Update app.py

app.py

@@ -19,9 +19,8 @@ except Exception as e:
     print(f"Error detecting GPU. Using CPU. Error: {e}")
     torch_device = torch.device("cpu")
 
-# Load the SemantiCodec model
+# Load the SemantiCodec model without specifying a device
 semanticodec = SemantiCodec(token_rate=100, semantic_vocab_size=32768)
-semanticodec.to(torch_device)
 
 # Global variables for cancellation
 cancel_encode = False
@@ -105,18 +104,16 @@ def decode_audio(encoded_file_path):
             tokens_numpy = np.frombuffer(tokens_numpy_bytes, dtype=np.int64).reshape(shape)
 
         # Move the tensor to the same device as the model
-        tokens = torch.from_numpy(tokens_numpy).to(device=semanticodec.device)
-
-        print(f"Tokens shape: {tokens.shape}, dtype: {tokens.dtype}, device: {tokens.device}")
-        print(f"Model device: {semanticodec.device}")
-
-        # Ensure all model parameters are on the correct device
-        semanticodec.to(semanticodec.device)
+        tokens = torch.from_numpy(tokens_numpy)
+        print(f"Tokens device: {tokens.device}")
+        print(f"Model device: {next(semanticodec.parameters()).device}")
 
         # Decode the audio
         with torch.no_grad():
             waveform = semanticodec.decode(tokens)
 
+        print(f"Waveform device: {waveform.device}")
+
         # Move waveform to CPU for saving
         waveform_cpu = waveform.cpu()
 
@@ -149,13 +146,9 @@ async def stream_decode_audio(encoded_file_path) -> Generator[tuple, None, None]
             tokens_numpy = np.frombuffer(tokens_numpy_bytes, dtype=np.int64).reshape(shape)
             
         # Move the tensor to the same device as the model
-        tokens = torch.from_numpy(tokens_numpy).to(device=semanticodec.device)
-
-        print(f"Streaming tokens shape: {tokens.shape}, dtype: {tokens.dtype}, device: {tokens.device}")
-        print(f"Model device: {semanticodec.device}")
-
-        # Ensure all model parameters are on the correct device
-        semanticodec.to(semanticodec.device)
+        tokens = torch.from_numpy(tokens_numpy)
+        print(f"Streaming tokens device: {tokens.device}")
+        print(f"Model device: {next(semanticodec.parameters()).device}")
 
         # Decode the audio in chunks
         chunk_size = sample_rate * 2  # Adjust chunk size as needed