Use INT4 encoder (MatMulNBits — WebGPU-supported); ORT 1.23 keeps webgpu fp16 fix for decoders

mega-asr.js

@@ -232,26 +232,15 @@ async function loadAll() {
   setProgress(30);
 
   // 3. ONNX sessions
-  // Audio encoder: try static INT8 (QLinearConv/QLinearMatMul, browser-OK),
-  // then fall back to INT4 (MatMulNBits) if static INT8 also has issues.
-  // The OLD audio_encoder_int8.onnx uses ConvInteger (dynamic quant) which
-  // onnxruntime-web does NOT support — skipping that.
-  setLoaderStatus("audio encoder (INT8 static → INT4 fallback) ...");
-  try {
-    state.encoder = await createSessionSimple(
-      `${HF_ROOT}/onnx/audio_encoder_int8_static.onnx`,
-      "audio_encoder INT8 (static)",
-      p => setProgress(30 + p * 10),
-    );
-  } catch (e) {
-    log(`static INT8 encoder unsupported (${e.message}); falling back to INT4`);
-    state.encoder = await createSession(
-      `${HF_ROOT}/onnx/audio_encoder_int4.onnx`,
-      `${HF_ROOT}/onnx/audio_encoder_int4.onnx.data`,
-      "audio_encoder INT4",
-      p => setProgress(30 + p * 10),
-    );
-  }
+  // Audio encoder: INT4 (MatMulNBits) — well-supported on WebGPU and WASM.
+  // Static INT8 (QLinearConv/QLinearMatMul) crashes onnxruntime-web on WebGPU.
+  setLoaderStatus("audio encoder INT4 ...");
+  state.encoder = await createSession(
+    `${HF_ROOT}/onnx/audio_encoder_int4.onnx`,
+    `${HF_ROOT}/onnx/audio_encoder_int4.onnx.data`,
+    "audio_encoder INT4",
+    p => setProgress(30 + p * 10),
+  );
   setProgress(40);
 
   setLoaderStatus("decoder prefill (~970 MB)...");