Upload new wheels for the newly supported Torch/CUDA versions?

[marcusaureliusfun]

The installation instructions here recommend installing prebuilt wheels, but this cuts off at torch 2.5.1 and cuda 12.6 https://kaolin.readthedocs.io/en/latest/notes/installation.html

Can new torch/cuda versions be supported there? Does this repo maintain those installation instructions, or does someone else?

[Caenorst]

Yes we are making an effort to update the wheels to the latest supported pytorch version at release. We usually don't try to support retroactively post-release as the process would a bit complicated. At the moment if you need to use pytorch 2.7.0 we recommend building from source.

[marcusaureliusfun]

Ah I only need pytorch 2.6.0. This may be asking a bit too much, but do you have a pointer to how to build this from source? I have a dev background but not in the open source world.

step 1: clone repo step 2: ?

[Caenorst]

I recommend having a look at our Dockerfile for in-depth guidance, but in short you can do the following: step 2: install dependencies as in https://github.com/NVIDIAGameWorks/kaolin/blob/master/tools/linux/Dockerfile.install#L73-L79 step 3: python setup.py develop (or python setup.py install)

[IgorAherne]

You can use the following .whl file (disguised as a .png) Right click it, save, then change extension to .whl

Compiled with 2.7 and Cuda 12.8 with the following setup.py:

# some useful environment variables:
#
# TORCH_CUDA_ARCH_LIST
#   specify which CUDA architectures to build for
#
# IGNORE_TORCH_VER
#   ignore version requirements for PyTorch

import os
from setuptools import setup, find_packages
import importlib
from pkg_resources import parse_version
import warnings
import logging
import glob
import sys
import subprocess  # Added import

# Define version constraints
TORCH_MIN_VER = '1.6.0'
TORCH_MAX_VER = '2.6.0'  # Updated to support newer PyTorch versions
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
# Check for PyTorch
torch_spec = importlib.util.find_spec("torch")
if torch_spec is None:
    raise ImportError(
        f"Kaolin requires PyTorch >= {TORCH_MIN_VER}, <= {TORCH_MAX_VER}. "
        "Please install it before proceeding."
    )
else:
    import torch
    torch_ver = parse_version(parse_version(torch.__version__).base_version)
    if torch_ver < parse_version(TORCH_MIN_VER) or torch_ver > parse_version(TORCH_MAX_VER):
        if IGNORE_TORCH_VER:
            warnings.warn(
                f"Kaolin is compatible with PyTorch >= {TORCH_MIN_VER}, <= {TORCH_MAX_VER}, "
                f"but found version {torch.__version__}. Continuing as IGNORE_TORCH_VER is set."
            )
        else:
            raise ImportError(
                f"Kaolin requires PyTorch >= {TORCH_MIN_VER}, <= {TORCH_MAX_VER}, "
                f"but found version {torch.__version__}. "
                "Set IGNORE_TORCH_VER=1 to proceed with this version."
            )

# Check for Cython
cython_spec = importlib.util.find_spec("cython")
if cython_spec is None:
    raise ImportError(
        f"Kaolin requires Cython >= {CYTHON_MIN_VER}. Please install it before proceeding."
    )
else:
    import Cython
    cython_ver = parse_version(Cython.__version__)
    if cython_ver < parse_version(CYTHON_MIN_VER):
        warnings.warn(
            f"Kaolin requires Cython >= {CYTHON_MIN_VER}, "
            f"but found version {Cython.__version__}. This may cause compatibility issues."
        )

# Check for NumPy
numpy_spec = importlib.util.find_spec("numpy")
if numpy_spec is None:
    raise ImportError(
        "Kaolin requires NumPy. Please install it before proceeding."
    )

import numpy
from torch.utils.cpp_extension import BuildExtension, CppExtension, CUDAExtension, CUDA_HOME

# Setup logging and working directory
cwd = os.path.dirname(os.path.abspath(__file__))
logger = logging.getLogger()
logging.basicConfig(format='%(levelname)s - %(message)s')

def get_cuda_bare_metal_version(cuda_dir):
    """Get CUDA version from nvcc."""
    raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True)
    output = raw_output.split()
    release_idx = output.index("release") + 1
    release = output[release_idx].split(".")
    return raw_output, release[0], release[1][0]

