Skip to content

PyTorch/XLA wheel for release branch build failed to find cudnn8.7.0 #5419

@ManfeiBai

Description

@ManfeiBai

Hi @mateuszlewko and @will-cromar,

I wonder how build process find specific cudnn version like libcudnn8=8.7.0.84-1+cuda11.8, do we installed it already and saved in a bucket or we could only access specific version with limitation?

now we want to build release branch(2.0) with cudnn8.7.0.84-1+cudnn11.8 like https://github.com/pytorch/xla/blob/r2.0/infra/ansible/config/cuda_deps.yaml, but failed with msg: no available installation candidate for libcudnn8=8.7.0.84-1+cuda11.8, we confirmed libcudnn8=8.7.0.84-1+cuda11.8 according to blog due to no access to nvidia yet.

So my question is how to enable building system find and install this specific verison of cudnn(libcudnn8=8.7.0.84-1+cuda11.8)?

Thanks.

cc: @JackCaoG @miladm @will-cromar

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions