Description
🐛 Describe the bug
PyTorch nightly binary build jobs are failing for ROCm as a side-effect of our AOTriton mem-efficient PR #124885
It seems to be due to the CMake version being older (3.18.2) in the nightly binary build jobs (vs 3.26.4 in the trunk build jobs):
e.g https://github.com/pytorch/pytorch/actions/runs/9434924774/job/25990255607
2024-06-09T11:21:37.0999495Z CMake Error at v2src/CMakeLists.txt:6 (cmake_path):
2024-06-09T11:21:37.1000356Z Unknown CMake command "cmake_path".2024-06-09T11:21:37.1000356Z Unknown CMake command "cmake_path".
2024-06-09T11:21:37.1000356Z Unknown CMake command "cmake_path".2024-06-09T11:21:37.1000356Z Unknown CMake command "cmake_path".
However, instead of bumping the CMake version for binary build jobs, we can get around this issue by applying the same approach as for Pytorch 6082 CI ie. Install AOTriton from tarball (instead of building from source).
Versions
nightly
cc @jeffdaily @sunway513 @pruthvistony @ROCmSupport @dllehr-amd @jataylo @hongxiayang