Module setup 

xil setup — build the venv a local ML worker runs under. Port of XILU023_setup.py.

• chatterbox — venv-chatterbox for the Chatterbox Turbo TTS worker: PyTorch, chatterbox-tts and pydub.
• whisper — venv-whisper for xil stem-verify: faster-whisper (CTranslate2, no PyTorch).
• mmaudio — venv-mmaudio for --sfx-backend mmaudio: a pinned clone of hkchengrex/MMAudio installed editable, then PyTorch re-pinned afterwards (MMAudio’s unbounded torch pulls a CUDA 13 build).

Structs

Args 🔒

Opts 🔒

Spec 🔒

One xil setup target. repo (URL, ref) is cloned next to the venv as MMAudio and installed editable before packages; torch is installed last so it wins over whatever packages pulled in.

Step 🔒

Enums

Device 🔒

Installer 🔒

What builds the venv.

Target 🔒

Constants

CHATTERBOX 🔒

MMAUDIO 🔒

MMAudio pins numpy<2.1, which has no Python 3.13 wheels: default to 3.12.

MMAUDIO_WEIGHTS 🔒

The checkpoint of the worker’s model (mmaudio_worker._DEFAULT_VARIANT).

WHISPER 🔒

Functions

command

The clap definition behind --help, for man pages.

detect_cuda 🔒

device_name 🔒

execute 🔒

find_installer 🔒

find_on_path 🔒

header 🔒

index_url 🔒

needs_clone 🔒

next_steps 🔒

no_gpu_warning 🔒

plan 🔒

repo_dir 🔒

The MMAudio clone sits beside its venv.

run

target_dir 🔒

Where the venv goes: --dir, else $XIL_CODEROOT, else the workspace. The same places resolve_venv_python searches.

use_cuda 🔒

venv_parent 🔒

The directory that holds the venv (Python’s os.path.dirname).

venv_python 🔒

verify 🔒

Some(gpu_available) when the venv imports everything the worker needs.