coreml : fix ANE optimized encoder (#1716)

coreml/whisper-encoder.mm

@@ -24,9 +24,9 @@ struct whisper_coreml_context * whisper_coreml_init(const char * path_model) {
 
     // select which device to run the Core ML model on
     MLModelConfiguration *config = [[MLModelConfiguration alloc] init];
-    config.computeUnits = MLComputeUnitsCPUAndGPU;
+    // config.computeUnits = MLComputeUnitsCPUAndGPU;
     //config.computeUnits = MLComputeUnitsCPUAndNeuralEngine;
-    //config.computeUnits = MLComputeUnitsAll;
+    config.computeUnits = MLComputeUnitsAll;
 
     const void * data = CFBridgingRetain([[whisper_encoder_impl alloc] initWithContentsOfURL:url_model configuration:config error:nil]);
 

models/convert-whisper-to-coreml.py

@@ -143,20 +143,7 @@ class AudioEncoderANE(AudioEncoder):
             x = block(x)
 
         x = self.ln_post(x)
-
-        # """
-        # TODO:
-        # I think we need to transpose the result here to make it fit whisper.cpp memory order.
-        # However, even doing this, the results are still wrong. Kind of less wrong compared to
-        # not transposing, but still wrong.
-
-        # Also, I don't know why the original OpenAI implementation does not need to transpose
-
-        # transpose to (batch_size, n_ctx, n_state)
-        # x : torch.Tensor, shape = (batch_size, n_state, 1, n_ctx)
-
-        # """
-        # x = x.transpose(1,3)
+        x = x.squeeze(2).transpose(1, 2)
 
         return x
 

models/generate-coreml-model.sh

@@ -23,7 +23,7 @@ if [[ $mname == "-h5" ]]; then
   echo $mpath
   python3 models/convert-h5-to-coreml.py --model-name $mname --model-path $mpath --encoder-only True
 else
-  python3 models/convert-whisper-to-coreml.py --model $mname --encoder-only True
+  python3 models/convert-whisper-to-coreml.py --model $mname --encoder-only True  --optimize-ane True
 fi
 
 xcrun coremlc compile models/coreml-encoder-${mname}.mlpackage models/