1

I'm trying to configure an Azure Batch pool to use "PyTorch from NVIDIA" VM images. I'm provisioning the pool using the Azure Python SDK and I'm getting the following error:

Reason: The specified imageReference with publisher nvidia offer pytorch_from_nvidia sku pytorch_from_nvidia_19_11_3 is not supported.

When I tried to provision the pool using the Azure Portal I was unable to find Nvidia among the supported publishers (see image below).

Is it possible to use Azure Batch with an Nvidia VM image?

VM publisher dropdown

ifeins
  • 869
  • 1
  • 9
  • 20

1 Answers1

0

I'm afraid there is a misunderstanding of the VM image on you. The Azure VM image does not decide the GPU which you need, it only decides the operating system which you would use, such as Ubuntu, Redhat or Windows Server. The VM size is what would decide the GPU or not for you. Take a look at the GPU optimized virtual machine sizes, which are designed to decide the capacity of GPU you gonna have.

So you don't need to find the VM image which has Nvidia GPU, what you need to do is find an OS that you need and select a size which has enough GPU capacity to meet your requirements.

Charles Xu
  • 29,862
  • 2
  • 22
  • 39
  • I know that the VM size (e.g. NC6) is what indicates whether an instance will have GPU or not. However there are some pre-built VM images such as "PyTorch from NVIDIA" which can be used in conjunction with VM size such as NC6 and they provide built-in Nvidia drivers and in this case an optimized version of PyTorch. So my question remains the same: Can I use Azure Batch with an NVIDIA VM image (such as "PyTorch from NVIDIA")? – ifeins Dec 19 '19 at 11:25
  • @ifeins Of course, yes. As you said if it's the image pre-build by Azure, but it does not show in the batch images. Then you can create a custom image yourself from the pre-build image which you want to use and then use the custom image by selecting the custom image type in batch. – Charles Xu Dec 20 '19 at 01:36
  • This also doesn't work. If you try to create a custom image from an Nvidia image then when you attempt to create a VM from it, you get an error that plan information is required. And when you configure a batch pool you cannot supply plan information (publisher, product, sku) when you use a custom image. – ifeins Dec 25 '19 at 09:29