# Handle CUDA availability
if not torch.cuda.is_available() and os.getenv('FORCE_CUDA', '0') == '1':
    logging.warning(
        "Torch did not find available GPUs. Assuming cross-compilation.\n"
        "Default architectures: Pascal (6.0, 6.1, 6.2), Volta (7.0), Turing (7.5),\n"
        "Ampere (8.0) if CUDA >= 11.0, Hopper (9.0) if CUDA >= 11.8, Blackwell (12.0) if CUDA >= 12.8\n"
        "Set TORCH_CUDA_ARCH_LIST for specific architectures."
    )
    if os.getenv("TORCH_CUDA_ARCH_LIST") is None:
        _, major, minor = get_cuda_bare_metal_version(CUDA_HOME)
        major, minor = int(major), int(minor)
        if major == 11:
            if minor == 0:
                os.environ["TORCH_CUDA_ARCH_LIST"] = "5.2;6.0;6.1;6.2;7.0;7.5;8.0"
            elif minor < 8:
                os.environ["TORCH_CUDA_ARCH_LIST"] = "5.2;6.0;6.1;6.2;7.0;7.5;8.0;8.6"
            else:
                os.environ["TORCH_CUDA_ARCH_LIST"] = "5.2;6.0;6.1;6.2;7.0;7.5;8.0;8.6;8.9;9.0"
        elif major == 12:
            if minor <= 6:
                os.environ["TORCH_CUDA_ARCH_LIST"] = "5.2;6.0;6.1;6.2;7.0;7.5;8.0;8.6;9.0"
            else:
                os.environ["TORCH_CUDA_ARCH_LIST"] = "5.2;6.0;6.1;6.2;7.0;7.5;8.0;8.6;9.0;12.0"
        else:
            os.environ["TORCH_CUDA_ARCH_LIST"] = "6.0;6.1;6.2;7.0;7.5"
        print(f'TORCH_CUDA_ARCH_LIST: {os.environ["TORCH_CUDA_ARCH_LIST"]}')
elif not torch.cuda.is_available():
    logging.warning(
        "No GPUs found. Installing with CPU support only. "
        "Set FORCE_CUDA=1 for GPU cross-compilation."
    )

# Package metadata
PACKAGE_NAME = 'kaolin'
DESCRIPTION = 'Kaolin: A PyTorch library for accelerating 3D deep learning research'
URL = 'https://github.com/NVIDIAGameWorks/kaolin'
AUTHOR = 'NVIDIA'
LICENSE = 'Apache License 2.0'
LONG_DESCRIPTION = """
Kaolin is a PyTorch library aiming to accelerate 3D deep learning research. Kaolin provides efficient implementations
of differentiable 3D modules for use in deep learning systems. With functionality to load and preprocess several popular
3D datasets, and native functions to manipulate meshes, pointclouds, signed distance functions, and voxel grids, Kaolin
mitigates the need to write wasteful boilerplate code. Kaolin packages together several differentiable graphics modules
including rendering, lighting, shading, and view warping. Kaolin also supports an array of loss functions and evaluation
metrics for seamless evaluation and provides visualization functionality to render the 3D results. Importantly, we curate
a comprehensive model zoo comprising many state-of-the-art 3D deep learning architectures, to serve as a starting point
for future research endeavours.
"""

# Version handling
version_txt = os.path.join(cwd, 'version.txt')
with open(version_txt) as f:
    version = f.readline().strip()

def write_version_file():
    """Write version to kaolin/version.py."""
    version_path = os.path.join(cwd, 'kaolin', 'version.py')
    with open(version_path, 'w') as f:
        f.write(f"__version__ = '{version}'\n")

write_version_file()

def get_requirements():
    """Read runtime dependencies from requirements files."""
    requirements = []
    with open(os.path.join(cwd, 'tools', 'viz_requirements.txt'), 'r') as f:
        requirements.extend(line.strip() for line in f)
    with open(os.path.join(cwd, 'tools', 'requirements.txt'), 'r') as f:
        requirements.extend(line.strip() for line in f)
    return requirements

def get_scripts():
    """Return list of scripts to install."""
    return ['kaolin/experimental/dash3d/kaolin-dash3d']

def get_extensions():
    """Define C++ and CUDA extensions."""
    extra_compile_args = {} # Initialize
    if sys.platform == 'win32':
        # MSVC specific optimization flags for C++
        extra_compile_args['cxx'] = ['/O2', '/std:c++17'] # Add /std:c++17 as seen in your log
    else:
        # GCC/Clang flags for C++
        extra_compile_args['cxx'] = ['-O3', '-std=c++17'] # Add -std=c++17 as seen in your log

    define_macros = []
    include_dirs = []
    sources = glob.glob('kaolin/csrc/**/*.cpp', recursive=True)
    # FORCE_CUDA is for cross-compilation in docker build
    is_cuda = torch.cuda.is_available() or os.getenv('FORCE_CUDA', '0') == '1'

