Refactor: Pure MLX implementation and repository restructuring

Author: Acelogic

.gitignore

@@ -26,7 +26,7 @@ Intermediates.noindex/
 *.xcresult/
 *.xccheckout
 *.xcscmblueprint
-xcuserdata/
+/xcuserdata/
 *.orig
 
 # Logs
@@ -38,10 +38,14 @@ xcuserdata/
 *.onnx
 *.index
 *.safetensors
+*.npz
 assets/audios/
 assets/datasets/
 logs/
 rvc/models/
+rvc_mlx/models/
+weights/
+test-audio/
 
 # Media
 *.mp3
@@ -51,5 +55,6 @@ rvc/models/
 *.wav
 *.png
 
+
 # Misc
 *.txt

Demos/iOS/RVCNative/RVCNativePackage/Sources/RVCNativeFeature/RVC/HubertModel.swift

@@ -310,6 +310,7 @@ public class HubertModel: Module {
         x = feature_projection(x)
         x = encoder(x)
         // x = final_proj(x) // Skip projection for RVC v2 (model expects 768-dim)
+        print("DEBUG: HubertModel output shape: \(x.shape)")
         return x
     }
 }

Demos/iOS/RVCNative/RVCNativePackage/Sources/RVCNativeFeature/RVC/RMVPE.swift

@@ -388,6 +388,7 @@ class RMVPE: Module {
         // audio: [T]
         let melProcessor = MelSpectrogram()
         let mel = melProcessor(audio) // [n_mels, T_frames] (Log Mel) (128, T)
+        print("DEBUG: Mel Spectrogram Stats: min \(mel.min().item(Float.self)), max \(mel.max().item(Float.self)), shape \(mel.shape)")
 
         // Model expects [N, T, n_mels] (transposed)
         // mel.T -> [T_frames, n_mels]. Add dim [1, T, n_mels]
@@ -398,6 +399,13 @@ class RMVPE: Module {
 
         // Decode
         let f0 = self.decode(hidden, thred: thred) // [1, T, 1]
+        
+        // DEBUG Stats
+        let f0_min = f0.min().item(Float.self)
+        let f0_max = f0.max().item(Float.self)
+        let f0_mean = f0.mean().item(Float.self)
+        print("DEBUG: RMVPE F0 Stats: min \(f0_min), max \(f0_max), mean \(f0_mean)")
+        
         return f0
     }
 }

Demos/iOS/RVCNative/RVCNativePackage/Sources/RVCNativeFeature/RVC/RVCInference.swift

@@ -298,6 +298,6 @@ import MLXNN
                 sid: sid
             )
 
-            return audioOut // [1, T_out, 1]
+        return audioOut // [1, T_out, 1]
         }
     }

Demos/iOS/RVCNative/RVCNativePackage/Sources/RVCNativeFeature/RVC/RVCModel.swift

@@ -310,17 +310,20 @@ class Generator: Module {
 
     func callAsFunction(_ x: MLXArray, f0: MLXArray, g: MLXArray? = nil) -> MLXArray {
         var out = conv_pre(x)
+        print("DEBUG: Generator.conv_pre out: [\(out.min().item(Float.self))...\(out.max().item(Float.self))]")
 
         // Add speaker conditioning if available
         if let g = g, let condLayer = cond {
             out = out + condLayer(g)
+            print("DEBUG: Generator.cond added: [\(out.min().item(Float.self))...\(out.max().item(Float.self))]")
         }
 
 
         // NSF Source Signal: [B, L*U, 1] (High Resolution)
         // upp = prod(upsampleRates) = 400
         let upp = 400
         let har_source = m_source(f0, upsampling_factor: upp)
+        print("DEBUG: Generator.har_source: [\(har_source.min().item(Float.self))...\(har_source.max().item(Float.self))]")
         // har_source is now [B, AudioLen, 1]
 
         var resIdx = 0
@@ -332,16 +335,12 @@ class Generator: Module {
             // Add NSF Noise
             // noise_conv reduces har_source resolution to match current 'out'
             let noise_conv = noise_convs[i]
-            // We need to transpose har_source to [B, 1, T] ?? No. MLX is [N, L, C].
-            // har_source is [N, T, 1]. noise_conv(1->C) -> [N, T_down, C].
-            // Should match 'out' shape.
             let n = noise_conv(har_source)
 
-            // Crop if necessary (Python does: if x.shape[1] != n.shape[1])
+            // Crop if necessary
             if out.shape[1] != n.shape[1] {
                 let minLen = min(out.shape[1], n.shape[1])
                 out = out[0..., 0..<minLen, 0...]
-                // n = n[0..., 0..<minLen, 0...] // n is let constant, but we add it.
                 out = out + n[0..., 0..<minLen, 0...]
             } else {
                 out = out + n
@@ -358,11 +357,15 @@ class Generator: Module {
             }
             // Average
             out = xs! / 3.0
+            print("DEBUG: Generator.ups[\(i)] (after resblocks) out: [\(out.min().item(Float.self))...\(out.max().item(Float.self))]")
         }
 
         out = leakyRelu(out)
         out = conv_post(out)
+        print("DEBUG: Generator.conv_post out: [\(out.min().item(Float.self))...\(out.max().item(Float.self))]")
+        
         out = tanh(out)
+        print("DEBUG: Generator.final (tanh): [\(out.min().item(Float.self))...\(out.max().item(Float.self))]")
         return out
     }
 }

Demos/iOS/RVCNative/RVCNativePackage/Sources/RVCNativeFeature/RVC/Synthesizer.swift

@@ -430,17 +430,20 @@ public class Synthesizer: Module {
 
         // Encode features
         let (m_p, logs_p, xMask) = enc_p(phone, pitch: pitch, lengths: phoneLengths)
-        print("DEBUG: TextEncoder output - m_p: \(m_p.shape), logs_p: \(logs_p.shape)")
+        print("DEBUG: TextEncoder output - m_p: \(m_p.shape) [\(m_p.min().item(Float.self))...\(m_p.max().item(Float.self))], logs_p: \(logs_p.shape) [\(logs_p.min().item(Float.self))...\(logs_p.max().item(Float.self))]")
 
         // Sample from encoded distribution
         let xMaskExpanded = xMask.expandedDimensions(axis: -1)
         let z_p = (m_p + exp(logs_p) * MLXRandom.normal(m_p.shape).asType(m_p.dtype) * 0.66666) * xMaskExpanded
+        print("DEBUG: z_p stats: min \(z_p.min().item(Float.self)), max \(z_p.max().item(Float.self))")
 
         // Flow reverse pass
         let z = flow(z_p, xMask: xMaskExpanded, g: g, reverse: true)
+        print("DEBUG: z (after flow) stats: min \(z.min().item(Float.self)), max \(z.max().item(Float.self))")
 
         // Decode to audio
         let output = dec(z * xMaskExpanded, f0: nsff0 ?? MLX.zeros([phone.shape[0], phone.shape[1], 1]), g: g)
+        print("DEBUG: Generator output stats: min \(output.min().item(Float.self)), max \(output.max().item(Float.self))")
 
         return output
     }

Demos/iOS/RVCNative/Vendor/mlx-swift

@@ -0,0 +1 @@
+Subproject commit 15b6ade64bd6aefdb35d732a90678c3b264ecfac