op_builder: use C++20 for nvcc on CUDA 13+

[bheu]

Problem

CUDAOpBuilder.nvcc_args() only passes -std=c++20 to nvcc when CUDA is 12.5+; on CUDA 13.x it falls back to -std=c++17. PyTorch >= 2.12 headers require C++20, so building DeepSpeed CUDA ops against CUDA 13 + torch >= 2.12 fails:

torch/include/torch/csrc/api/include/torch/all.h:5:2: error: #error C++20 or later compatible compiler is required to use PyTorch.
torch/include/ATen/ATen.h:5:2: error: #error C++20 or later compatible compiler is required to use ATen.
torch/include/c10/util/intrusive_ptr.h(775): error: namespace "std" has no member "strong_ordering"
...
error: command '/usr/bin/ccache' failed with exit code 255

Note that cxx_args() already selects -std=c++20 for torch >= 2.12 (see #8238), so only the nvcc path is broken.

Fix

Extend the C++20 condition in nvcc_args() to also cover CUDA 13+:

if (cuda_major == 12 and cuda_minor >= 5) or cuda_major >= 13:

Verification

Built DeepSpeed from source (DS_BUILD_OPS=1 pip install .) with CUDA 13.3, PyTorch 2.14.0, gcc-13: the build fails before this change with the errors above and completes successfully with it.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 11250f027b

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".