mirror of
https://github.com/nomadkaraoke/karaoke-gen.git
synced 2026-11-20 13:07:57 +09:00
## Summary First production make-it quick version (job `405eed71`, Rihanna "Push Up On Me") failed with `render timed out`: - **Render never ended**: the GPU image ships ffmpeg **4.4** (Ubuntu 22.04). With the guide-vocals `amix` in the filtergraph plus the infinite lavfi colour + looped PNG inputs, it ignores `-shortest` — reproduced in `ubuntu:22.04`: a 20 s song produced 540 s+ of video until killed. ffmpeg 8 (local dev) stops fine, which is why local tests passed. Fix: every input and the output are bounded with `-t <duration>` (colour source `d=`). Verified in `ubuntu:22.04`: 3.7 s render, exactly 20.0 s audio+video. - **Quick separation too slow**: the baked roformer took 14 s load + 124 s for a 3:15 song on the L4. Switched to fastgen's `UVR-MDX-NET-Inst_HQ_4.onnx`. It's downloaded on load until the next GPU base rebuild bakes it (`download_models.py` `EXTRA_MODELS`), with a fallback to the baked roformer if the load fails. Prod timeline observed for that job (for the <4 min goal): job created 05:22:12 → **audio-download-job container took 3m46s to start** (existing issue, 1m40–4m across recent executions; separate follow-up) → audio in GCS 05:26:46 → GPU worker started 05:27:10 → quick separation done 05:29:49 → render hung. ## Testing - `backend/tests/test_quick_version.py`: 40 pass (new: every stream bounded to the duration; model fallback; both-fail raises). - Docker `ubuntu:22.04` ffmpeg 4.4.2 reproduction before/after (see above). 🤖 Generated with [Claude Code](https://claude.com/claude-code) <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **New Features** * Quick audio separation now uses a faster model by default, with an alternate model available if the primary model cannot be loaded. * **Bug Fixes** * Quick video rendering is now limited to the song’s duration, helping prevent renders from continuing beyond the audio or failing to finish. * Model loading failures are handled more reliably: if both the primary and alternate models fail to load, the error is reported. <!-- end of auto-generated comment: release notes by coderabbit.ai --> Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>