Environment Setup
Overview
All platform images run Ubuntu; most are Ubuntu 18.04, with a few running Ubuntu 20.04.
The platform includes pre-installed images for the following frameworks and versions; instances using these images will come with the corresponding framework software pre-installed. If the pre-installed framework versions or Python versions listed below do not meet your needs, please continue reading below for instructions on configuring other versions of the frameworks or Python.
| Framework | Framework Version | Python Version | CUDA Version |
|---|---|---|---|
| PyTorch | 1.1.0 | 3.7 | 10.0 |
| PyTorch | 1.5.1 | 3.8 | 10.1 |
| PyTorch | 1.6.0 | 3.8 | 10.1 |
| PyTorch | 1.7.0 | 3.8 | 11.0 |
| PyTorch | 1.8.1 | 3.8 | 11.1 |
| PyTorch | 1.9.0 | 3.8 | 11.1 |
| PyTorch | 1.10.0 | 3.8 | 11.3 |
| PyTorch | 1.11.0 | 3.8 | 11.3 |
| PyTorch | 2.0.0 | 3.8 | 11.8 |
| PyTorch | 2.1.0 | 3.10 | 12.1 |
| PyTorch | 2.1.2 | 3.10 | 11.8 |
| PyTorch | 2.3.0 | 3.12 | 12.1 |
| PyTorch | 2.5.1 | 3.12 | 12.4 |
| PyTorch | 2.7.0 | 3.12 | 12.8 |
| PyTorch | 2.8.0 | 3.12 | 12.8 |
| TensorFlow | 1.15.5 | 3.8 | 11.4 |
| TensorFlow | 2.5.0 | 3.8 | 11.2 |
| TensorFlow | 2.9.0 | 3.8 | 11.2 |
| Miniconda | conda3 | 3.7 | 9.0 |
| Miniconda | conda3 | 3.8 | 10.1 |
| Miniconda | conda3 | 3.8 | 10.2 |
| Miniconda | conda3 | 3.8 | 11.1 |
| Miniconda | conda3 | 3.8 | 11.3 |
| Miniconda | conda3 | 3.8 | 11.3(cudagl) |
| Miniconda | conda3 | 3.8 | 11.6 |
| Miniconda | conda3 | 3.8 | 11.8 |
| Miniconda | conda3 | 3.10 | 11.8 |
| tritonserver | 24.12 | 3.12 | 12.6 |
| JAX | 0.3.10 | 3.8 | 11.1 |
| PaddlePaddle | 2.2.0 | 3.8 | 11.2 |
| PaddlePaddle | 2.4.0 | 3.8 | 11.2 |
| TensorRT | 8.5.1 | 3.8 | 11.8 |
| TensorRT | 8.6.1 | 3.8 | 11.8 |
| Gromacs | 2022.2 | 3.8 | 11.4 |
| Gromacs | 2023.2 | 3.10 | 11.8 |
Installing Other Versions of Python: Reference
Installing Other Versions of CUDA: Reference
Installing PyTorch: Reference
When installing using the official Torch conda command in China, the version installed is typically the CPU version (which has bugs) rather than the CUDA version. Therefore, it is recommended to use pip for installation. If using the official Torch pip command, omit the -f/--index-url parameter so that the installation can use local pip repositories, resulting in faster speeds.
Installing TensorFlow: Reference
Recommended Usage:
- First, check whether the platform image contains the versions of frameworks you need, such as Torch and TensorFlow. If so, use the platform’s built-in image as your first choice.
- If the platform does not have a suitable version of frameworks such as PyTorch or TensorFlow, check which CUDA version your framework requires. For example, PyTorch 1.9.0 requires CUDA 11.1. In that case, you can select a platform image with Miniconda and CUDA 11.1,then install the frameworks you need within that image to avoid the hassle of installing the CUDA Toolkit. (The platform’s built-in CUDA includes .h header files, which is more convenient if you need to recompile code.)
- If none of the above conditions are met, you can choose any Miniconda image and install the relevant frameworks, CUDA, or even other versions of Python yourself after booting up.