Repository navigation
Fix the bugs reported against 0.19.0 - #737
Merged
Merged
Conversation
added 9 commits
September 30, 2026 12:25
available_torch_devices() caught only ImportError. A torch whose DLLs fail to load raises OSError instead (WinError 127 from a CUDA DLL left behind on Windows, #723), which killed startup when the device setting was "auto" and made every GET /api/settings a 500 otherwise. The probe now catches any failure, falls back to ["cpu"], logs the cause once with its traceback, and remembers it. /api/settings carries it as torch_error, and Settings replaces the device description with a warning in red: separation cannot run on any device until torch is repaired, CPU included, so a quiet fallback would hide the real state. tests/test_torch_unloadable.py installs an import hook that makes `import torch` raise that OSError; the probe and settings tests fail on the old code. Closes #730
The Now Playing heart was the only way to favourite, and daw.css hides that whole card below 1460 px. A maximised laptop window is under that, so there was no way to favourite at all, and no way to un-favourite from the Favorites view. - Library and Favorites rows get a heart next to the delete button. - Every heart goes through toggleFavorite(), which saves, repaints the Now Playing heart when it is the open track, and re-renders the list, so the hearts cannot disagree with each other or the store. - Row actions float over the row's right edge on hover or keyboard focus. In the flow the hidden delete button took about 29 px from every title at rest; a second button would have taken it again. Delete was also invisible under keyboard focus before. - A favourite shows a small heart in its subline at rest. tests/e2e/favorite-reach.spec.mjs runs at 1280 and 1600 px; all four fail on the old code. Closes #724
A CUDA 13 driver was offered cu128 alone. cu128 installs torch 2.8, whose builds start at sm_61 on Windows and sm_70 on Linux (TORCH_CUDA_ARCH_LIST in pytorch's release/2.8 .ci scripts), so a GTX 970 (sm_52) installed it, failed verification with "no kernel image is available", and dropped to CPU. - A card below compute capability 7 is never offered cu128; it gets cu124 then cu118, the torch 2.6 builds, which carry sm_50 on both platforms (release/2.6 scripts), and an sm_50 binary runs on 5.x. - A CUDA 13 driver on a newer card keeps cu128 first, with cu124 and cu118 behind it instead of CPU. New tests: a_pre_volta_card_is_never_offered_cu128 (the old code returned ["cu128"] for 5.2 under 13.0) and a_cuda_13_driver_falls_back_to_the_2_6_builds. The published-wheel check now covers caps 5.2 and 6.1. Closes #732
The analysis strip said the key three times: "D maj" on the Key card, "Major" under it, and "Major" again on a Scale card of its own, while the Key card was too narrow for a longer label. - keyLabel.js formatKey() turns the stored key and scale into one label, "B minor" or "B harmonic minor", from a translated template and mode names in all ten languages (plus the European Portuguese spelling). Note names stay letters everywhere, as DAWs write them. - The Scale card, the Major/Minor sub-line and an unused confidence label are gone; the confidence ring stays. - Key takes the Scale card's grid column, so the row keeps the seven edges the presence row lines up with. Measured at a 950 px window it still has 115 px for a label that can need 155, so the key wraps to a second line rather than clip; a short key stays on one line. - Display only: stored key, scale and key_confidence are unchanged, so nothing needs re-analysis. tests/js/key-label.test.mjs checks every language; the new tests/e2e/key-card.spec.mjs checks en, de and fr at 1366 and 950 px for the label and that it is not clipped, and fails on the old code. Closes #736
