Skip to content

Update NCCL building block and bump default version to 2.29.7-1 #536

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
samcmill merged 1 commit into from
Mar 13, 2026
Merged

Update NCCL building block and bump default version to 2.29.7-1 #536

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions docs/building_blocks.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2958,7 +2958,7 @@ source. The default value is False.
use the latest commit on the default branch for the repository.

- __cuda__: Flag to specify the CUDA version of the package to download.
The default is `11.6`. This option is ignored if build is True.
The default is `13.2`. This option is ignored if build is True.

- __environment__: Boolean flag to specify whether the environment
(`CPATH`, `LD_LIBRARY_PATH`, `LIBRARY_PATH`, and `PATH`) should be
Expand All @@ -2980,7 +2980,7 @@ repository. The default is empty, i.e., use the release package
specified by `version`.

- __version__: The version of NCCL to install. The default value is
`2.12.10-1`.
`2.29.7-1`.

__Examples__

Expand Down
77 changes: 57 additions & 20 deletions hpccm/building_blocks/nccl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,8 +31,7 @@
from hpccm.building_blocks.base import bb_base
from hpccm.building_blocks.generic_build import generic_build
from hpccm.building_blocks.packages import packages
from hpccm.common import linux_distro
from hpccm.config import get_cpu_architecture
from hpccm.common import cpu_arch, linux_distro
from hpccm.primitives.comment import comment
from hpccm.primitives.copy import copy
from hpccm.primitives.environment import environment
Expand All @@ -56,7 +55,7 @@ class nccl(bb_base, hpccm.templates.downloader, hpccm.templates.envvars,
use the latest commit on the default branch for the repository.

cuda: Flag to specify the CUDA version of the package to download.
The default is `11.6`. This option is ignored if build is True.
The default is `13.2`. This option is ignored if build is True.

environment: Boolean flag to specify whether the environment
(`CPATH`, `LD_LIBRARY_PATH`, `LIBRARY_PATH`, and `PATH`) should be
Expand All @@ -78,7 +77,7 @@ class nccl(bb_base, hpccm.templates.downloader, hpccm.templates.envvars,
specified by `version`.

version: The version of NCCL to install. The default value is
`2.12.10-1`.
`2.29.7-1`.

# Examples

Expand All @@ -97,39 +96,47 @@ def __init__(self, **kwargs):

super(nccl, self).__init__(**kwargs)

self.__arch_label = '' # Filled in by __cpu_arch
self.__baseurl = kwargs.pop('baseurl', 'https://github.com/NVIDIA/nccl/archive')
self.__build = kwargs.pop('build', False)
self.__build_environment = '' # Filled in by __configure
self.__default_repository = 'https://github.com/NVIDIA/nccl.git'
self.__distro_label = '' # Filled in by __distro
self.__cuda = kwargs.pop('cuda', '11.6')
self.__cuda = kwargs.pop('cuda', '13.2')
self.__make_variables = kwargs.pop('make_variables', {})
self.__ospackages = kwargs.pop('ospackages', [])
self.__prefix = kwargs.pop('prefix', '/usr/local/nccl')
self.__repo_key = '' # Filled in by __repo_key
self.__src_directory = kwargs.pop('src_directory', None)
self.__version = kwargs.pop('version', '2.12.10-1')
self.__version = kwargs.pop('version', '2.29.7-1')
self.__wd = kwargs.get('wd', hpccm.config.g_wd) # working directory

if not self.__build:
# Install prebuild package

# Set the CPU architecture specific parameters
self.__cpu_arch()

# Set the Linux distribution specific parameters
self.__distro()

# Set the repo key
self.__repo()

self += comment('NCCL {}'.format(self.__version))
self += packages(ospackages=self.__ospackages)
self += packages(
apt=['libnccl2={0}+cuda{1}'.format(self.__version,
self.__cuda),
'libnccl-dev={0}+cuda{1}'.format(self.__version,
self.__cuda)],
apt_keys=['https://developer.download.nvidia.com/compute/cuda/repos/{0}/{1}/3bf863cc.pub'.format(self.__distro_label, get_cpu_architecture())],
apt_repositories=['deb [signed-by=/usr/share/keyrings/3bf863cc.gpg] https://developer.download.nvidia.com/compute/cuda/repos/{0}/{1} /'.format(self.__distro_label, get_cpu_architecture())],
apt_keys=['https://developer.download.nvidia.com/compute/cuda/repos/{0}/{1}/{2}'.format(self.__distro_label, self.__arch_label, self.__repo_key)],
apt_repositories=['deb [signed-by=/usr/share/keyrings/{2}] https://developer.download.nvidia.com/compute/cuda/repos/{0}/{1} /'.format(self.__distro_label, self.__arch_label, self.__repo_key.replace('.pub', '.gpg'))],
yum=['libnccl-{0}+cuda{1}'.format(self.__version, self.__cuda),
'libnccl-devel-{0}+cuda{1}'.format(self.__version,
self.__cuda)],
yum_keys=['https://developer.download.nvidia.com/compute/cuda/repos/{0}/{1}/3bf863cc.pub'.format(self.__distro_label, get_cpu_architecture())],
yum_repositories=['https://developer.download.nvidia.com/compute/cuda/repos/{0}/{1}'.format(self.__distro_label, get_cpu_architecture())])
yum_keys=['https://developer.download.nvidia.com/compute/cuda/repos/{0}/{1}/{2}'.format(self.__distro_label, self.__arch_label, self.__repo_key)],
yum_repositories=['https://developer.download.nvidia.com/compute/cuda/repos/{0}/{1}'.format(self.__distro_label, self.__arch_label)])

