Lunderberg commented on PR #16302: URL: https://github.com/apache/tvm/pull/16302#issuecomment-1874258412
@vinx13 I agree that is the better long-term option. However, to avoid it will require dropping support for cmake versions older than 3.18. For these versions, the `NVCC_FLAGS` options are the only way to change the target architecture. The `CMAKE_CUDA_ARCHITECTURES` option was only added in 3.18 (and the `native` option is only available in cmake 3.24+). It looks like we currently set `cmake_minimum_required(VERSION 3.18)` in the TVM `CMakeLists.txt`. The cutlass kernels have older version requirement, but those should be updatable. That said, I think the long-term switch from `NVCC_FLAGS` to `CMAKE_CUDA_ARCHITECTURES` should be in a later PR. Updating the cutlass repos to use `CMAKE_CUDA_ARCHITECTURES` would be a multi-step process, first updating them and later updating TVM to point to the new versions. Because this currently prevents compiling on TVM main, I would like to add this PR first, before the breakage impacts too many others. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
