Information about availability in Docker containers

[ayushchatur]

Hello, I wonder whether these libraries are part of deep learning containers ? Specifically, Pytorch containers https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-23-07.html

Or whether these are part of CUDA runtime ?

[JanuszL]

Hi @ayushchatur,

Most of the libraries are part of the CUDA toolkit and in effect, they are included in the mentioned container. Some of them are not like nvTIFF. Do you have any specific library in mind?

[ayushchatur]

@JanuszL My ultimate goal is to leverage cusparselt with PyTorch cpp extension to leverage 2:4 sparsity and on tensor cores . Hence I am going a bit methodically. Here is what i plan to do:

1. Try and test (compile and test) cusparselt cuda kernel with the given examples in your documentation.
2. Benchmark with and without sparsity on CUDA and tensor core respectively.
3. Do point 3. for different precision types
4. Call this cuda kernel with PyTorch

Therefore, My first task is to follow the example

For it I need to find the path for

CUSPARSELT_PATH=<cusparseLt_path>

Hence i was wondering where can i find the path for it inside the container

[JanuszL]

@ayushchatur - I think you can install the library you need in the mentioned container and experiment with it there.

[fbusato]

here a good starting point https://pytorch.org/tutorials/prototype/semi_structured_sparse.html