{ai}[foss/2024a] PyTorch v2.9.1 w/ CUDA 12.6.0

easybuild/easyconfigs/c/cuDNN/cuDNN-9.10.2.21-CUDA-12.6.0.eb

@@ -0,0 +1,34 @@
+name = 'cuDNN'
+version = '9.10.2.21'
+versionsuffix = '-CUDA-%(cudaver)s'
+homepage = 'https://developer.nvidia.com/cudnn'
+description = """The NVIDIA CUDA Deep Neural Network library (cuDNN) is
+a GPU-accelerated library of primitives for deep neural networks."""
+
+toolchain = SYSTEM
+
+source_urls = [
+    'https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/linux-%(cudnnarch)s/'
+]
+# note: cuDNN is tied to specific to CUDA versions,
+# see also https://docs.nvidia.com/deeplearning/cudnn/support-matrix/index.html#cudnn-cuda-hardware-versions
+sources = ['%(namelower)s-linux-%(cudnnarch)s-%(version)s_cuda%(cudamajver)s-archive.tar.xz']
+checksums = [{
+    '%(namelower)s-linux-sbsa-%(version)s_cuda%(cudamajver)s-archive.tar.xz':
+        '4d57dceba3be27a68b078ce8630525bf40ab7f1b546eb45d0b363c3eeb55f8fa',
+    '%(namelower)s-linux-x86_64-%(version)s_cuda%(cudamajver)s-archive.tar.xz':
+        'd0defcbc4c6dad711ff4cb66d254036a300c9071b07c7b64199aacab534313c1',
+}]
+
+dependencies = [('CUDA', '12.6.0')]
+
+sanity_check_paths = {
+    'files': [
+        'include/cudnn.h', 'lib64/libcudnn_adv_static.a', 'lib64/libcudnn_cnn_static.a',
+        'lib64/libcudnn_engines_precompiled_static.a', 'lib64/libcudnn_engines_runtime_compiled_static.a',
+        'lib64/libcudnn_graph_static.a', 'lib64/libcudnn_heuristic_static.a', 'lib64/libcudnn_ops_static.a',
+    ],
+    'dirs': ['include', 'lib64'],
+}
+
+moduleclass = 'numlib'

easybuild/easyconfigs/n/NCCL/NCCL-2.27.5-GCCcore-13.3.0-CUDA-12.6.0.eb

@@ -0,0 +1,26 @@
+name = 'NCCL'
+version = '2.27.5'
+versionsuffix = '-CUDA-%(cudaver)s'
+
+homepage = 'https://developer.nvidia.com/nccl'
+description = """The NVIDIA Collective Communications Library (NCCL) implements multi-GPU and multi-node collective
+communication primitives that are performance optimized for NVIDIA GPUs."""
+
+toolchain = {'name': 'GCCcore', 'version': '13.3.0'}
+
+github_account = 'NVIDIA'
+source_urls = [GITHUB_SOURCE]
+sources = ['v%(version)s-1.tar.gz']
+checksums = ['e8a8972fc7f7517703510ef23608d41f6484db5331fca37827b4af3f66995344']
+
+builddependencies = [('binutils', '2.42')]
+
+dependencies = [
+    ('CUDA', '12.6.0', '', SYSTEM),
+    ('UCX-CUDA', '1.16.0', versionsuffix),
+]
+
+# default CUDA compute capabilities to use (override via --cuda-compute-capabilities)
+cuda_compute_capabilities = ['5.0', '6.0', '7.0', '7.5', '8.0', '8.6', '9.0']
+
+moduleclass = 'lib'

easybuild/easyconfigs/p/PyTorch/PyTorch-2.9.0_readd-support-for-nvidia-cutlass-python-package.patch

