update

app.py

@@ -25,6 +25,7 @@ import gradio as gr
 import librosa
 import torch
 import torchaudio
+import devicetorch
 
 torchaudio.set_audio_backend("soundfile")
 
@@ -190,6 +191,9 @@ def inference(
     if torch.cuda.is_available():
         torch.cuda.empty_cache()
         gc.collect()
+    elif torch.backends.mps.is_available():
+        torch.mps.empty_cache()
+        gc.collect()
 
 
 def inference_with_auto_rerank(
@@ -341,6 +345,9 @@ def change_if_load_asr_model(if_load):
         if torch.cuda.is_available():
             torch.cuda.empty_cache()
             gc.collect()
+        elif torch.backends.mps.is_available():
+            torch.mps.empty_cache()
+            gc.collect()
         return gr.Checkbox(label="Load faster whisper model", value=if_load)
 
 
@@ -602,7 +609,8 @@ def parse_args():
         default="checkpoints/fish-speech-1.4/firefly-gan-vq-fsq-8x1024-21hz-generator.pth",
     )
     parser.add_argument("--decoder-config-name", type=str, default="firefly_gan_vq")
-    parser.add_argument("--device", type=str, default="cuda")
+    #parser.add_argument("--device", type=str, default="cuda")
+    parser.add_argument("--device", type=str, default=devicetorch.get(torch))
     parser.add_argument("--half", action="store_true")
     parser.add_argument("--compile", action="store_true",default=True)
     parser.add_argument("--max-gradio-length", type=int, default=0)

requirements.txt

@@ -1,5 +1,5 @@
-torch==2.3.0
-torchaudio
+#torch==2.3.0
+#torchaudio
 transformers>=4.35.2
 datasets>=2.14.5
 lightning>=2.1.0
@@ -22,9 +22,11 @@ vector_quantize_pytorch>=1.14.7
 samplerate>=0.2.1
 resampy>=0.4.3
 spaces>=0.26.1
-einx[torch]==0.2.0
+#einx[torch]==0.2.0
+einx[torch]
 opencc
 faster-whisper
 ormsgpack
 ffmpeg
-soundfile
+soundfile
+devicetorch