LlamaFactory Adapter Plugin

Configure environment variables

Before running a training task, configure these environment variables in bash:

Example:

export ECO_CLIENT_ID="userxx@xx.com"
export ECO_API_KEY="XXXX"

You can also write them to ~/.bashrc and activate the shell configuration:

vim ~/.bashrc
source ~/.bashrc

Security note: never commit ECO_API_KEY to repositories, shared logs, or public tickets. Use environment variables or a controlled secret manager for collaboration.

Start a training task

Replace these values based on your environment:

• Data path
• Model path
• Training argument file path
• accelerate configuration file path

Complete training example:

export ECO_CLIENT_ID="userxx@xx.com" && \
export ECO_API_KEY="XXXX" && \
CUDA_VISIBLE_DEVICES=0,1 accelerate launch --config_file fsdp_config.yaml \
  --main_process_port 29501 src/train.py train_args.yaml

If the task needs to preserve the current training state so it can continue after being stopped, set the following value in train_args.yaml:

save_only_model: false

This setting saves optimizer state, learning-rate scheduler state, model weights, and other training state information.

Plugin log reference

1. Plugin imported and initialized successfully

The log includes:

[EcoPhase] ✅ EcoTrainAPI initialized.

2. Plugin enabled

The log includes:

[EcoPhase] ✅ API is enabled.

3. Plugin inactive

The log includes:

[EcoPhase] 🛑 API is disabled.

4. Early stop triggered

The system automatically saves the model and prints a training summary, for example:

✅ Task early stopped at step 200/2000. Reduction: 90.0%. Saved GPU-hours: 1.03.

This means:

Notes

1. The plugin initialization code must be inserted at the correct position in trainer.py.
2. Before running, confirm that the code uses enabled=True.
3. Check API is enabled / API is disabled first in the logs.
4. The plugin currently supports data parallelism only. Other parallel modes are not supported yet.
5. The plugin currently supports Linux only.