Nikita Shulga

As a workaround for pytorch we suppress some of the error coming from asmjit in OSS PyTorch build, see https://github.com/pytorch/pytorch/blob/8a7b4b85592dc7736cc29ede0f5f47b5f52b8965/cmake/Dependencies.cmake#L825 ``` # See https://github.com/pytorch/pytorch/issues/74352 target_compile_options_if_supported(asmjit -Wno-deprecated-copy) target_compile_options_if_supported(asmjit -Wno-unused-but-set-variable) ```

It sounds like prerequisite for this one would be to move existing CI workflow to GitHub action (I'm not sure if CircleCI supports repository dispatch) Also, can you please elaborate...

@atalman referenced document is not available to the general public, so please copy relevant parts of what you expect one needs to accomplished to close this task

I were able to compile and install simple workflow using my personal M1, see https://github.com/malfet/deleteme/runs/6295939723?check_suite_focus=true#step:3:7 Will try to replicate this behavior on our cloud instances

And I can successfully build torchvision natively on M1 (but can't upload it yet), see WIP https://github.com/pytorch/vision/pull/5948 / https://github.com/pytorch/vision/runs/6299765936?check_suite_focus=true

Update: we link statically with cudnn while shipping to conda, because neither https://anaconda.org/anaconda/cudnn nor https://anaconda.org/nvidia/cudnn has versions we depends on. (And CuDNN for 11.1 is much bigger than the one...

@soumith we can prune CuDNN for 11.1, as it results in unusable library, see following comment, which reproduces the problem with CuBLAS, but CuDNN is similarly affected: https://github.com/pytorch/pytorch/issues/53336#issuecomment-791849506

@atalman referenced document is not available to the general public, so please copy relevant parts of what you expect one needs to accomplished to close this task. And we do...

6Mb increase sounds fine (though undesirable, but we can think about it later and 11.7 should help there as well) But I though this code would not build for older...

Should we close the issue then? (Especially if this case is now covered by unittests?) Trying to triage when the issue was fixed(1.13.0.dev20220812 has the bug, while 1.13.0.dev20220813 does not)...