# Isolated venv for ACE-Step (ACE-Step/ACE-Step-v1-3.5B).
#
# ACE-Step is a text-to-music model (Apache-2.0 weights) that ships its own
# package. It pins a torch/diffusers/transformers combination the coderai
# server cannot hold -- the server runs transformers 5.x and a much newer
# diffusers -- so it runs in its own venv and coderai talks to it over HTTP,
# the same shape as CrisperWhisper and NeMo-Canary
# (codai/api/acestep_worker.py + tools/acestep_service.py, both still to be
# written). Do NOT merge these into requirements.txt.
#
# NOT YET INSTALLED. The `acestep` package is not on PyPI, so it installs from
# the upstream git repository, and nothing here has been run or verified. The
# audio_gen loader therefore refuses an ACE-Step model with an explicit
# message rather than loading it as a MusicGen and failing with a shape error
# minutes later. Install this venv, then introspect the package before
# writing the service -- the API below the pin is not guessed at here.
#
# CUDA build for the 3090; cu124 wheels run on this box's newer driver.

--extra-index-url https://download.pytorch.org/whl/cu124
torch==2.5.1+cu124
torchaudio==2.5.1+cu124
soundfile>=0.12.0

# ACE-Step itself, from the upstream repository. Pin the commit you install so
# a rebuild is reproducible:
#   acestep @ git+https://github.com/ace-step/ACE-Step.git@<commit>
