PYTORCH_BUILD_NUMBER=1, I could build locally version
1.10.0 with PyTorch 1.10.0-rc3 from source. However, the cuda113 torchvision rc requests PyTorch version
1.10.0+cu113, and the PyTorch build fails if I export
PYTORCH_BUILD_VERSION=1.10.0+cu113. I looked into the python script, and yet could only find this one method to specify the PyTorch version. I wondering how PyTorch team built the wheel with the cuda version tag. Similarly for CPU binaries, how did the public binary come with
1.10.0+cpu? Could someone help here? Thanks!
1.10.0+cu113 works without issues in my current linux container now.
Regarding the scripts mentioned by @ptrblck , I think these scripts are run when promoting built binaries, not when building them. The vision script is under conda, and there isn’t one for torch. Thanks for the help though.