Add rocm checkup for check_binary.sh

[walterddr]

No description provided.

[walterddr]

@jeffdaily @seemethere

[jeffdaily]

The corresponding PR on the pytorch side is pytorch/pytorch#44250. It is failing to run the build_libtorch.sh script because build_amd.py is not run first to hipify the sources.

This log shows a sample failing build:
https://app.circleci.com/pipelines/github/pytorch/pytorch/210639/workflows/72c5caa6-c1dc-4fbc-b092-b95fcac1a3aa/jobs/7276996

Suggested diff for this PR to run the hipify script.

diff --git a/manywheel/build_libtorch.sh b/manywheel/build_libtorch.sh
index ffd56cf..5b392f2 100644
--- a/manywheel/build_libtorch.sh
+++ b/manywheel/build_libtorch.sh
@@ -110,6 +110,11 @@ else
     export _GLIBCXX_USE_CXX11_ABI=0
 fi

+if [[ "$DESIRED_CUDA" == *"rocm"* ]]; then
+    echo "Calling build_amd.py at $(date)"
+    python tools/amd_build/build_amd.py
+fi
+
 echo "Calling setup.py install at $(date)"

 if [[ $LIBTORCH_VARIANT = *"static"* ]]; then