backend: support Metal

README.md

@@ -61,6 +61,8 @@ cmake -DGGML_METAL=ON -DBUILD_SHARED_LIBS=Off ..
 cmake --build . --config Release
 ```
 
+Once built, pass the `--n-gpu-layers` flag with a value greather than 0 to the executable.
+
 ### Using cuBLAS
 
 The inference can be offloaded on a CUDA backend with cuBLAS.

decoder.h

@@ -12,61 +12,65 @@
 
 struct encodec_decoder_block {
     // upsampling layers
-    struct ggml_tensor *us_conv_w;
-    struct ggml_tensor *us_conv_b;
+    struct ggml_tensor * us_conv_w;
+    struct ggml_tensor * us_conv_b;
 
     // conv1
-    struct ggml_tensor *conv_1_w;
-    struct ggml_tensor *conv_1_b;
+    struct ggml_tensor * conv_1_w;
+    struct ggml_tensor * conv_1_b;
 
     // conv2
-    struct ggml_tensor *conv_2_w;
-    struct ggml_tensor *conv_2_b;
+    struct ggml_tensor * conv_2_w;
+    struct ggml_tensor * conv_2_b;
 
     // shortcut
-    struct ggml_tensor *conv_sc_w;
-    struct ggml_tensor *conv_sc_b;
+    struct ggml_tensor * conv_sc_w;
+    struct ggml_tensor * conv_sc_b;
 };
 
 struct encodec_decoder {
-    struct ggml_tensor *init_conv_w;
-    struct ggml_tensor *init_conv_b;
+    struct ggml_tensor * init_conv_w;
+    struct ggml_tensor * init_conv_b;
 
     encodec_lstm lstm;
 
-    struct ggml_tensor *final_conv_w;
-    struct ggml_tensor *final_conv_b;
+    struct ggml_tensor * final_conv_w;
+    struct ggml_tensor * final_conv_b;
 
     std::vector<encodec_decoder_block> blocks;
 };
 
-struct ggml_tensor *encodec_forward_decoder(
-    const struct encodec_decoder *decoder, struct ggml_context *ctx0,
-    struct ggml_tensor *quantized_out, const int *ratios, const int kernel_size, const int res_kernel_size,
-    const int stride) {
+struct ggml_tensor * encodec_forward_decoder(
+    const struct encodec_decoder * decoder,
+             struct ggml_context * ctx0,
+              struct ggml_tensor * quantized_out,
+                       const int * ratios,
+                       const int   kernel_size,
+                       const int   res_kernel_size,
+                       const int   stride) {
 
     if (!quantized_out) {
         fprintf(stderr, "%s: null input tensor\n", __func__);
         return NULL;
     }
 
-    struct ggml_tensor *inpL = strided_conv_1d(
+    struct ggml_tensor * inpL = strided_conv_1d(
         ctx0, quantized_out, decoder->init_conv_w, decoder->init_conv_b, stride);
 
     // lstm
     {
-        struct ggml_tensor *cur = inpL;
+        struct ggml_tensor * cur = inpL;
 
         const encodec_lstm lstm = decoder->lstm;
 
         // first lstm layer
         char l0_prefix[7] = "dec_l0";
-        struct ggml_tensor *hs1 = forward_pass_lstm_unilayer(
+        struct ggml_tensor * hs1 = forward_pass_lstm_unilayer(
             ctx0, cur, lstm.l0_ih_w, lstm.l0_hh_w, lstm.l0_ih_b, lstm.l0_hh_b, l0_prefix);
 
         // second lstm layer
         char l1_prefix[7] = "dec_l1";
-        struct ggml_tensor *out = forward_pass_lstm_unilayer(
+        struct ggml_tensor * out = forward_pass_lstm_unilayer(
             ctx0, hs1, lstm.l1_ih_w, lstm.l1_hh_w, lstm.l1_ih_b, lstm.l1_hh_b, l1_prefix);
 
         inpL = ggml_add(ctx0, inpL, out);
@@ -81,10 +85,10 @@ struct ggml_tensor *encodec_forward_decoder(
         inpL = strided_conv_transpose_1d(
             ctx0, inpL, block.us_conv_w, block.us_conv_b, ratios[layer_ix]);
 
-        struct ggml_tensor *current = inpL;
+        struct ggml_tensor * current = inpL;
 
         // shortcut
-        struct ggml_tensor *shortcut = strided_conv_1d(
+        struct ggml_tensor * shortcut = strided_conv_1d(
             ctx0, inpL, block.conv_sc_w, block.conv_sc_b, stride);
 
         // conv1
@@ -106,7 +110,7 @@ struct ggml_tensor *encodec_forward_decoder(
     // final conv
     inpL = ggml_elu(ctx0, inpL);
 
-    struct ggml_tensor *decoded_inp = strided_conv_1d(
+    struct ggml_tensor * decoded_inp = strided_conv_1d(
         ctx0, inpL, decoder->final_conv_w, decoder->final_conv_b, stride);
 
     return decoded_inp;