如何在 AWS Sagemaker 中通过 Conda 指定和安装最新版本的 PyTorch？

Question

I'm attempting to use a recent version of PyTorch (1.7.0) in a Conda environment on Sagemaker by specifying the package version in an environment.yml file. However, I'm getting a ResolvePackageNotFound error. Note that I'm just working in a Jupyter notebook with a kernel corresponding to this Conda environment. I'm not using a deep learning image.

Steps to reproduce:

Save the code below as a.yml file, navigate to where this.yml file is stored, and run conda env create environment.yml .

name: test_env
dependencies:
  - numpy
  - pandas
  - pytorch>=1.7.0
  - torchvision
  - scipy
  - ipykernel
  - torchvision

I've tried this on instances of various types (ml.p2.xlarge, ml.p3.2xlarge, and ml.p3.8xlarge) and have gotten the same error each time. I tried it with conda version 4.8.3 and 4.9.2 as well. If I specify pytorch>=1.5.0 , I'm able to create the environment successfully.

Does anyone have any idea why I can't create an environment successfully with more recent versions of PyTorch? Based on this documentation , I wonder if Sagemaker has preinstalled certain versions of PyTorch, and something is going awry when I attempt to use a more recent version.

Answer 1

Sagemaker instances do not always have support for latest packages. Check this link for the list of supported images in Sagemaker instances.