Environment Variables
Configure AITraining behavior through environment variables.
Authentication
Hugging Face
export HF_TOKEN="hf_xxxxxxxxxxxxxxxxxxxxx"
- Private models
- Private datasets
- Pushing to Hub
Weights & Biases
export WANDB_API_KEY="xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
--log wandb.
GPU Configuration
Select GPUs
# Use only GPU 0
export CUDA_VISIBLE_DEVICES=0
# Use GPUs 0 and 1
export CUDA_VISIBLE_DEVICES=0,1
# Disable GPU (CPU only)
export CUDA_VISIBLE_DEVICES=""
Memory Management
# Limit memory fragmentation and enable expandable segments
export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512,expandable_segments:True
GPU Override
# Force CPU training (override GPU detection)
export AUTOTRAIN_FORCE_NUM_GPUS=0
# Force single GPU
export AUTOTRAIN_FORCE_NUM_GPUS=1
Mac (Apple Silicon) Settings
MPS Control
# Disable MPS (force CPU on Mac)
export AUTOTRAIN_DISABLE_MPS=1
# Force enable MPS even with incompatible settings
export AUTOTRAIN_ENABLE_MPS=1
When MPS is disabled or incompatible, AITraining automatically sets PYTORCH_ENABLE_MPS_FALLBACK=1 for CPU fallback.
Security Settings
Remote Code Execution
# Disable trust_remote_code for all model loading (more secure)
export ALLOW_REMOTE_CODE=false
Setting ALLOW_REMOTE_CODE=false may prevent loading models that require custom code (like some multimodal models). Default is true.
Logging
W&B Settings
export WANDB_PROJECT=my-project
export WANDB_ENTITY=my-team
export WANDB_MODE=offline # Don't sync to cloud
Hub Settings
Default Username
export HF_USERNAME=my-username
Cache Directory
export HF_HOME=/path/to/cache
export TRANSFORMERS_CACHE=/path/to/cache
Training Settings
Disable Telemetry
export HF_HUB_DISABLE_TELEMETRY=1
Offline Mode
export TRANSFORMERS_OFFLINE=1
export HF_HUB_OFFLINE=1
Usage Examples
Full Setup Script
Create a setup script setup_env.sh:
#!/bin/bash
# Authentication
export HF_TOKEN="hf_your_token_here"
export WANDB_API_KEY="your_wandb_key"
# GPU settings
export CUDA_VISIBLE_DEVICES=0,1
# Cache
export HF_HOME=~/.cache/huggingface
# W&B project settings
export WANDB_PROJECT=my-project
source setup_env.sh
aitraining llm --train ...
Using .env Files
Create a .env file:
HF_TOKEN=hf_xxxxx
WANDB_API_KEY=xxxxx
CUDA_VISIBLE_DEVICES=0
export $(cat .env | xargs)
aitraining --config training.yaml
Internal Variables
These are set automatically by AITraining:
| Variable | Value | Purpose |
|---|
TF_CPP_MIN_LOG_LEVEL | 3 | Suppress TensorFlow warnings |
TOKENIZERS_PARALLELISM | false | Disable tokenizer parallelism warnings |
BITSANDBYTES_NOWELCOME | 1 | Suppress bitsandbytes welcome message |
AUTOTRAIN_TUI_MODE | 1 | Set when running in TUI mode |
Debug Variables
| Variable | Default | Purpose |
|---|
PAUSE_ON_FAILURE | 1 | Pause Space on training failure (for Spaces backend) |
Priority
When using --config, the config file takes full control of training parameters. Environment variables are used for authentication and system settings (like HF_TOKEN, WANDB_API_KEY, CUDA_VISIBLE_DEVICES).
Next Steps
