[release/2.3] Use preinstalled aotriton if it exists - Githubissues

ROCm / pytorch

Tensors and Dynamic neural networks in Python with strong GPU acceleration

http://pytorch.org

Other

219 stars 51 forks source link

[release/2.3] Use preinstalled aotriton if it exists #1523

Closed jithunnair-amd closed 2 months ago

jithunnair-amd commented 2 months ago

Needs https://github.com/ROCm/pytorch/pull/1506 to be merged first

Add .ci/docker/aotriton_version.txt for details of aotriton tarball
Add logic to cmake/External/aotriton.cmake to use preinstalled aotriton if it exists
Update logic in cmake/External/aotriton.cmake to generate/use shared library instead of static archive
Add logic to .ci/docker/ubuntu-rocm/Dockerfile and .ci/docker/centos-rocm/Dockerfile to install aotriton from tarball
Use https://github.com/ROCm/aotriton/commit/99f540a954e80f446ec6980f108e8c25408f1823 from release/0.4 branch for manylinux2_28 and other fixes

TODO:

[x] Generate gcc-8 based LLVM tarball for triton used by aotriton for release/2.3: https://compute-artifactory.amd.com/artifactory/rocm-generic-local/triton-llvm/aotriton_pytorch_2.3/llvm-49af6502-almalinux8-x64_gcc8.tar.gz
[x] Generate manylinux2_28-based shared library tarball for release/2.3-compatible aotriton: https://github.com/ROCm/aotriton/releases/download/0.4.1b/aotriton-0.4.1b-manylinux_2_28_x86_64-rocm6-shared.tar.bz2

Tested via: http://ml-ci-internal.amd.com:8080/job/pytorch/job/dev/job/manylinux_rocm_wheels_test/156/