Version of pytorch used in modules doesn't support Nvidia sm90 driver

Support
1

Description

Version of pytorch used in modules doesn’t support Nvidia sm90 driver. The version of pytorch used in vector modules only support Nvidia sm86 or older drivers.

Server Setup Information

  • Weaviate Server Version: 1.26.4
  • Deployment Method: k8s
  • Multi Node? Number of Running Nodes: 3
  • Client Language and Version:
  • Multitenancy?:

Any additional Information

Python error saying version of pytorch doesn’t support sm90 driver.

I see PyTorch 2.0.1 being used has no support for sm_90:

Torch: 2.0.1+cu121
Torch: [‘sm_50’, ‘sm_60’, ‘sm_70’, ‘sm_75’, ‘sm_80’, ‘sm_86’]

I see PyTorch 2.1.0 has sm_90:

Torch: 2.1.0+cu121
Torch: [‘sm_50’, ‘sm_60’, ‘sm_70’, ‘sm_75’, ‘sm_80’, ‘sm_86’, ‘sm_90’]

You can check on this with:

print(torch.version)
print(torch.cuda.get_arch_list())

Problem is modules contain complied code using pytorch 2.0.1, it will need to be re-complied for pytorch 2.1

2

Should be ```
Torch: 2.0.1+cu117

3

Hi, is any work being done to update the weaviate modules to use a new version of Pytorch to support the Nvidia gpu using the sm_90 driver.