HML ENV

Introduction

The hml-env provides a development environment for interdisciplinary research between high-energy physics and artificial intelligence. Based on the included software and libraries, it is divided into three environments:

1. base: Provides the fundamental software for phenomenological research in high-energy physics and can generate simulated collision events.

2. lite: Builds upon base by adding support for TensorFlow 2.16.

3. dev: Builds upon lite by adding support for a complete machine learning framework, enabling comprehensive research.

Configurations

	base	lite	dev
Ubuntu 22.04	✔	✔	✔
CUDA 12.2.2		✔	✔
cuDNN 8		✔	✔
ROOT 6.28.12	✔	✔	✔
Delphes 3.5.0	✔	✔	✔
Madgraph5 (MG5) 3.4.2	✔	✔	✔
HepMC 2.06.09 (by MG5)	✔	✔	✔
LHAPDF 6.5.4 (by MG5)	✔	✔	✔
PYTHIA 8.311 (by MG5)	✔	✔	✔
Tensorflow 2.16.1 (include Keras 3)		✔	✔
PyTorch 2.3.0			✔
Jax 0.4.28			✔

• lite and dev are built based on the image nvidia/cuda:12.2.2-cudnn8-devel-ubuntu22.04. Before starting the container, you need to install the NVIDIA Container Toolkit. Please refer to here for installation instructions.

• Delphes is installed separately for performance considerations during compilation. The software installed by MG5 has also undergone some minor modifications to utilize all CPU cores during compilation.

• All three machine learning frameworks support GPU and can be used as backends for Keras versions 3 and above.

• The memory growth of TensorFlow and Jax is limited to prevent a single program from occupying the entire GPU memory at once.

Use cases

1. Launching the Container and Generating Events with Madgraph5:

   docker run -it --rm star9daisy/hml-env:3.0.0-base

   After entering the container, type mg5_aMC to open the Madgraph5 command-line interface.

2. Starting the Container in the Background and Attaching with VSCode:

   docker run -itd --gpus all star9daisy/hml-env:3.0.0-dev

   After starting, attach VSCode to the running container, allowing you to open the container as if it were a local directory.

3. The image includes an SSH service, and you can add port mappings and passwords when starting the container:

   docker run -itd --gpus all --publish <port>:22 --env PASSWORD=<password> star9daisy/hml-env:3.0.0-dev

   When this container is running on a remote server, doing so can save a step: changing from local → remote → container to local → container. Use the following command to SSH directly into the container:

   ssh -p <port> root@<remote server ip>