Skip to content

Changes to enable manylinux2_28 wheels for ROCm#51

Merged
jithunnair-amd merged 40 commits intoROCm:rocm6.3from
jithunnair-amd:rocm_manylinux2_28
Aug 6, 2024
Merged

Changes to enable manylinux2_28 wheels for ROCm#51
jithunnair-amd merged 40 commits intoROCm:rocm6.3from
jithunnair-amd:rocm_manylinux2_28

Conversation

@jithunnair-amd
Copy link
Copy Markdown
Collaborator

@jithunnair-amd jithunnair-amd commented Jul 15, 2024 •
edited
Loading

  • Add manylinux2_28 case to manywheel/build_docker.sh
  • Update steps to install libraries such as magma for ROCm
  • Clean up old-ROCm-version code
  • Use _GLIBCXX_USE_CXX11_ABI=1: 9daa4e4
  • Install numpy<2.0.0 for PyTorch versions<2.4: 6bf4bfd
  • Add conditions for ROCm6.2 and ROCm6.3 in common/install_miopen.sh
  • Upgrade CMake version to 3.28.4 (need minimum of 3.21 to get enable_langugage(HIP) support, but 3.21 has a bug that inserts dynamic linker flags -Wl,no-as-needed in static linker flags causing build failure)
  • Skip unnecessary installation of epel rpm from expired link for CentOS7
  • Install epel-release-7 from cached rpm to enable ius-release to be installed for Almalinux8

TODO:

  • Cherry-pick changes to rocm6.2 branch - Not needed since ROCm6.2.x stays on CentOS7-based flow

Tested via: http://ml-ci-internal.amd.com:8080/job/pytorch/job/dev/job/manylinux_rocm_wheels_test/158/

jithunnair-amd and others added 16 commits July 12, 2024 02:56
@jithunnair-amd
Update ROCm instructions for manylinux
456cf76
@jithunnair-amd
Sync devtoolset version to 11 with CUDA manylinux_2_28 entry
2022877
@jithunnair-amd
Update Dockerfile for ROCm
961e12f
@jithunnair-amd
Add almalinux case
2a4dc44
@jithunnair-amd
Use | to OR different conditions and avoid repetition
674fbc7
@jithunnair-amd
Use devtoolset 11 in alignment with CUDA manylinux config
03f18ff
@pruthvistony @jithunnair-amd
Magma build preinstalls
45dc537 
(cherry picked from commit 9a803c9)
@pruthvistony @jithunnair-amd
Add symlink to python
22dabc1 
(cherry picked from commit 5c4bf9f)
@jithunnair-amd
Add ROCM_PATH
c7007f7
@jithunnair-amd
Use cpu_final to get the devtoolset PATH/LD_LIBRARY_PATH settings and…
dd28584 
… pip cmake install
@jithunnair-amd
Move creation of python link to install_rocm_magma.sh since it's only…
e827147 
… needed for Magma
@jithunnair-amd
Cmake updates
0771a0d
@jithunnair-amd
Add Almalinux case for bundling system libs
92c00a6
@jithunnair-amd
Set _GLIBCXX_USE_CXX11_ABI=1 for ROCm since RHEL8 ROCm packages refer…
9daa4e4 
… to cxx11-abi symbols, so PyTorch build fails if _GLIBCXX_USE_CXX11_ABI=0
@jithunnair-amd
Need * to have a match
a777638
@jithunnair-amd
Install numpy<2.0.0 for older release branches
6bf4bfd
@jithunnair-amd
Copy link
Copy Markdown
Collaborator Author

jithunnair-amd commented Jul 17, 2024

For "Install numpy<2.0.0 for PyTorch versions<2.4: 6bf4bfd":

[2024-07-17T20:18:01.140Z] +++ printf %3d%03d%03d%03d 2 4
[2024-07-17T20:18:01.140Z] ++ [[   1013001000 -ge   2004000000 ]]
[2024-07-17T20:18:01.140Z] ++ retry pip install -q 'numpy<2.0.0'
[2024-07-17T20:18:01.140Z] ++ pip install -q 'numpy<2.0.0'

