[GitHub] [tvm] junrushao1994 commented on pull request #9544: [Target] decouple set_cuda_target_arch with autotvm

[GitBox <gi...@apache.org>]

junrushao1994 commented on pull request #9544:
URL: https://github.com/apache/tvm/pull/9544#issuecomment-974736441


   Some background: @comaniac and I have been discussing (complaining) the intrusive design of autotvm's `tvm_callback_cuda_compile` hook, which makes it hardly possible to specify the cuda arch without introducing dependency to autotvm. For example, even meta schedule has to import autotvm: https://github.com/apache/tvm/blob/main/python/tvm/meta_schedule/builder/local_builder.py#L214-L215
   
   Therefore, we are very grateful that you take the initial stab addressing this problem!
   
   On the other hand, the introduction of `cuda_scope` is probably yet another new interface that would potentially lead to future confusion. I would love to hear about more options if we could possibly make things more concise. Is it possible to detect the "arch" flag in `Target.current()` in the `nvcc.py::compile_cuda`?


-- 
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: commits-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org