search

k2-fsa / k2

FSA/FST algorithms, differentiable, with PyTorch compatibility.
https://k2-fsa.github.io/k2
Apache License 2.0
1.1k stars 214 forks source link

Provide manylinux wheels #1227

Closed csukuangfj closed 1 year ago

csukuangfj commented 1 year ago

We are providing pre-compiled k2 wheels that are built on Ubuntu 20.04. The consequence is that users with older linux systems, e.g., Ubuntu 18.04, CentOS 7, etc, cannot use them since the wheels link symbols from high versions of libstdc++.so and libc.so.

The fix is to provide pre-compiled k2 wheels that are able to run on many linux systems, especially on older linux systems, e.g., Ubuntu 18.04, CentOS 7, etc.

I just managed to build a manylinux2014 wheel k2-1.24.3.dev20230715+cuda10.2.torch1.10.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl

It can be downloaded from https://huggingface.co/csukuangfj/k2/resolve/main/cuda/k2-1.24.3.dev20230715%2Bcuda10.2.torch1.10.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl

Would be nice if someone can test it on CentOS 7 or some similar older systems. If it works well, I will provide wheels for other combinations of PyTorch and CUDA.

desh2608 commented 1 year ago

@salah-zaiem perhaps this would work for your cluster?

csukuangfj commented 1 year ago

I have updated the doc to include links to manylinux wheels. Please see

Here is a screenshot: Screenshot 2023-07-19 at 15 51 52