Closed jeroen-mostert closed 1 week ago
The good news, if you can call it that, is that a clean pull of the onnxruntime repo built with build.sh
(so no special options and not enabling ROCm) fails as well, but with different errors than the above. Unfortunately onnxruntime's own internal build system is a bit of an ogre, so troubleshooting it is a pain. There are no apparent relevant issues in the upstream repo that look relevant.
Update: after adding -Wno-template-id-cdtor
and adding the patch for the _M_Manager
warning a vanilla build fails with the same error, invoking as build.sh --config Release --enable_training --build_wheel --skip_tests --build_shared_lib
. This at least show the issue is not due to the ROCm-specific bits or SDK patches. The build succeeds with --config Debug
, but that of course says little about the validity of the warning.
Well shucks. I was just about to prepare a patch when I noticed @lamikr beat me to the punch and I could have saved myself the trouble. :P Patch 9 for onnxruntime fixes this. I do have a suggestion on how to make it better, but that'll be a separate pull. Closing this.
Attempting to build the
wip/rocm_sdk_builder_612
branch with the Python patch from https://github.com/lamikr/rocm_sdk_builder/pull/70 applied (which is likely unrelated, but mentioned for completeness) on a fully up-to-date Manjaro unstable produces a peculiar error that I've found myself unable to troubleshoot. The header path (/usr/include/absl
) seems to indicate it's using a global include file from the system's GCC, rather than clang or hipcc; I'm not sure whether that's intended or not.And it's been a while since I've written C++ but the error itself is a mystery to me as well: it seems to complain about an uninitialized member being present when invoking a copy constructor, but the instances are being initialized using the default constructor.
This is gcc 14.1.1 20240522.