Merge pull request #23606 from Flamefire/20250808105302_new_pr_nvidia-cutlass3800

{lib}[gfbf/2024a] nvidia-cutlass v3.8.0.0 w/ CUDA 12.6.0

Author: boegel

easybuild/easyconfigs/n/nvidia-cutlass/nvidia-cutlass-3.8.0.0-gfbf-2024a-CUDA-12.6.0.eb

@@ -0,0 +1,66 @@
+easyblock = 'PythonBundle'
+
+name = 'nvidia-cutlass'
+version = '3.8.0.0'
+versionsuffix = '-CUDA-%(cudaver)s'
+
+homepage = 'https://pypi.org/project/nvidia-cutlass'
+description = """
+CUTLASS is a collection of CUDA C++ template abstractions for implementing high-performance
+matrix-matrix multiplication (GEMM) and related computations at all levels and scales within CUDA.
+It incorporates strategies for hierarchical decomposition and data movement similar to those used
+to implement cuBLAS and cuDNN.
+CUTLASS decomposes these "moving parts" into reusable, modular software components abstracted by C++ template classes.
+Primitives for different levels of a conceptual parallelization hierarchy can be specialized and tuned
+via custom tiling sizes, data types, and other algorithmic policy.
+The resulting flexibility simplifies their use as building blocks within custom kernels and applications.
+"""
+
+toolchain = {'name': 'gfbf', 'version': '2024a'}
+
+builddependencies = [
+    ('poetry', '1.8.3'),
+]
+
+dependencies = [
+    ('CUDA', '12.6.0', '', SYSTEM),
+    ('CUDA-Python', '12.6.2.post1', versionsuffix),
+    ('Python', '3.12.3'),
+    ('Python-bundle-PyPI', '2024.06'),
+    ('SciPy-bundle', '2024.05'),
+    ('networkx', '3.4.2'),
+    ('pydot', '3.0.3'),
+]
+
+exts_list = [
+    ('treelib', '1.8.0', {
+        'sources': [SOURCE_TAR_GZ],
+        'checksums': ['e1be2c6b66ffbfae85079fc4c76fb4909946d01d915ee29ff6795de53aed5d55'],
+    }),
+    (name, version, {
+        'source_tmpl': 'nvidia_cutlass-%(version)s-py3-none-any.whl',
+        'post_install_patches': [{
+            'name': 'nvidia-cutlass-3.8.0.0_fix-BytesWarning.patch',
+            'sourcepath': 'lib/python%(pyshortver)s/site-packages/cutlass',
+            'level': 3,
+        }],
+        'checksums': [
+            '013147221a63500205da233ae02e6262463917f3fe39cb09efbca37bfd1c39f9',
+            {'nvidia-cutlass-3.8.0.0_fix-BytesWarning.patch':
+             '63eb47894340c0ea03d0d2faaa49c1979915f903b5bc2ced17f8e0dd5ab854ed'},
+        ],
+        'modulename': 'cutlass',
+    }),
+]
+
+sanity_check_commands = [
+    'python -sc "import cutlass_library"',
+    'python -bb -sc "' + '; '.join((
+        'import cutlass',
+        # These serves as a smoke test, e.g. nvcc_version() was incompatible with -bb
+        "assert cutlass.nvcc_version().startswith('%(cudamajver)s')",
+        "assert cutlass.cuda_install_path() == '$EBROOTCUDA'",
+    )) + '"',
+]
+
+moduleclass = 'lib'

easybuild/easyconfigs/n/nvidia-cutlass/nvidia-cutlass-3.8.0.0_fix-BytesWarning.patch

@@ -0,0 +1,39 @@
+When neither `text` nor `encoding` is specified to `subprocess.run` then `stdout` will be of type `bytes`.
+A subsequent `str(stdout)` causes a `BytesWarning` which might result in errors, e.g. if `python -bb` is used.
+See https://github.com/NVIDIA/cutlass/pull/2682
+
+Fixes hard failures in PyTorch tests that do use `-bb`.
+
+Author: Alexander Grund (TU Dresden)
+
+diff --git a/python/cutlass_cppgen/__init__.py b/python/cutlass_cppgen/__init__.py
+index 9bdd259c02..0e28ff55fd 100644
+--- a/python/cutlass_cppgen/__init__.py
++++ b/python/cutlass_cppgen/__init__.py
+@@ -39,11 +39,11 @@
+ def _cuda_install_path_from_nvcc() -> str:
+     import subprocess
+     # Attempt to detect CUDA_INSTALL_PATH based on location of NVCC
+-    result = subprocess.run(['/usr/bin/which', 'nvcc'], capture_output=True)
++    result = subprocess.run(['/usr/bin/which', 'nvcc'], capture_output=True, text=True)
+     if result.returncode != 0:
+         raise Exception(f'Unable to find nvcc via `which` utility.')
+ 
+-    cuda_install_path = result.stdout.decode('utf-8').split('/bin/nvcc')[0]
++    cuda_install_path = result.stdout.split('/bin/nvcc')[0]
+     if not os.path.isdir(cuda_install_path):
+         raise Exception(f'Environment variable "CUDA_INSTALL_PATH" is not defined, '
+                         f'and default path of {cuda_install_path} does not exist.')
+@@ -63,10 +63,10 @@ def nvcc_version():
+         import subprocess
+ 
+         # Attempt to get NVCC version
+-        result = subprocess.run(['nvcc', '--version'], capture_output=True)
++        result = subprocess.run(['nvcc', '--version'], capture_output=True, text=True)
+         if result.returncode != 0:
+             raise Exception('Unable to run `nvcc --version')
+-        _NVCC_VERSION = str(result.stdout).split(" release ")[-1].split(",")[0]
++        _NVCC_VERSION = result.stdout.split(" release ")[-1].split(",")[0]
+     return _NVCC_VERSION
+ 
+ _CUDA_INSTALL_PATH = None