Environment variables¶

KonfAI uses a mix of user-facing and internal runtime environment variables.

User-facing variables¶

`CUDA_VISIBLE_DEVICES`¶

Controls which GPUs are visible to PyTorch and therefore to KonfAI.

KonfAI also rewrites this variable internally when you pass --gpu.

`KONFAI_API_TOKEN`¶

Bearer token used by:

konfai-apps in remote mode
konfai-apps-server in bearer-auth mode

Hugging Face authentication¶

The repository and CI also rely on Hugging Face-hosted assets. KonfAI itself uses huggingface_hub, so standard Hugging Face authentication variables may be relevant in practice, but they are not KonfAI-specific.

Runtime variables set by KonfAI¶

These variables are normally set by the CLI wrappers and are not expected to be managed manually in day-to-day usage.

Variable	Set by	Purpose
`KONFAI_config_file`	train/predict/evaluate wrappers	Active YAML file path.
`KONFAI_ROOT`	train/predict/evaluate wrappers	Root config object: `Trainer`, `Predictor`, or `Evaluator`.
`KONFAI_STATE`	train/predict/evaluate wrappers	Active workflow state.
`KONFAI_CHECKPOINTS_DIRECTORY`	training wrapper	Checkpoint output directory.
`KONFAI_STATISTICS_DIRECTORY`	training wrapper	Statistics output directory.
`KONFAI_PREDICTIONS_DIRECTORY`	prediction wrapper	Prediction output directory.
`KONFAI_EVALUATIONS_DIRECTORY`	evaluation wrapper	Evaluation output directory.
`KONFAI_OVERWRITE`	distributed wrapper	Mirrors the `--overwrite` flag.
`KONFAI_TENSORBOARD_PORT`	distributed wrapper	Selected TensorBoard port.
`KONFAI_VERBOSE`	distributed wrapper	Mirrors the inverse of `--quiet`.
`KONFAI_CLUSTER`	cluster wrapper	Marks cluster execution.

Internal debug/config variables¶

The codebase also references internal variables such as:

KONFAI_CONFIG_MODE
KONFAI_CONFIG_PATH
KONFAI_CONFIG_VARIABLE
KONFAI_APPS_CONFIG
KONFAI_DEBUG
KONFAI_DEBUG_LAST_LAYER

These are part of KonfAI’s internal execution model and are best treated as implementation details unless you are actively extending the framework.