@@ -0,0 +1,124 @@
+Allow use of the NVIDIA CUTLASS Python package if installed.
+See https://github.com/pytorch/pytorch/pull/160180
+
+Author: Alexander Grund (TU Dresden)
+
+diff -ur a/torch/_inductor/codecache.py b/torch/_inductor/codecache.py
+--- a/torch/_inductor/codecache.py	2025-10-15 19:15:08.000000000 +0200
++++ b/torch/_inductor/codecache.py	2025-10-24 18:07:49.519431015 +0200
+@@ -3628,13 +3628,15 @@
+     return "nvcc"
+
+
+-def _cutlass_path() -> str:
++def _cutlass_path() -> Optional[str]:
+     if config.is_fbcode():
+         from libfb.py import parutil
+
+         return parutil.get_dir_path("cutlass-4-headers")
+     else:
+-        return config.cuda.cutlass_dir
++        from torch._inductor.codegen.cuda.cutlass_utils import try_import_cutlass
++
++        return config.cuda.cutlass_dir if try_import_cutlass() else None
+
+
+ def _cutlass_paths() -> list[str]:
+@@ -3649,6 +3651,8 @@
+ def _clone_cutlass_paths(build_root: str) -> list[str]:
+     paths = _cutlass_paths()
+     cutlass_root = _cutlass_path()
++    if cutlass_root is None:
++        return []
+     for path in _cutlass_paths():
+         old_path = os.path.join(cutlass_root, path)
+         new_path = os.path.join(build_root, path)
+@@ -3657,10 +3661,12 @@
+
+
+ def _cutlass_include_paths() -> list[str]:
+-    cutlass_path = _cutlass_path()
++    cutlass_root = _cutlass_path()
++    if cutlass_root is None:
++        return []
+     return [
+         # Use realpath to get canonical absolute paths, in order not to mess up cache keys
+-        os.path.realpath(os.path.join(cutlass_path, path))
++        os.path.realpath(os.path.join(cutlass_root, path))
+         for path in _cutlass_paths()
+     ]
+
+diff -ur a/torch/_inductor/codegen/cuda/cutlass_utils.py b/torch/_inductor/codegen/cuda/cutlass_utils.py
+--- a/torch/_inductor/codegen/cuda/cutlass_utils.py	2025-10-15 19:15:08.000000000 +0200
++++ b/torch/_inductor/codegen/cuda/cutlass_utils.py	2025-10-24 18:07:49.520431003 +0200
+@@ -1,6 +1,7 @@
+ # mypy: allow-untyped-defs
+ import atexit
+ import functools
++import importlib.metadata
+ import logging
+ import os
+ import shutil
+@@ -15,6 +16,7 @@
+ import torch
+ from torch._inductor.runtime.runtime_utils import dynamo_timed
+ from torch._inductor.utils import clear_on_fresh_cache
++from torch._vendor.packaging.version import Version
+ from torch.utils._ordered_set import OrderedSet
+
+ from ... import config
+@@ -73,7 +75,9 @@
+     """
+     We want to support three ways of passing in CUTLASS:
+     1. fbcode, handled by the internal build system.
+-    2. User specifies cutlass_dir. The default is ../third_party/cutlass/,
++    2. pip install nvidia-cutlass, which provides the cutlass_library package
++       and the header files in the cutlass_library/source directory.
++    3. User specifies cutlass_dir. The default is ../third_party/cutlass/,
+        which is the directory when developers build from source.
+     """
+     if config.is_fbcode():
+@@ -89,6 +93,34 @@
+
+         return True
+
++    try:
++        cutlass_version = Version(importlib.metadata.version("cutlass"))
++        if cutlass_version < Version("3.7"):
++            log.warning("CUTLASS version < 3.7 is not recommended.")
++
++        import cutlass_library  # type: ignore[import-not-found]  # noqa: F811
++
++        log.debug(
++            "Found cutlass_library in python search path, overriding config.cuda.cutlass_dir"
++        )
++        cutlass_library_dir = os.path.dirname(cutlass_library.__file__)
++        assert os.path.isdir(cutlass_library_dir), (
++            f"{cutlass_library_dir} is not a directory"
++        )
++        config.cuda.cutlass_dir = os.path.abspath(
++            os.path.join(
++                cutlass_library_dir,
++                "source",
++            )
++        )
++
++        return True
++    except (ModuleNotFoundError, importlib.metadata.PackageNotFoundError):
++        log.debug(
++            "cutlass_library not found in sys.path, trying to import from config.cuda.cutlass_dir",
++            exc_info=True,
++        )
++
+     # Copy CUTLASS python scripts to a temp dir and add the temp dir to Python search path.
+     # This is a temporary hack to avoid CUTLASS module naming conflicts.
+     # TODO(ipiszy): remove this hack when CUTLASS solves Python scripts packaging structure issues.
+@@ -156,7 +188,7 @@
+                 )
+
+         try:
+-            import cutlass  # noqa: F401, F811
++            import cutlass  # noqa: F401
+             import cutlass_library.generator  # noqa: F401
+             import cutlass_library.library  # noqa: F401
+             import cutlass_library.manifest  # noqa: F401