A CPU result born from a failure is re-probed on every launch (#247), which is right for a network drop. For a GPU that every offered CUDA build refuses with "no kernel image is available", it meant downloading and installing CUDA torch, failing verification and restoring CPU torch on every launch: six cycles in the #723 reporter's setup.log. - verify_cuda_torch() now says why it failed; the no-kernel case is matched on CUDA's exact wording, so driver, memory or install problems are never mistaken for it. - When every candidate installed and was refused for that reason, setup records cuda-unsupported-gpu. Any other mix keeps today's reasons. - The setup screen treats that reason as settled and says plainly why the GPU is not used. A new app version re-runs setup anyway, which is when different builds could be on offer. New tests: a_gpu_every_build_refuses_is_settled_as_unsupported and only_cudas_own_wording_counts_as_no_kernel_image. Closes #733
The CUDA install laid the CUDA wheels over the CPU ones, and the restore laid the CPU wheels back over those, both with --ignore-installed. Overlaying never removes files only the CUDA wheel has. On Windows torch loads every DLL in torch/lib, so the leftover c10_cuda.dll of one torch version failed to link against the restored CPU DLLs, and `import torch` raised WinError 127 for good while setup reported success (#723). New module torchswap: - Before the first real CUDA install the bundled torch is moved aside into .stemdeck-torch-backup next to site-packages. What belongs to torch is read from pip's RECORD files; on a real install that is only torch, functorch, torchgen, torchaudio, torio, torchvision and their dist-infos. - A manifest is written and synced before anything moves, so a crash at any point is recoverable; the next setup run restores it first. - Each candidate installs into clean directories; a failure deletes the CUDA build and renames the original back, with no network. Success deletes the backup. After any restore `import torch` is now checked. Without a backup (the move failed, or the install broke before this fix) the CPU wheels are reinstalled, and if torch still does not import, everything torch owns is deleted and they are installed once more into clean directories. That is what repairs an install already carrying the stray DLL. Whether a working CUDA build is already present is now asked once, before anything moves, instead of inside each install. Eight new tests on real temp directories, including the reporter's case (no c10_cuda.dll after rollback, c10.dll byte for byte) and a snapshot interrupted part way. Closes #731
Key detection multiplied each key's profile correlation by how loud its
root was. In a minor song the relative major's root is the minor chord's
third and sits in most of a typical loop's chords, so it is usually
louder: every i-VI-III-VII loop came out as its relative major at up to
100% confidence ("Plug in Baby", B minor, read D major). Harmonic minor
was never reported, and confidence changed with the input's level.
- Keys are ranked by correlation alone, which is level-invariant.
- The keys close to the best (within 0.1) and their relatives are then
weighed against the song's edges: how well the first and, when the
whole song was analysed, last 4 s of music fit each key's tonic triad
and bass note. A song starts and ends on its tonic far more often than
anywhere else, which is what a whole-song histogram cannot show.
- A minor key is reported as harmonic minor when its raised seventh
outweighs the flat seventh 1.3 to 1.
- After separation the key is detected again from the stems: harmony
stems without drums, the bass stem as the bass line, and the whole
song. Any failure keeps the estimate from the mix.
- Chroma uses 93 ms frames: 4.1 s instead of 8.6 s on a 9.5 min track's
stems, same score. The stems pass takes about 1 s per minute of audio.
Measured on tests/keyset.py, 33 synthetic songs with known keys (chord
loops over a bass line, 11 progressions in 3 keys), key and scale both
right:
before 21 of 33 (27 keys right)
after 30 of 33 from the mix, 30 of 33 from stems
The three misses are vi-IV-I-V, which starts and ends on its minor
chord. A real 9.5 min track (E minor) reads the same before and after.
Already analysed tracks keep their stored key; nothing is re-analysed.
tests/test_key_detection.py scores the set and the reporter's loop, and
fails on the old detector.
