Await tensor.getData() for WebGPU outputs (audio_embeds + logits) so data is actually copied back to CPU

mega-asr.js

@@ -322,7 +322,8 @@ async function transcribe({ mel, dims, T_mel }) {
   setStatus("audio encoder ...");
   const melTensor = new ort.Tensor("float32", mel, dims);
   const encOut = await state.encoder.run({ mel: melTensor });
-  const audioEmbedsAll = encOut.audio_embeds.data; // Float32Array (1*390*2048,)
+  // For WebGPU outputs we must await getData() to bring values back to CPU.
+  const audioEmbedsAll = await encOut.audio_embeds.getData(true); // Float32Array (1*390*2048,)
   const audioEmbedsDims = encOut.audio_embeds.dims; // [1, 390, 2048]
   const realChunks = Math.floor((T_mel + 99) / 100);
   const lastChunkMel = T_mel - (realChunks - 1) * 100;
@@ -375,7 +376,9 @@ async function transcribe({ mel, dims, T_mel }) {
 
   // 5. greedy decode
   setStatus("decoding ...");
-  let logits = prefillOut.logits.data; // (1, L, VOCAB)
+  // WebGPU outputs live in GPU memory — must call getData() (async) to bring
+  // them back to CPU. CPU/WASM tensors return their data array synchronously.
+  let logits = await prefillOut.logits.getData(true); // (1, L, VOCAB)
   const logitsDims = prefillOut.logits.dims;
   // Diagnostic: dump top-5 of last logit so we can see what the decoder predicted
   {
@@ -419,7 +422,7 @@ async function transcribe({ mel, dims, T_mel }) {
       feeds[`past.${i}.value`] = kvs[2 * i + 1];
     }
     const out = await state.step.run(feeds);
-    logits = out.logits.data;
+    logits = await out.logits.getData(true);
     nid = argmax(logits, 0, VOCAB);
     gen.push(nid);
     curLen += 1;