use gpu when possible

app.py

@@ -1,9 +1,10 @@
+import spaces
 import gradio as gr
 import torch
 import numpy as np
 from synthesize import synthesize
 
-
+@spaces.GPU
 def text_to_speech(text, speaker_id, cfg_scale, num_sampling_steps):
     audio, sample_rate = synthesize(
         text=text,

gaussian_diffusion.py

@@ -202,7 +202,7 @@ class GaussianDiffusion:
         )
 
         # convert all numpy arrays to torch tensors
-        DEVICE = th.device("cpu")
+        DEVICE = th.device("cuda") if th.cuda.is_available() else th.device("cpu")
         self.betas = th.from_numpy(self.betas).to(DEVICE)
         self.alphas_cumprod = th.from_numpy(self.alphas_cumprod).to(DEVICE)
         self.alphas_cumprod_prev = th.from_numpy(self.alphas_cumprod_prev).to(DEVICE)

requirements.txt

@@ -1,7 +1,7 @@
 # This file was autogenerated by uv via the following command:
-#    uv pip compile pyproject.toml -o requirements.txt --python-platform x86_64-unknown-linux-gnu --emit-index-url --extra-index-url https://download.pytorch.org/whl/cpu
+#    uv pip compile pyproject.toml -o requirements.txt --python-platform x86_64-unknown-linux-gnu --emit-index-url --extra-index-url https://download.pytorch.org/whl/cu118
 --index-url https://pypi.org/simple
---extra-index-url https://download.pytorch.org/whl/cpu
+--extra-index-url https://download.pytorch.org/whl/cu118
 
 aiofiles==23.2.1
     # via gradio
@@ -174,13 +174,13 @@ sympy==1.13.1
     # via torch
 tomlkit==0.13.2
     # via gradio
-torch==2.5.1+cpu
+torch==2.5.1
     # via
     #   diffusion-speech-360h (pyproject.toml)
     #   encodec
     #   torchaudio
     #   vocos
-torchaudio==2.5.1+cpu
+torchaudio==2.5.1
     # via
     #   encodec
     #   vocos