Environment setup
Use separate Python 3.12.13 environments for Actor/Judge and Editor.
One-command setup
Create both environments for single-image inference:
bash scripts/setup_envs.sh --profile inference
Create and verify the CPU release-test environment:
bash scripts/setup_envs.sh --profile test
Create the training and test environments together after obtaining the FlashAttention wheel that matches Python, PyTorch, and CUDA:
FLASH_ATTN_WHEEL=/absolute/path/to/validated_flash_attn.whl \
bash scripts/setup_envs.sh --profile all
Use --profile training to omit the CPU test environment, --no-verify to
skip post-install checks, or --dry-run to inspect every command without
changing the machine. The training and all profiles create .env from
.env.example when needed, but local model and dataset paths still have to be
filled in.
Actor and Judge
conda env create -f environment/actor-judge.yml
conda activate mr_iqa_actor_judge
python -m pip install --upgrade pip
python -m pip install -r requirements/actor-judge.txt
Full training/evaluation requires a CUDA 13.0-compatible driver and eight visible NVIDIA GPUs.
The launchers also require a prebuilt FlashAttention wheel. Configure it in
your private .env using .env.example; the launcher validates the artifact
automatically. Install that same wheel into the Actor/Judge environment before
the first preflight:
python -m pip install /path/to/validated_flash_attn.whl
python -c 'import flash_attn; print(flash_attn.__version__)'
The wheel must match the Python, PyTorch, and CUDA ABI.
Editor
Conda setup:
conda env create -f environment/editor.yml
conda activate mr_iqa_editor
python -m pip install --upgrade pip
python -m pip install -r requirements/editor.txt
Set DIFFUSERS_VENV and DIFFUSERS_MODEL_PATH in .env.
CPU release checks
The repository's format, privacy, data-integrity, and contract tests do not load model weights or initialize CUDA:
python -m venv .venv/release-test
source .venv/release-test/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements/test.txt
python -m pip install torch==2.11.0 \
--index-url https://download.pytorch.org/whl/cpu
bash scripts/test_release.sh
Use bash scripts/test_release.sh --static before installing test dependencies.
The one-command test profile uses environment/test.yml and runs the same
suite after installation.
Configuration and provenance
Copy .env.example to .env and fill in local paths. Never commit .env or
tokens. Launchers validate configured artifacts automatically.
Capture a sanitized runtime manifest alongside every run without recording a hostname, username, or local path:
python environment/capture_runtime.py --role actor-judge > runtime-actor-judge.json
python environment/capture_runtime.py --role editor > runtime-editor.json