ML Frameworks
PyTorch and TensorFlow are available on both Pegasus and Triton/Summit, but the environment must match the system architecture and GPU software stack.
This page covers non-container GPU framework workflows. For container-based GPU workflows, see the Apptainer GPU Containers page.
Important
GPU framework code must run on GPU compute nodes. Request GPU resources through LSF before running PyTorch or TensorFlow GPU workloads.
Availability
Framework |
Pegasus |
Triton/Summit |
Notes |
|---|---|---|---|
PyTorch |
Available |
Available |
Use architecture-compatible Python, Conda, module, JupyterHub, or container environments. |
TensorFlow |
Available |
Available |
GPU support depends on CUDA, cuDNN, driver, and architecture compatibility. |
Start a GPU Session
Pegasus H100 example:
bsub -P <project> -Is -q gpu_h100 -W 00:30 -n 1 -gpu "num=1" bash
Triton/Summit example:
bsub -P <project> -Is -q interactive -W 00:30 -n 1 -gpu "num=1" bash
Check the host GPU:
hostname
uname -m
nvidia-smi
Architecture Notes
Pegasus uses x86_64 architecture. Standard linux-64 Python, Conda, and
framework packages are generally the starting point.
Triton/Summit uses ppc64le architecture. Standard amd64 or x86_64
packages and images may not work. Use Power-compatible environments such as
Open-CE, RocketCE, compatible Conda packages, or Apptainer images built for
ppc64le.
Caution
Do not assume that a PyTorch or TensorFlow environment created on Pegasus will work on Triton/Summit. The architecture and package availability differ.
Check PyTorch
After starting a GPU session and loading or activating the desired environment:
python -c "import torch; print('torch:', torch.__version__); print('cuda available:', torch.cuda.is_available()); print('gpu:', torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'no GPU')"
A GPU-capable PyTorch environment should report:
cuda available: True
Check TensorFlow
After starting a GPU session and loading or activating the desired environment:
python -c "import tensorflow as tf; print('tensorflow:', tf.__version__); print('gpus:', tf.config.list_physical_devices('GPU'))"
A GPU-capable TensorFlow environment should return one or more GPU devices.
Caution
TensorFlow GPU environments are sensitive to CUDA, cuDNN, and driver compatibility. If TensorFlow imports but shows an empty GPU list, check the GPU allocation, environment, architecture, and framework build.
Pegasus Notes
Pegasus GPU workflows may use queues such as gpu_h100 when the project has
access. Pegasus uses x86_64 architecture.
Use supported modules, Miniforge environments, JupyterHub kernels, or Apptainer containers that match the Pegasus GPU software stack.
Triton/Summit Notes
Triton/Summit GPU workflows may use the interactive queue with
-gpu "num=1" when the project has access. Triton/Summit uses ppc64le
architecture.
Use Power-compatible environments such as Open-CE/RocketCE or other
ppc64le-compatible framework builds.
Recommended Practice
Check
nvidia-smion the host before checking the framework.Use the correct GPU queue and project allocation.
Use architecture-compatible packages or containers.
Use JupyterHub for interactive testing when appropriate.
Use LSF batch jobs for production GPU runs.
Use Apptainer when a portable or isolated framework stack is required.