Update CUDA compiler matrix (#86360)
Switch GCC/Clang max versions to be exclusive as the `include/crt/host_config.h` checks the major version only for the upper bound. This allows to be less restrictive and match the checks in the aforementioned header.
Also update the versions using that header in the CUDA SDKs.
Follow up to #82860
I noticed this as PyTorch 1.12.1 with CUDA 11.3.1 and GCC 10.3 was failing in the `test_cpp_extensions*` tests.
Example for CUDA 11.3.1 from the SDK header:
```
#if __GNUC__ > 11
// Error out
...
#if (__clang_major__ >= 12) || (__clang_major__ < 3) || ((__clang_major__ == 3) && (__clang_minor__ < 3))
// Error out
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/86360
Approved by: https://github.com/ezyang
diff --git a/torch/utils/cpp_extension.py b/torch/utils/cpp_extension.py
index 7209352..d74ef9a 100644
--- a/torch/utils/cpp_extension.py
+++ b/torch/utils/cpp_extension.py
@@ -18,7 +18,7 @@
from ._cpp_extension_versioner import ExtensionVersioner
from .hipify import hipify_python
from .hipify.hipify_python import GeneratedFileCleaner
-from typing import List, Optional, Union, Tuple
+from typing import Dict, List, Optional, Union, Tuple
from torch.torch_version import TorchVersion
from setuptools.command.build_ext import build_ext
@@ -42,29 +42,35 @@
MINIMUM_GCC_VERSION = (5, 0, 0)
MINIMUM_MSVC_VERSION = (19, 0, 24215)
+VersionRange = Tuple[Tuple[int, ...], Tuple[int, ...]]
+VersionMap = Dict[str, VersionRange]
# The following values were taken from the following GitHub gist that
# summarizes the minimum valid major versions of g++/clang++ for each supported
# CUDA version: https://gist.github.com/ax3l/9489132
-CUDA_GCC_VERSIONS = {
- '10.2': (MINIMUM_GCC_VERSION, (8, 0, 0)),
- '11.1': (MINIMUM_GCC_VERSION, (10, 0, 0)),
- '11.2': (MINIMUM_GCC_VERSION, (10, 2, 1)),
- '11.3': (MINIMUM_GCC_VERSION, (10, 2, 1)),
- '11.4': ((6, 0, 0), (11, 5, 0)),
- '11.5': ((6, 0, 0), (11, 5, 0)),
- '11.6': ((6, 0, 0), (11, 5, 0)),
- '11.7': ((6, 0, 0), (11, 5, 0)),
+# Or from include/crt/host_config.h in the CUDA SDK
+# The second value is the exclusive(!) upper bound, i.e. min <= version < max
+CUDA_GCC_VERSIONS: VersionMap = {
+ '10.2': (MINIMUM_GCC_VERSION, (9, 0)),
+ '11.0': (MINIMUM_GCC_VERSION, (10, 0)),
+ '11.1': (MINIMUM_GCC_VERSION, (11, 0)),
+ '11.2': (MINIMUM_GCC_VERSION, (11, 0)),
+ '11.3': (MINIMUM_GCC_VERSION, (11, 0)),
+ '11.4': ((6, 0, 0), (12, 0)),
+ '11.5': ((6, 0, 0), (12, 0)),
+ '11.6': ((6, 0, 0), (12, 0)),
+ '11.7': ((6, 0, 0), (12, 0)),
}
-CUDA_CLANG_VERSIONS = {
- '10.2': ((3, 3, 0), (8, 0, 0)),
- '11.1': ((6, 0, 0), (9, 0, 0)),
- '11.2': ((6, 0, 0), (9, 0, 0)),
- '11.3': ((6, 0, 0), (11, 0, 0)),
- '11.4': ((6, 0, 0), (11, 0, 0)),
- '11.5': ((6, 0, 0), (12, 0, 0)),
- '11.6': ((6, 0, 0), (12, 0, 0)),
- '11.7': ((6, 0, 0), (13, 0, 0)),
+MINIMUM_CLANG_VERSION = (3, 3, 0)
+CUDA_CLANG_VERSIONS: VersionMap = {
+ '10.2': (MINIMUM_CLANG_VERSION, (9, 0)),
+ '11.1': (MINIMUM_CLANG_VERSION, (11, 0)),
+ '11.2': (MINIMUM_CLANG_VERSION, (12, 0)),
+ '11.3': (MINIMUM_CLANG_VERSION, (12, 0)),
+ '11.4': (MINIMUM_CLANG_VERSION, (13, 0)),
+ '11.5': (MINIMUM_CLANG_VERSION, (13, 0)),
+ '11.6': (MINIMUM_CLANG_VERSION, (14, 0)),
+ '11.7': (MINIMUM_CLANG_VERSION, (14, 0)),
}
__all__ = ["get_default_build_root", "check_compiler_ok_for_platform", "get_compiler_abi_compatibility_and_version", "BuildExtension",
@@ -388,20 +394,19 @@
_is_binary_build()):
return
- cuda_compiler_bounds = CUDA_CLANG_VERSIONS if compiler_name.startswith('clang') else CUDA_GCC_VERSIONS
+ cuda_compiler_bounds: VersionMap = CUDA_CLANG_VERSIONS if compiler_name.startswith('clang') else CUDA_GCC_VERSIONS
if cuda_str_version not in cuda_compiler_bounds:
warnings.warn(f'There are no {compiler_name} version bounds defined for CUDA version {cuda_str_version}')
else:
- min_compiler_version, max_compiler_version = cuda_compiler_bounds[cuda_str_version]
- # Special case for 11.4.0, which has lower compiler bounds that 11.4.1
+ min_compiler_version, max_excl_compiler_version = cuda_compiler_bounds[cuda_str_version]
+ # Special case for 11.4.0, which has lower compiler bounds than 11.4.1
if "V11.4.48" in cuda_version_str and cuda_compiler_bounds == CUDA_GCC_VERSIONS:
- max_compiler_version = (10, 0, 0)
+ max_excl_compiler_version = (11, 0)
min_compiler_version_str = '.'.join(map(str, min_compiler_version))
- max_compiler_version_str = '.'.join(map(str, max_compiler_version))
+ max_excl_compiler_version_str = '.'.join(map(str, max_excl_compiler_version))
- version_bound_str = f'>={min_compiler_version_str}'
- version_bound_str = f'{version_bound_str}, <={max_compiler_version_str}'
+ version_bound_str = f'>={min_compiler_version_str}, <{max_excl_compiler_version_str}'
if compiler_version < TorchVersion(min_compiler_version_str):
raise RuntimeError(
@@ -409,10 +414,10 @@
f'than the minimum required version by CUDA {cuda_str_version} ({min_compiler_version_str}). '
f'Please make sure to use an adequate version of {compiler_name} ({version_bound_str}).'
)
- if compiler_version > TorchVersion(max_compiler_version_str):
+ if compiler_version >= TorchVersion(max_excl_compiler_version_str):
raise RuntimeError(
f'The current installed version of {compiler_name} ({compiler_version}) is greater '
- f'than the maximum required version by CUDA {cuda_str_version} ({max_compiler_version_str}). '
+ f'than the maximum required version by CUDA {cuda_str_version}. '
f'Please make sure to use an adequate version of {compiler_name} ({version_bound_str}).'
)
