Use CUDA for audio decoding unless you are on 'mps' and need to use CPU (#98)

Previously, both CUDA and MPS tensors were being moved to CPU for audio
decoding. This was only necessary for MPS devices due to embedding operation
limitations. Now CUDA tensors remain on GPU for decoding, improving performance
while maintaining MPS compatibility.

Author: vvaidy

examples/generation.py

@@ -374,7 +374,7 @@ def generate(
         concat_audio_out_ids = torch.concat(audio_out_ids_l, dim=1)
 
         # Fix MPS compatibility: detach and move to CPU before decoding
-        if concat_audio_out_ids.device.type in ["mps", "cuda"]:
+        if concat_audio_out_ids.device.type == "mps":
             concat_audio_out_ids_cpu = concat_audio_out_ids.detach().cpu()
         else:
             concat_audio_out_ids_cpu = concat_audio_out_ids