@jithunnair-amd
Copy link
Copy Markdown
Collaborator Author

jithunnair-amd commented Jul 17, 2024 •
edited
Loading

For "Use _GLIBCXX_USE_CXX11_ABI=1: 9daa4e4"

[2024-07-17T20:18:04.647Z] �[0m++ [[ '' == *\c\x\x\1\1\-\a\b\i* ]]
[2024-07-17T20:18:04.647Z] ++ [[ rocm6.1.0 == *\r\o\c\m* ]]
[2024-07-17T20:18:04.647Z] ++ export _GLIBCXX_USE_CXX11_ABI=1
[2024-07-17T20:18:04.647Z] ++ _GLIBCXX_USE_CXX11_ABI=1
...
[2024-07-17T20:18:13.348Z] cmake3 -GNinja -DBUILD_DEBUG_INFO=0 -DBUILD_ENVIRONMENT=manywheel 3.9 rocm6.1.0 -DBUILD_JNI=OFF -DBUILD_LIBTORCH_CPU_WITH_DEBUG=0 -DBUILD_PYTHON=True -DBUILD_PYTHONLESS= -DBUILD_TEST=True -DCAFFE2_STATIC_LINK_CUDA=1 -DCMAKE_ARGS= -DCMAKE_BUILD_TYPE=Release -DCMAKE_INCLUDE_PATH=/opt/openssl/include:/opt/intel/include: -DCMAKE_INSTALL_PREFIX=/pytorch/torch -DCMAKE_LIBRARY_PATH=/opt/intel/lib:/lib: -DCMAKE_PREFIX_PATH=/opt/python/cp39-cp39/lib/python3.9/site-packages -DGLIBCXX_USE_CXX11_ABI=1 -DINSTALL_TEST=0 -DJAVA_HOME= -DNUMPY_INCLUDE_DIR=/opt/python/cp39-cp39/lib/python3.9/site-packages/numpy/core/include -DOPENSSL_ROOT_DIR=/opt/openssl -DPYTHON_EXECUTABLE=/opt/python/cp39-cp39/bin/python -DPYTHON_INCLUDE_DIR=/opt/python/cp39-cp39/include/python3.9 -DPYTHON_LIBRARY=/opt/_internal/cpython-3.9.19/lib/libpython3.9.a -DTORCH_BUILD_VERSION=1.13.1+rocm6.1.0 -DUSE_FBGEMM=1 -DUSE_GLOO_WITH_OPENSSL=ON -DUSE_GOLD_LINKER=OFF -DUSE_KINETO=1 -DUSE_NCCL=1 -DUSE_NUMPY=True -DUSE_RCCL=1 -DUSE_STATIC_CUDNN=1 -DUSE_STATIC_NCCL=1 /pytorch
...
[2024-07-17T20:18:48.375Z] --   CXX flags                 :  -D_GLIBCXX_USE_CXX11_ABI=1 -Wno-deprecated -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -fopenmp -Wnon-virtual-dtor

@jithunnair-amd jithunnair-amd force-pushed the rocm_manylinux2_28 branch 3 times, most recently from d5321d2 to 4e19c2f Compare July 19, 2024 22:18
@jithunnair-amd
Enable MIOpen build from source for ROCm6.1 to see if it resolves seg…
820f4d1 
…fault issue

Remove existing sqlite3.pc to make cget happy
jithunnair-amd and others added 16 commits August 1, 2024 18:55
@jithunnair-amd jithunnair-amd marked this pull request as ready for review August 5, 2024 08:36
pruthvistony
pruthvistony approved these changes Aug 5, 2024
View reviewed changes
@jithunnair-amd jithunnair-amd merged commit 9eb77a8 into ROCm:rocm6.3 Aug 6, 2024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@pruthvistony pruthvistony pruthvistony approved these changes

@jeffdaily jeffdaily Awaiting requested review from jeffdaily

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@jithunnair-amd @pruthvistony