6 +10278,7 @@ var FaceCompositor = class {
   constructor(config) {
     this.emotionResolver = new EmotionResolver();
     // Pre-allocated buffers
+    this.outputBuffer = new Float32Array(52);
     this.smoothedUpper = new Float32Array(52);
     this.smoothedLower = new Float32Array(52);
     this.lifeBuffer = new Float32Array(52);
@@ -10296,10 +10296,12 @@ var FaceCompositor = class {
    *
    * @param base - A2E raw output (Float32Array[52], LAM_BLENDSHAPES order)
    * @param input - Per-frame input (deltaTime, emotion, life layer params)
-   * @returns Float32Array[52] with all values clamped to [0, 1]
+   * @param target - Optional pre-allocated output buffer (avoids per-frame allocation).
+   *   When omitted, an internal buffer is used (valid until next compose() call).
+   * @returns Blendshapes (Float32Array[52] clamped [0,1]) and head rotation deltas
    */
-  compose(base, input) {
-    const out = new Float32Array(52);
+  compose(base, input, target) {
+    const out = target ?? this.outputBuffer;
     out.set(base);
     const emotion = input.emotion ?? this.stickyEmotion;
     if (emotion) {
@@ -10321,7 +10323,14 @@ var FaceCompositor = class {
         out[i] *= 1 + this.smoothedLower[i] * bilabialSuppress;
       }
     }
-    this.lifeLayer.updateToArray(input.deltaTime, input, this.lifeBuffer);
+    const lifeResult = this.lifeLayer.update(input.deltaTime, input);
+    this.lifeBuffer.fill(0);
+    for (const [name, value] of Object.entries(lifeResult.blendshapes)) {
+      const idx = BS_INDEX2.get(name);
+      if (idx !== void 0) {
+        this.lifeBuffer[idx] = value;
+      }
+    }
     for (let i = 0; i < 52; i++) {
       if (IS_EYE_CHANNEL[i]) {
         out[i] = this.lifeBuffer[i];
@@ -10336,7 +10345,7 @@ var FaceCompositor = class {
       if (out[i] < 0) out[i] = 0;
       else if (out[i] > 1) out[i] = 1;
     }
-    return out;
+    return { blendshapes: out, headDelta: lifeResult.headDelta };
   }
   /**
    * Set sticky emotion (used when input.emotion is not provided).
@@ -10623,6 +10632,7 @@ var VoicePipeline = class extends EventEmitter {
           new Promise((r) => setTimeout(() => r("timeout"), timeoutMs))
         ]);
         if (lamLoadResult === "timeout") {
+          console.error(`[VoicePipeline] LAM TIMEOUT after ${timeoutMs}ms \u2014 forcing CPU fallback`);
           logger19.warn(`LAM GPU load timed out after ${timeoutMs}ms, falling back to CPU`);
           await lam.dispose();
           lam = createA2E({