    if is_cuda:
        define_macros += [("WITH_CUDA", None), ("THRUST_IGNORE_CUB_VERSION_CHECK", None)]
        sources += glob.glob('kaolin/csrc/**/*.cu', recursive=True)
        extension = CUDAExtension
        extra_compile_args['nvcc'] = [
            '--use_fast_math',
            '-O3',
            '-DWITH_CUDA',
            '--allow-unsupported-compiler',  #else my VisualStudio 2022 isn't recognized.
        ]
        include_dirs = get_include_dirs()
    else:
        extension = CppExtension

    extensions = [
        extension(
            name='kaolin._C',
            sources=sources,
            define_macros=define_macros,
            extra_compile_args=extra_compile_args,
            include_dirs=include_dirs
        )
    ]
    
    # Replace cudart with cudart_static
    for ext in extensions:
        ext.libraries = ['cudart_static' if x == 'cudart' else x for x in ext.libraries]

    # Cython extensions
    use_cython = True
    ext_suffix = '.pyx' if use_cython else '.cpp'
    cython_extensions = [
        CppExtension(
            'kaolin.ops.conversions.mise',
            sources=[f'kaolin/cython/ops/conversions/mise{ext_suffix}'],
        ),
    ]
    
    if use_cython:
        from Cython.Build import cythonize
        from Cython.Compiler import Options
        compiler_directives = Options.get_directive_defaults()
        compiler_directives["emit_code_comments"] = False
        cython_extensions = cythonize(
            cython_extensions,
            language='c++',
            compiler_directives=compiler_directives
        )
    
    return extensions + cython_extensions

def get_include_dirs():
    """Get include directories for CUDA builds."""
    include_dirs = []
    if torch.cuda.is_available() or os.getenv('FORCE_CUDA', '0') == '1':
        _, major, _ = get_cuda_bare_metal_version(CUDA_HOME)
        if "CUB_HOME" in os.environ:
            logging.warning(f"Including CUB_HOME: {os.environ['CUB_HOME']}")
            include_dirs.append(os.environ["CUB_HOME"])
        elif int(major) < 11:
            logging.warning(f"Including default CUB: {os.path.join(cwd, 'third_party/cub')}")
            include_dirs.append(os.path.join(cwd, 'third_party/cub'))
    return include_dirs

if __name__ == '__main__':
    setup(
        name=PACKAGE_NAME,
        version=version,
        author=AUTHOR,
        description=DESCRIPTION,
        url=URL,
        long_description=LONG_DESCRIPTION,
        license=LICENSE,
        python_requires='~=3.7',
        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={'build_ext': BuildExtension.with_options(no_python_abi_suffix=True)}
    )

[yangmingweiwtu]

您可以使用以下.whl文件（伪装为 .png） 右键单击它，保存，然后将扩展名更改为 .whl

使用 2.7 和 Cuda 12.8 进行编译，setup.py 如下：

# some useful environment variables:
#
# TORCH_CUDA_ARCH_LIST
#   specify which CUDA architectures to build for
#
# IGNORE_TORCH_VER
#   ignore version requirements for PyTorch

import os
from setuptools import setup, find_packages
import importlib
from pkg_resources import parse_version
import warnings
import logging
import glob
import sys
import subprocess  # Added import

# Define version constraints
TORCH_MIN_VER = '1.6.0'
TORCH_MAX_VER = '2.6.0'  # Updated to support newer PyTorch versions
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
# Check for PyTorch
torch_spec = importlib.util.find_spec("torch")
if torch_spec is None:
    raise ImportError(
        f"Kaolin requires PyTorch >= {TORCH_MIN_VER}, <= {TORCH_MAX_VER}. "
        "Please install it before proceeding."
    )
else:
    import torch
    torch_ver = parse_version(parse_version(torch.__version__).base_version)
    if torch_ver < parse_version(TORCH_MIN_VER) or torch_ver > parse_version(TORCH_MAX_VER):
        if IGNORE_TORCH_VER:
            warnings.warn(
                f"Kaolin is compatible with PyTorch >= {TORCH_MIN_VER}, <= {TORCH_MAX_VER}, "
                f"but found version {torch.__version__}. Continuing as IGNORE_TORCH_VER is set."
            )
        else:
            raise ImportError(
                f"Kaolin requires PyTorch >= {TORCH_MIN_VER}, <= {TORCH_MAX_VER}, "
                f"but found version {torch.__version__}. "
                "Set IGNORE_TORCH_VER=1 to proceed with this version."
            )

