ARM & PYTORCH 2.6 &CUDA 12.8 #873
Conversation
|
It is working fine on jetson agx orin cuda 12.8 pytorch 2.6.0 https://github.com/dusty-nv/jetson-containers/tree/master/packages/pytorch/torch3d |
There was a problem hiding this comment.
Thank you so much for the PR. I'm gonna add few commits on top just to fix the doc issue and the comments that git removed.
Is there a specific reason to remove the no_python_abi_suffix option? Also the title claim this is for 12.8. How are most of you using 12.8? nvcr or nightly build? It will be difficult for us to have wheels before pytorch make a release. We will try to see how we can ship wheels for ARM but for now it's great to enable at least build from source.
setup.py
Outdated
| CYTHON_MIN_VER = '0.29.37' | ||
| IGNORE_TORCH_VER = os.getenv('IGNORE_TORCH_VER') is not None | ||
|
|
||
| # Module required before installation | ||
| # trying to install it ahead turned out to be too unstable. |
There was a problem hiding this comment.
can you keep this comment?
setup.py
Outdated
| extra_compile_args = {'cxx': ['-O3']} | ||
| define_macros = [] | ||
| include_dirs = [] | ||
| sources = glob.glob('kaolin/csrc/**/*.cpp', recursive=True) | ||
| # FORCE_CUDA is for cross-compilation in docker build |
There was a problem hiding this comment.
Can you keep this comment?
setup.py
Outdated
| packages=find_packages(exclude=('docs', 'tests', 'examples')), | ||
| scripts=get_scripts(), | ||
| include_package_data=True, | ||
| install_requires=get_requirements(), | ||
| zip_safe=False, | ||
| ext_modules=get_extensions(), | ||
| cmdclass={ |
There was a problem hiding this comment.
Please keep the options, is there a specific reason to remove it?
We build our pytorch in jetson-containers (Nvidia) with the official branch. |
|
I pushed my master by mistake and cause it to close the PR, now I can't push anything on you branch anymore 😂 I pushed my modifications on top of your work here: https://github.com/Caenorst/kaolin/tree/arm do you mind reopen the PR with this? |
Ah it’s because seems that in my branch are not the cahnges anymore. You can continue if you want. |
This pull request includes updates to CUDA dispatch macros and improvements to the
setup.pyscript for the Kaolin library. The most important changes include updating the dispatch macros to usescalar_type()and enhancing thesetup.pyscript for better version handling and logging.CUDA dispatch macro updates:
AT_DISPATCH_FLOATING_TYPES_AND_HALFmacros to usescalar_type()instead oftype()inkaolin/csrc/ops/spc/point_utils_cuda.cu(interpolate_trilinear_cuda_impl,coords_to_trilinear_cuda_impl) [1] [2].AT_DISPATCH_FLOATING_TYPES_AND_HALFmacros to usescalar_type()instead oftype()inkaolin/csrc/ops/spc/query_cuda.cu(query_cuda_impl,query_multiscale_cuda_impl,query_cuda_impl_empty) [1] [2] [3].AT_DISPATCH_FLOATING_TYPES_AND_HALFandAT_DISPATCH_INTEGRAL_TYPESmacros to usescalar_type()inkaolin/csrc/render/spc/raytrace_cuda.cu(mark_pack_boundaries_cuda_impl,diff_cuda_impl,sum_reduce_cuda_impl,cumsum_cuda_impl,cumprod_cuda_impl) [1] [2] [3] [4] [5] [6] [7].Setup script improvements:
setup.pyto better handle version constraints and logging for PyTorch and Cython dependencies [1] [2].setup.py[1] [2].