Closed flip111 closed 4 weeks ago
lamikr
commented
1 month ago Did you found a solution to this one yet? If not, I will try tomorrow install the latest ubuntu and use it instead of 2023.10 version I have used for build testing from Ubuntu. It could be some msgpackc- devel dep file that needs to be installed.
flip111
commented
1 month ago It might be the missing package libmsgpuck-dev at first i couldn't find it because of puck instead of pack, odd but it's like that. I will install it then recompile again.
Question: is recompiling from the point it stopped last time supported or do i need to wipe the build dir?
EDIT: already quickly found out it's not libmsgpuck-dev and libmsgpack-cxx-dev was already installed
flip111
commented
1 month ago I think it's due to renaming of the library
https://packages.ubuntu.com/mantic/libmsgpack-cxx-dev https://packages.ubuntu.com/noble/libmsgpack-cxx-dev
lamikr
commented
1 month ago Just checked that on my ubuntu 39, I have libmsgpack-cxx-dev installed. I will setup Ubuntu 40 to try this out.
# apt list --installed | grep msgpack
WARNING: apt does not have a stable CLI interface. Use with caution in scripts.
libmsgpack-cxx-dev/mantic,now 4.1.3-1 amd64 [installed,automatic]
libmsgpack-dev/mantic,now 4.0.0-3 amd64 [installed]
libmsgpackc2/mantic,now 4.0.0-3 amd64 [installed,automatic]@flip111 Usually the build can just continue from where it failed, unless there are fixes to the older packages. Inside each build folder there are .result_* files and you can remove those if you want to force certain steps to be re-run for some package. (Like "rm -rf builddir/007_02_rocminfo/.result_install" for example)
Once some step has been succesfully done, it creates that file. Python apps are little more problematic as at the moment, they are modifying also the content in source directory. (In pytorch aten and hipify for example modify some cuda-files to amd friendly format in pytorch source directory and at the moment I do not have any clean solution in babs.sh itself for resetting those cleanly)
TheRikke
commented
1 month ago I could resolve the issue by creating a link to the new library name:
sudo ln -s /usr/lib/x86_64-linux-gnu/libmsgpackc.so.2.0.0 /usr/lib/x86_64-linux-gnu/libmsgpackc-cxx.so
eLBart0-DTG
commented
1 month ago I could resolve the issue by creating a link to the new library name:
sudo ln -s /usr/lib/x86_64-linux-gnu/libmsgpackc.so.2.0.0 /usr/lib/x86_64-linux-gnu/libmsgpackc-cxx.so
I can confirm this works. Of course it's not the best solution, but you can definitely try it, if you don't want to wait for an official work-around.
lamikr
commented
1 month ago Thanks for the findings, I have not had time yet to work on Ubuntu 24.04. WIll plan to do that once last 2 Fedora 40 problems are solved. There are discussion from same kind of thing on
https://github.com/lamikr/rocm_sdk_builder/pull/27
Maybe the project CMakefiles could be changed to optionally try to search both the msgpack and msgpack-cxx options for AMDMIGraphX?
lamikr
commented
4 weeks ago Pushed fix does not require creating symlink for testing. Tested already on Ubuntu 24.04, now doing just other test build on other distro to check it does not break anything
OS: Ubuntu 24.04 GPU: Sapphire Radeon RX 7900 XTX Nitro+ Vapor-X Aktiv PCIe 4.0 x16