Closes #726
A separate "key" stage timing broke the timings' order, which tests/test_identify_regressions.py holds as a contract: new entries may only follow the existing stages. The pass is short and already inside post, so it is counted there. Refs #726
An ignored test, run by hand: copies a real site-packages torch (1.2 GB here), moves it aside, lays a CUDA build with a stray c10_cuda.dll into its place, rolls back, and compares every file's SHA-256 with the original. Passed against the dev venv's torch 2.6.0+cpu, which then imported and ran a tensor op from the restored copy. Refs #731
Merged
thcp
added a commit
that referenced
this pull request
Oct 2, 2026
Release 0.19.1: everything on `next-release` since v0.19.0. - #727: slowed playback without stutter, doubled kicks or echo (Signalsmith Stretch as the tempo stage). - #737: the bugs reported against 0.19.0 (CUDA setup, torch that cannot load, favourites in a narrow window, minor keys, the key label). - #740: the remaining open issues (top bar and Extract stems, the card at every width, favourites on the phone, the lyrics lookup on the server, an adjustable slow speed, Rust advisories). - #741: trivy ignores the two urllib3 CVEs until the lock next moves. - #743: a 0.5x button beside the slow speed. Verified on local Windows NVIDIA builds: 0.19.1.dev0 passed 12 of 12 (#727, #737), and 0.19.1.dev1 passed 18 of 18 automated checks including #740, in an isolated profile. `uv.lock` is unchanged since v0.19.0, so the in-app update works. Closes #722 Closes #728 Closes #729 Closes #723 Closes #724 Closes #726 Closes #730 Closes #731 Closes #732 Closes #733 Closes #736 Closes #701 Closes #738 Closes #725 Closes #735 Closes #734 Closes #719 Closes #742
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #723
Closes #724
Closes #726
Closes #730
Closes #731
Closes #732
Closes #733
Closes #736
The bugs reported against 0.19.0 by other people, one commit per issue. Found along the way and tracked, but not in this PR: #734 (favourites on the phone UI, needs server storage) and #735 (About this song unreachable below 1460 px).
CUDA setup on Windows (#723, split into #730 to #733)
The reporter's GTX 970 went through a chain of four failures:
cu128only, which installs torch 2.8. Its Windows builds start at sm_61 and its Linux builds at sm_70 (checked against pytorch's release/2.8.ciscripts), so a Maxwell card could never pass. Cards below compute capability 7 now getcu124thencu118(torch 2.6, which carries sm_50 on both platforms), and newer cards under a CUDA 13 driver fall back to those instead of CPU.c10_cuda.dllmadeimport torchfail for good. The bundled torch is now moved aside before the first CUDA install (what belongs to torch is read from pip'sRECORDfiles, with a manifest synced to disk first so a crash is recoverable) and renamed back on failure: exact, and no network. After any restore,import torchis checked, and an install already carrying the stray DLL is cleared and reinstalled into clean directories.ImportError, so thatOSErrorkilled startup and made/api/settingsa 500. It now falls back to CPU, logs the cause once, and Settings says in red that separation cannot run until torch is repaired, in all ten languages.Favourites below 1460 px (#724)
The Now Playing heart was the only way to favourite, and that card is hidden below 1460 px. Library and Favorites rows now have a heart next to delete, all hearts go through one
toggleFavorite(), and row actions float over the row edge on hover or keyboard focus, which gives titles back the ~29 px the hidden delete button always took. The reporter's other request, removing the music icon, was declined earlier in #636 and is not changed.Key detection (#726) and the key label (#736)
tests/keyset.py, 33 synthetic songs with known keys, key and scale both right: 21 of 33 before, 30 of 33 after, on the mix and on stems. The three misses arevi-IV-I-V, which starts and ends on its minor chord. The stems pass costs about 1 s per minute of audio.Tests
test_torch_unloadable.py,test_key_detection.py(with the labelled set)favorite-reach.spec.mjs(1280 and 1600 px),key-card.spec.mjs(en, de, fr at 1366 and 950 px)key-label.test.mjs, every languageGate
uv.lockunchanged-D warnings, 128 testsVerification beyond the tests
c10_cuda.dll, and rolled back. Every file's SHA-256 matched the original, and the restored torch imported and ran.STEMDECK_ACCEPTANCE_SKIP_LONG=1:Not caused by this branch
uv.lock; fixing it means a lock change, which sends desktop users a full download, so it is a separate decision.