fix(python tcp server): use CPU on macOS as a tmp fix against the memory leak for TTS

tcp_server/src/lib/tts/api.py

@@ -11,6 +11,7 @@ import os
 from . import utils
 from .models import SynthesizerTrn
 from .split_utils import split_sentence
+from ..utils import is_macos
 
 # torch.backends.cudnn.enabled = False
 
@@ -40,6 +41,13 @@ class TTS(nn.Module):
         if 'mps' in device:
             assert torch.backends.mps.is_available()
 
+        if is_macos():
+            """
+            Temporary fix.
+            Force CPU device for macOS because of the memory leak where cache does not want to clear up on MPS
+            """
+            device = 'cpu'
+
         self.log(f'Device: {device}')
 
         hps = utils.get_hparams_from_file(config_path)
@@ -172,11 +180,6 @@ class TTS(nn.Module):
             else:
                 audio_list.append(audio)
 
-        if self.device == 'cuda':
-            torch.cuda.empty_cache()
-        if self.device == 'mps':
-            torch.mps.empty_cache()
-
         toc = time.perf_counter()
         self.log(f"Time taken to generate audio: {toc - tic:0.4f} seconds")
 
@@ -184,6 +187,12 @@ class TTS(nn.Module):
         if not stream:
             audio = self.audio_numpy_concat(audio_list, sr=self.hps.data.sampling_rate, speed=speed)
 
+            del audio_list
+            if self.device == 'cuda':
+                torch.cuda.empty_cache()
+            if self.device == 'mps':
+                torch.mps.empty_cache()
+
             if output_path is None:
                 return audio
             else: