Skip to content

Fix MPS audio decoding failure with explicit dtype preservation #22

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
lmangani merged 3 commits into from
Jan 30, 2026
Merged

Fix MPS audio decoding failure with explicit dtype preservation #22

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1431967
Initial plan
Copilot Jan 30, 2026
033a8a4
Fix MPS FloatTensor error by preserving torch.long dtype during frame…
Copilot Jan 30, 2026
f148354
Update comment to be more accurate about MPS dtype behavior
Copilot Jan 30, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion backend/app/services/music_service.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -804,7 +804,9 @@ def _pad_audio_token(token):
progress = int((i + 1) / max_audio_frames * 100)
callback(progress, f"Generating audio... {i + 1}/{max_audio_frames} frames")

frames = torch.stack(frames).permute(1, 2, 0).squeeze(0).cpu() # Move to CPU immediately
# Stack frames and explicitly preserve torch.long dtype (critical for MPS compatibility)
# Explicitly ensure torch.long dtype is preserved (defensive fix for potential MPS backend issues)
frames = torch.stack(frames).permute(1, 2, 0).squeeze(0).to(dtype=torch.long).cpu()

# Sequential offload: Move HeartMuLa to CPU before loading HeartCodec
# This allows fitting on smaller GPUs (12GB) by never having both models in VRAM
Expand Down