Merge pull request #21 from audiohacking/copilot/fix-decoding-audio-error-again

Fix MPS tensor dtype in audio generation pipeline

Author: lmangani

backend/app/services/music_service.py

@@ -757,6 +757,8 @@ def generate_with_callback(inputs, callback=None, **kwargs):
                 continuous_segments=continuous_segment,
                 starts=starts,
             )
+        # Convert to long immediately after generation (tokens must be integers)
+        curr_token = curr_token.long()
         frames.append(curr_token[0:1,])
 
         def _pad_audio_token(token):
@@ -791,6 +793,8 @@ def _pad_audio_token(token):
                     continuous_segments=None,
                     starts=None,
                 )
+            # Convert to long immediately after generation (tokens must be integers)
+            curr_token = curr_token.long()
             if torch.any(curr_token[0:1, :] >= pipeline.config.audio_eos_id):
                 break
             frames.append(curr_token[0:1,])