Merge pull request #10 from audiohacking/copilot/fix-generation-failed-issue

Fix MPS out of memory error by disabling PyTorch memory watermark

Author: lmangani

backend/app/services/music_service.py

@@ -1858,6 +1858,7 @@ def _pipeline_callback(progress, msg):
 
                 def _run_pipeline():
                     os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
+                    os.environ["PYTORCH_MPS_HIGH_WATERMARK_RATIO"] = "0.0"
 
                     sound_tags = request.tags if request.tags and request.tags.strip() else "pop music"
 

launcher.py

@@ -55,6 +55,7 @@ def setup_environment():
     os.environ["HEARTMULA_GENERATED_AUDIO_DIR"] = str(generated_audio_dir)
     os.environ["HEARTMULA_REF_AUDIO_DIR"] = str(ref_audio_dir)
     os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"  # Enable Metal Performance Shaders
+    os.environ["PYTORCH_MPS_HIGH_WATERMARK_RATIO"] = "0.0"  # Disable MPS memory limit to prevent OOM errors
 
     # Change working directory to app bundle for frontend serving
     # Backend will use absolute paths from environment variables