Select the device rather than hard-coding CUDA

stable_audio_runtime.py

@@ -6,6 +6,7 @@ import sys
 import tempfile
 import threading
 from pathlib import Path
+import torch
 
 MODEL_REPO = "stabilityai/stable-audio-3-small-music"
 MODEL_REVISION = "0fef1392cd842149a2b6d445e181c97608faac06"
@@ -72,7 +73,12 @@ def _load_model():
         finally:
             pretrained.hf_hub_download = original_download
 
-        model = model.to("cuda").to(torch.float16)
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+        model = model.to(device)
+        if device == "cuda":
+            # Half precision is a GPU optimization, and CPU kernels for
+            # float16 are either missing or slower than float32
+            model = model.to(torch.float16)
         model.eval().requires_grad_(False)
         _MODEL = model
         _MODEL_CONFIG = config