update

app.py

@@ -51,7 +51,7 @@ def infer_music(lrc, ref_audio_path, steps, file_type, max_frames=2048):
                                start_time=start_time,
                                file_type=file_type
                                )
-    torch.cuda.empty_cache()
+    devicetorch.empty_cache(torch)
     gc.collect()
     print(">4")
 
@@ -207,7 +207,7 @@ with gr.Blocks(css=css) as demo:
                                     interactive=True,
                                     elem_id="step_slider"
                                 )
-                        file_type = gr.Dropdown(["wav", "mp3", "ogg"], label="Output Format", value="wav")
+                        file_type = gr.Dropdown(["wav", "mp3", "ogg"], label="Output Format", value="mp3")
                     
 
 

diffrhythm/infer/infer.py

@@ -134,7 +134,11 @@ if __name__ == "__main__":
     parser.add_argument('--output-dir', type=str, default="example/output")
     args = parser.parse_args()
     
-    device = 'cuda'
+    device = "cpu"
+    if torch.cuda.is_available():
+        device = "cuda"
+    elif torch.mps.is_available():
+        device = "mps"
     
     audio_length = args.audio_length
     if audio_length == 95:

diffrhythm/infer/infer_utils.py

@@ -169,8 +169,7 @@ def get_lrc_token(text, tokenizer, device):
     return lrc_emb, normalized_start_time
 
 def load_checkpoint(model, ckpt_path, device, use_ema=True):
-    if device == "cuda":
-        model = model.half()
+    model = model.half()
 
     ckpt_type = ckpt_path.split(".")[-1]
     if ckpt_type == "safetensors":

requirements.txt

@@ -12,7 +12,8 @@ pandas==2.2.3
 pylance==0.23.2
 ema-pytorch==0.7.7
 prefigure==0.0.10
-bitsandbytes==0.45.3
+#bitsandbytes==0.45.3
+bitsandbytes
 muq==0.1.0
 mutagen==1.47.0
 pyopenjtalk==0.4.0
@@ -21,7 +22,7 @@ jieba==0.42.1
 cn2an==0.5.23
 pypinyin==0.53.0
 #onnxruntime==1.20.1
-onnxruntime-gpu
+#onnxruntime-gpu
 Unidecode==1.3.8
 phonemizer==3.3.0
 LangSegment==0.3.5