# Check for Cython
cython_spec = importlib.util.find_spec("cython")
if cython_spec is None:
    raise ImportError(
        f"Kaolin requires Cython >= {CYTHON_MIN_VER}. Please install it before proceeding."
    )
else:
    import Cython
    cython_ver = parse_version(Cython.__version__)
    if cython_ver < parse_version(CYTHON_MIN_VER):
        warnings.warn(
            f"Kaolin requires Cython >= {CYTHON_MIN_VER}, "
            f"but found version {Cython.__version__}. This may cause compatibility issues."
        )

# Check for NumPy
numpy_spec = importlib.util.find_spec("numpy")
if numpy_spec is None:
    raise ImportError(
        "Kaolin requires NumPy. Please install it before proceeding."
    )

import numpy
from torch.utils.cpp_extension import BuildExtension, CppExtension, CUDAExtension, CUDA_HOME

# Setup logging and working directory
cwd = os.path.dirname(os.path.abspath(__file__))
logger = logging.getLogger()
logging.basicConfig(format='%(levelname)s - %(message)s')

def get_cuda_bare_metal_version(cuda_dir):
    """Get CUDA version from nvcc."""
    raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True)
    output = raw_output.split()
    release_idx = output.index("release") + 1
    release = output[release_idx].split(".")
    return raw_output, release[0], release[1][0]

# Handle CUDA availability
if not torch.cuda.is_available() and os.getenv('FORCE_CUDA', '0') == '1':
    logging.warning(
        "Torch did not find available GPUs. Assuming cross-compilation.\n"
        "Default architectures: Pascal (6.0, 6.1, 6.2), Volta (7.0), Turing (7.5),\n"
        "Ampere (8.0) if CUDA >= 11.0, Hopper (9.0) if CUDA >= 11.8, Blackwell (12.0) if CUDA >= 12.8\n"
        "Set TORCH_CUDA_ARCH_LIST for specific architectures."
    )
    if os.getenv("TORCH_CUDA_ARCH_LIST") is None:
        _, major, minor = get_cuda_bare_metal_version(CUDA_HOME)
        major, minor = int(major), int(minor)
        if major == 11:
            if minor == 0:
                os.environ["TORCH_CUDA_ARCH_LIST"] = "5.2;6.0;6.1;6.2;7.0;7.5;8.0"
            elif minor < 8:
                os.environ["TORCH_CUDA_ARCH_LIST"] = "5.2;6.0;6.1;6.2;7.0;7.5;8.0;8.6"
            else:
                os.environ["TORCH_CUDA_ARCH_LIST"] = "5.2;6.0;6.1;6.2;7.0;7.5;8.0;8.6;8.9;9.0"
        elif major == 12:
            if minor <= 6:
                os.environ["TORCH_CUDA_ARCH_LIST"] = "5.2;6.0;6.1;6.2;7.0;7.5;8.0;8.6;9.0"
            else:
                os.environ["TORCH_CUDA_ARCH_LIST"] = "5.2;6.0;6.1;6.2;7.0;7.5;8.0;8.6;9.0;12.0"
        else:
            os.environ["TORCH_CUDA_ARCH_LIST"] = "6.0;6.1;6.2;7.0;7.5"
        print(f'TORCH_CUDA_ARCH_LIST: {os.environ["TORCH_CUDA_ARCH_LIST"]}')
elif not torch.cuda.is_available():
    logging.warning(
        "No GPUs found. Installing with CPU support only. "
        "Set FORCE_CUDA=1 for GPU cross-compilation."
    )

# Package metadata
PACKAGE_NAME = 'kaolin'
DESCRIPTION = 'Kaolin: A PyTorch library for accelerating 3D deep learning research'
URL = 'https://github.com/NVIDIAGameWorks/kaolin'
AUTHOR = 'NVIDIA'
LICENSE = 'Apache License 2.0'
LONG_DESCRIPTION = """
Kaolin is a PyTorch library aiming to accelerate 3D deep learning research. Kaolin provides efficient implementations
of differentiable 3D modules for use in deep learning systems. With functionality to load and preprocess several popular
3D datasets, and native functions to manipulate meshes, pointclouds, signed distance functions, and voxel grids, Kaolin
mitigates the need to write wasteful boilerplate code. Kaolin packages together several differentiable graphics modules
including rendering, lighting, shading, and view warping. Kaolin also supports an array of loss functions and evaluation
metrics for seamless evaluation and provides visualization functionality to render the 3D results. Importantly, we curate
a comprehensive model zoo comprising many state-of-the-art 3D deep learning architectures, to serve as a starting point
for future research endeavours.
"""

