-
Notifications
You must be signed in to change notification settings - Fork 14
Troubleshooting
Static glitch output on the Medium model. Flash Attention is not installed correctly. Verify it with uv run python -c "from flash_attn import flash_attn_func; import flash_attn; print(flash_attn.__version__)" and reinstall a wheel matching the Python, torch, and CUDA combination from kingbri1/flash-attention.
"API UNREACHABLE" banner. The backend is not listening on port 8600. Test it with curl http://localhost:8600/api/health. On Windows, .\theDAW.bat clears stale processes automatically.
Out-of-memory on the Medium model. The Medium pipeline needs roughly 8 GB of VRAM. The small model, a shorter duration, or freeing competing CUDA processes resolves it.
Library slow or failing to save. Confirm the backend is running on port 8600, since the list loads once it reports ready, and free disk space if writes begin to fail.
Magenta sidecar shows empty port_src/. The upstream engine source is a git submodule. Clone theDAW with --recurse-submodules, or for the standalone studio use the release ZIP.
User Guide §23 has the full matrix.