Currently, ANT supports single-node multi-GPU settings, with multi-node support planned for future development.

The primary objective of ANT is to efficiently schedule jobs and allocate the requested GPU resources.

ANT is built and tested with the following dependencies:

Assuming you have a conda installation, the necessary environment can be created by running:

conda create -n ant2 python=3.11 conda-forge::nodejs==24.4.1 -y
conda activate ant2
pip install -r requirements.txt
bash setup.sh

setup.sh will build the frontend and generate necessary certificates.

Finally, launch ANT using:

python run.sh [gpu_ids separated by comma]

# Example (Selecting the first 4 GPUs):
python run.py  0,1,2,3

By default, this will load the configuration from config/default.json and host a web interface at https://0.0.0.0:6060. The backend status can be checked by curling as follows:

curl --insecure https://0.0.0.0:6060/api

Head over to the Create New Task tab and type the following in the commands box:

echo "Hello World from ANT!"

Hit the SUBMIT button and watch your commands got executed! ANT will also automatically save your stdout logs (similar to using tee or >>). Under default configurations, the logs will be saved at ./ant_runner_logs.

Intuitively, you can view all ongoing and completed tasks in their respectives tabs. There, you can easily view terminal logs, download, copy-commands, etc.

ANT supports any single-line command. For sequential execution of multiple commands, please use &&.

If your conda environment is necessary for your job, please use conda run instead of conda activate. Example:

cd /path/to/my/project && conda run --live-stream -n my_env python ...

Note that --live-stream is necessary for the conda run to live-stream the output to stdout. Otherwise, no output will be printed.

ANT features a built-in randomizer, particularly useful for distributed training that requires assigning a specific port.

# Randomizing integer
{rand int 4000 5000}

# Randomizing float
{rand float 1.45 5.65}

# Note that this syntax can be substituted like an f-string in your commands. Example:
PORT={rand int 4000 5000} python myscript.py
python myscript.py --seed {rand float 3.4 6.4}

ANT also support queuing multiple commands. To achieve this, select the "Multi" queue mode in the Create New Task page. Multiple commands can be seperated using new lines & each command can be extended to the following lines by adding \ at the end (just like you would on terminals). Lines with leading # will be ignored.

To configure running parameters, there two arguments can be used: ant_n_gpus : int = 1 & ant_task_id : str = uuid.uuid4()

# Running three commands with partially-defined parameters:
ant_n_gpus=4 ant_task_id="first_task" python first_task_.py \
--dataset my_dataset \
--batch_size 4
ant_n_gpus=2 python second_task.py \
--batch_size 8
python thrid_task.py

Note that if multiple ANT arguments present, the only the last one will take effect. If none is present, the default value (randomized task_id & 0 n_gpus) will be used

In previous versions of ant, commands can be very long and tedious to set up, hence we have integrated several special environment variables to improve QOL.

Hence, instead of appending:

cd /my/work/dir && /home/anaconda/bin/conda run --live-stream -n my_env mycommand

You can simply use the following environment variable in the Create New Task page:

Environment variables will be saved internally and applied to all commands if Multi Queue mode is selected.

This feature monitors GPU usage and detects if a GPU is being utilized by processes outside of ANT. If the GPU's average usage or memory utilization exceeds 50% for a consecutive 20-second period, ANT will mark the GPU as BUSY.

Enable this behavior by setting ADGS_enabled=true in your config. This feature is not enabled by default.

• Multi-node support

Web Template: Creative Tim.