Skip to content

linux-cli.md

Latest commit

 

History

History
90 lines (71 loc) · 3.34 KB

linux-cli.md

File metadata and controls

90 lines (71 loc) · 3.34 KB
Raw

Installation, commands and configuration

Use an existing Python environment with the appropriate PyTorch build. A package installation creates comfyquant and comfyui-native-quantizer console scripts in that environment's bin directory:

/path/to/environment/bin/python -m pip install /path/to/package.whl
/path/to/environment/bin/comfyquant --version
/path/to/environment/bin/python -m native_quantizer --no-config --list-formats

Activate the environment or add its bin directory to PATH to use the short commands. For a local wheel rebuilt with the same version, force replacement without changing compute dependencies:

/path/to/environment/bin/python -m pip install --no-index --no-deps \
  --force-reinstall /path/to/package.whl

Check installed imports and run a synthetic conversion outside the source checkout. An editable install loads the checkout and does not validate wheel contents.

Shell completion and source-checkout launchers

comfyquant --completions fish, bash or zsh emits completion definitions. Place them in the shell's normal completion directory.

The optional scripts/install_linux.py helper installs user-local launchers, completion files and XDG defaults from a source checkout:

python scripts/install_linux.py --offline-default

Its generated launchers load that checkout directly; keep the checkout present. For installed-wheel testing, use the environment's console scripts instead. --python chooses the interpreter. --cuda-default verifies CUDA before saving that preference. The helper does not install packages, download models or modify shell startup files.

Defaults

CLI defaults live in $XDG_CONFIG_HOME/comfyui-native-quantizer/config.json, or ~/.config/comfyui-native-quantizer/config.json when XDG is unset:

{
  "schema": 1,
  "device": "auto",
  "target_runtime": "auto",
  "max_memory": "512M",
  "offline": true,
  "optimizer": "none",
  "mse_objective": "auto",
  "mse_steps": 17,
  "mse_iterations": 3
}

CLI flags override the file. --config PATH selects another file; --no-config uses built-in defaults. Invalid or unknown values fail clearly. Overwrite and resume cannot be enabled through this file. Python API requests remain explicit.

device=auto uses an available accelerator or CPU. Explicit CUDA/ROCm requests fail when unavailable. --device cpu overrides a saved accelerator preference. The memory budget estimates temporary conversion work, excluding mmap pages, framework overhead and allocator caches.

offline=true prevents model downloads; a deliberate --no-offline overrides it. Repository metadata browsing remains available. Development checks and benchmarks use synthetic fixtures rather than downloading weights.

Inspection and scripting

comfyquant --no-config --diagnose --json
comfyquant --no-config --list-formats --json
comfyquant --no-config --list-architectures --json
comfyquant --no-config --audit-architectures /path/to/ComfyUI --json
comfyquant --no-config input.safetensors --inspect --json
comfyquant --no-config --verify-output output.safetensors --json

Help, version, catalogs, completion and architecture audits start without PyTorch. Exit codes are 0 for success, 1 for operation failure, 2 for usage errors and 130 for keyboard cancellation. Conversion progress uses stderr; JSONL events are available with --events-jsonl PATH.