# Version handling
version_txt = os.path.join(cwd, 'version.txt')
with open(version_txt) as f:
    version = f.readline().strip()

def write_version_file():
    """Write version to kaolin/version.py."""
    version_path = os.path.join(cwd, 'kaolin', 'version.py')
    with open(version_path, 'w') as f:
        f.write(f"__version__ = '{version}'\n")

write_version_file()

def get_requirements():
    """Read runtime dependencies from requirements files."""
    requirements = []
    with open(os.path.join(cwd, 'tools', 'viz_requirements.txt'), 'r') as f:
        requirements.extend(line.strip() for line in f)
    with open(os.path.join(cwd, 'tools', 'requirements.txt'), 'r') as f:
        requirements.extend(line.strip() for line in f)
    return requirements

def get_scripts():
    """Return list of scripts to install."""
    return ['kaolin/experimental/dash3d/kaolin-dash3d']

def get_extensions():
    """Define C++ and CUDA extensions."""
    extra_compile_args = {} # Initialize
    if sys.platform == 'win32':
        # MSVC specific optimization flags for C++
        extra_compile_args['cxx'] = ['/O2', '/std:c++17'] # Add /std:c++17 as seen in your log
    else:
        # GCC/Clang flags for C++
        extra_compile_args['cxx'] = ['-O3', '-std=c++17'] # Add -std=c++17 as seen in your log

    define_macros = []
    include_dirs = []
    sources = glob.glob('kaolin/csrc/**/*.cpp', recursive=True)
    # FORCE_CUDA is for cross-compilation in docker build
    is_cuda = torch.cuda.is_available() or os.getenv('FORCE_CUDA', '0') == '1'

    if is_cuda:
        define_macros += [("WITH_CUDA", None), ("THRUST_IGNORE_CUB_VERSION_CHECK", None)]
        sources += glob.glob('kaolin/csrc/**/*.cu', recursive=True)
        extension = CUDAExtension
        extra_compile_args['nvcc'] = [
            '--use_fast_math',
            '-O3',
            '-DWITH_CUDA',
            '--allow-unsupported-compiler',  #else my VisualStudio 2022 isn't recognized.
        ]
        include_dirs = get_include_dirs()
    else:
        extension = CppExtension

    extensions = [
        extension(
            name='kaolin._C',
            sources=sources,
            define_macros=define_macros,
            extra_compile_args=extra_compile_args,
            include_dirs=include_dirs
        )
    ]
    
    # Replace cudart with cudart_static
    for ext in extensions:
        ext.libraries = ['cudart_static' if x == 'cudart' else x for x in ext.libraries]

    # Cython extensions
    use_cython = True
    ext_suffix = '.pyx' if use_cython else '.cpp'
    cython_extensions = [
        CppExtension(
            'kaolin.ops.conversions.mise',
            sources=[f'kaolin/cython/ops/conversions/mise{ext_suffix}'],
        ),
    ]
    
    if use_cython:
        from Cython.Build import cythonize
        from Cython.Compiler import Options
        compiler_directives = Options.get_directive_defaults()
        compiler_directives["emit_code_comments"] = False
        cython_extensions = cythonize(
            cython_extensions,
            language='c++',
            compiler_directives=compiler_directives
        )
    
    return extensions + cython_extensions

def get_include_dirs():
    """Get include directories for CUDA builds."""
    include_dirs = []
    if torch.cuda.is_available() or os.getenv('FORCE_CUDA', '0') == '1':
        _, major, _ = get_cuda_bare_metal_version(CUDA_HOME)
        if "CUB_HOME" in os.environ:
            logging.warning(f"Including CUB_HOME: {os.environ['CUB_HOME']}")
            include_dirs.append(os.environ["CUB_HOME"])
        elif int(major) < 11:
            logging.warning(f"Including default CUB: {os.path.join(cwd, 'third_party/cub')}")
            include_dirs.append(os.path.join(cwd, 'third_party/cub'))
    return include_dirs

if __name__ == '__main__':
    setup(
        name=PACKAGE_NAME,
        version=version,
        author=AUTHOR,
        description=DESCRIPTION,
        url=URL,
        long_description=LONG_DESCRIPTION,
        license=LICENSE,
        python_requires='~=3.7',
        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={'build_ext': BuildExtension.with_options(no_python_abi_suffix=True)}
    )

Hello, can you explain in more detail what is disguised as a png file?

[Caenorst]

Hello, we now support the latest pytorch cuda version with kaolin v0.18.0. Closing as it is now addressed.