else:
# Build from source
Expand Down Expand Up @@ -166,6 +173,17 @@ def __init__(self, **kwargs):
self += packages(ospackages=self.__ospackages)
self += self.__bb

def __cpu_arch(self):
"""Based on the CPU architecture, set values accordingly. A user
specified value overrides any defaults."""

if hpccm.config.g_cpu_arch == cpu_arch.AARCH64:
self.__arch_label = 'sbsa'
elif hpccm.config.g_cpu_arch == cpu_arch.X86_64:
self.__arch_label = 'x86_64'
else: # pragma: no cover
raise RuntimeError('Unknown CPU architecture')

def __configure(self):
"""Setup build options based on user parameters"""

Expand All @@ -192,16 +210,20 @@ def __distro(self):
self.__ospackages = ['apt-transport-https', 'ca-certificates',
'gnupg', 'wget']

if hpccm.config.g_linux_version >= Version('18.0'):
self.__distro_label = 'ubuntu1804'
if hpccm.config.g_linux_version >= Version('24.0'):
self.__distro_label = 'ubuntu2404'
elif hpccm.config.g_linux_version >= Version('22.0'):
self.__distro_label = 'ubuntu2204'
else:
self.__distro_label = 'ubuntu1604'
self.__distro_label = 'ubuntu2004'

elif hpccm.config.g_linux_distro == linux_distro.CENTOS:
if hpccm.config.g_linux_version >= Version('8.0'):
self.__distro_label = 'rhel8'
if hpccm.config.g_linux_version >= Version('10.0'):
self.__distro_label = 'rhel10'
elif hpccm.config.g_linux_version >= Version('9.0'):
self.__distro_label = 'rhel9'
else:
self.__distro_label = 'rhel7'
self.__distro_label = 'rhel8'

else: # pragma: no cover
raise RuntimeError('Unknown Linux distribution')
Expand All @@ -225,6 +247,21 @@ def __download(self):
if not self.repository and not self.url:
self.url = '{0}/v{1}.tar.gz'.format(self.__baseurl, self.__version)

def __repo(self):
"""Based on the Linux distribution and CPU architecture, set values
accordingly. A user specified value overrides any defaults.
"""

if self.__distro_label.startswith('ubuntu'):
self.__repo_key = '3bf863cc.pub'
elif self.__distro_label.startswith('rhel'):
if hpccm.config.g_linux_version >= Version('10.0'):
self.__repo_key = 'CDF6BA43.pub'
else:
self.__repo_key = 'D42D0685.pub'
else: # pragma: no cover
raise RuntimeError('Unknown repository')

def runtime(self, _from='0'):
"""Generate the set of instructions to install the runtime specific
components from a build in a previous stage.
Expand All @@ -246,10 +283,10 @@ def runtime(self, _from='0'):
self.rt += packages(
apt=['libnccl2={0}+cuda{1}'.format(self.__version,
self.__cuda)],
apt_keys=['https://developer.download.nvidia.com/compute/cuda/repos/{0}/{1}/3bf863cc.pub'.format(self.__distro_label, get_cpu_architecture())],
apt_repositories=['deb [signed-by=/usr/share/keyrings/3bf863cc.gpg] https://developer.download.nvidia.com/compute/cuda/repos/{0}/{1} /'.format(self.__distro_label, get_cpu_architecture())],
apt_keys=['https://developer.download.nvidia.com/compute/cuda/repos/{0}/{1}/{2}'.format(self.__distro_label, self.__arch_label, self.__repo_key)],
apt_repositories=['deb [signed-by=/usr/share/keyrings/{2}] https://developer.download.nvidia.com/compute/cuda/repos/{0}/{1} /'.format(self.__distro_label, self.__arch_label, self.__repo_key.replace('.pub', '.gpg'))],
yum=['libnccl-{0}+cuda{1}'.format(self.__version, self.__cuda)],
yum_keys=['https://developer.download.nvidia.com/compute/cuda/repos/{0}/{1}/3bf863cc.pub'.format(self.__distro_label, get_cpu_architecture())],
yum_repositories=['https://developer.download.nvidia.com/compute/cuda/repos/{0}/{1}'.format(self.__distro_label, get_cpu_architecture())])
yum_keys=['https://developer.download.nvidia.com/compute/cuda/repos/{0}/{1}/{2}'.format(self.__distro_label, self.__arch_label, self.__repo_key)],
yum_repositories=['https://developer.download.nvidia.com/compute/cuda/repos/{0}/{1}'.format(self.__distro_label, self.__arch_label)])

return str(self.rt)
Loading
Loading