start.sh mirrors start.bat: finds Python 3.10+, creates the venv,
installs PyTorch (CUDA index on Linux, the MPS build on macOS, or the CPU
index with --cpu), installs ACE-Step, reports the device and launches the
UI. Same flags, and the defaults can also come from the environment
(PORT=7870 ./start.sh). On macOS it passes --bf16 false, which the README
already calls for. Dropped start.sh from .gitignore, where it sat among
the upstream author's local scratch files.
Both launchers also gain two fixes found while testing on WSL:
- A venv is only accepted if pip works in it, not merely if the
interpreter exists. A directory left by an interrupted install looked
ready and then failed several steps later with a misleading "check your
internet connection". Such a venv is now recreated, and if creation
fails on Debian/Ubuntu the error points at python3-venv, which is the
actual cause there.
- The launch banner announced the URL as if the server were already up,
while model loading still had a minute to go. It now says the interface
will be available once "Running on local URL" appears.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
infer.py could only render a random example from examples/input_params:
there was no way to pass your own prompt, lyrics, duration or seed, so
using it for anything specific meant writing a separate script. Add
--prompt, --lyrics/--lyrics_file, --duration, --steps, --guidance_scale,
--scheduler, --cfg_type, --omega_scale, --seed and --format alongside the
existing runtime flags. Without --prompt the old random-example behaviour
is kept, so existing invocations are unaffected.
Also move the training-only packages out of the default install.
datasets, pytorch_lightning, matplotlib, tensorboard and tensorboardX are
imported by trainer.py and convert2hf_dataset.py, never on the inference
path, but every user was installing them — and datasets==3.4.1 is a hard
pin that drags constraints onto huggingface-hub. They now live in
requirements-train.txt behind the existing (previously ineffective)
"train" extra: pip install -e ".[train]".
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
create_text2music_ui() sorted the saved *_input_params.json files with
int(name.split('_')[1]), which assumes the generated
output_<timestamp>_<idx>_ shape. Output names are user-controlled — via
infer.py --output_path or a save_path from the UI — so any other name
raised ValueError while the Blocks were being built and took the whole
interface down before it could start:
ValueError: invalid literal for int() with base 10: 'base'
Sort by mtime instead. That is what "previous generated input params"
means anyway (newest first) and it works for any filename; a file that
disappears between listdir() and getmtime() sorts last rather than
raising.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
torchaudio 2.11 routes torchaudio.save()/load() through TorchCodec and
ignores the `backend` argument, so every generation died at the save
step with "ImportError: TorchCodec is required for save_with_torchcodec"
after the diffusion had already finished. Reference-audio loading
(audio2audio, repaint, extend) and the training dataset loader hit the
same wall.
soundfile is already a required dependency and covers all four output
formats the UI offers, so use it directly rather than pulling in
TorchCodec and its native FFmpeg stack:
- pipeline_ace_step.save_wav_file(): sf.write(), transposing
(channels, samples) -> (samples, channels); drops the now-unused
torchaudio import
- MusicDCAE.load_audio() and text2music_dataset: sf.read(dtype=float32,
always_2d=True), transposed back to (channels, samples)
torchaudio is still used for Resample/MelScale transforms, which are
unaffected.
Verified end to end: 10s generation on an RTX 3060 in 9.7s, output is
valid non-silent 48kHz stereo; load_audio round-trips it; wav/mp3/ogg/
flac all write and read back.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Gradio was unpinned, so a fresh install pulled Gradio 6, where
gr.Audio no longer accepts show_download_button — the UI crashed on
startup with TypeError. Move forward to Gradio 6 instead of pinning
back to 5:
- requirements: gradio>=6.0.0
- drop the removed show_download_button kwarg (4 call sites)
- requirements: transformers>=4.57.0 — Gradio 6 requires
huggingface-hub>=1.0, which transformers==4.50.0 forbids. Only
UMT5EncoderModel and AutoTokenizer are used, so the bump is safe.
Verified: pip check clean, UI renders and server callbacks work on
gradio 6.26.0 / transformers 5.16.1 / huggingface-hub 1.30.0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- README_RU.md: full Russian translation, linked from README.md
- start.bat: one-click setup + launch for Windows. Finds Python 3.10+,
creates the venv, installs PyTorch (CUDA or CPU) and ACE-Step, checks
the GPU, then starts the web UI. Flags: --lowvram, --cpu, --share,
--port, --device, --listen, --reinstall, --update, --setup, --help.
- Both READMEs document the launcher.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- Added automatic shape alignment for target_latents and x0
- Handles both shorter (padding) and longer (trimming) cases
- Fixes crash in extend mode with long audio files
- Minimal impact on audio quality (~0.05-0.15 sec)
Resolves issue where extend mode fails with AssertionError
when target_latents shape doesn't match x0 shape after
padding/trimming operations.
Leonid PershinandClaude Opus 5
Gong Junmin
copilot-swe-agent[bot]andChuxiJ
jackj
WebChatAppAi
Gong Junmin
rkfg
chuxij
Sayo
Samrat Barai
sean
Michael Hedman
GeorgeDrb