[release/3.4][CUDA] Support Paddle on CUDA 13.2

[ShigureNyako]

PR Category

Environment Adaptation

PR Types

Improvements

Description

devPR:#78720

Backport #78720 to release/3.4 as a single squashed commit.

Source head: aa3920f0c43660700e1ba68e63a4ddb38fbc908b.
Backport commit: 363ad9abe2d0aac09f73132862990171032594f1.

Changes included:

• Add CUDA 13.x build branch/arch configuration and CUDA 13.x nvcc flags.
• Restrict bundled flash-attn enablement to CUDA 12.3-12.9 and update third_party/flashattn to the official PaddlePaddle commit required by CUDA 13.2.
• Add CUDA 13.2 Python package requirements, TensorRT cu13 handling, manylinux CUDA 13.2 Dockerfile/stage, and install_cuda.sh 13.2 support.
• Add nvcc 13.x workarounds for kernel registration and range/arange explicit instantiations.

Validation:

• git diff HEAD --check
• bash -n tools/dockerfile/manylinux/common/install_cuda.sh
• python3 -m py_compile setup.py
• Targeted consistency check for CUDA 13.2 markers across the touched CMake/Python/Dockerfile/shell files
• Commit hook checks passed: ruff, clang-format, cmake-format, CMake lint, whitespace/conflict/private-key checks, and related repository pre-commit hooks

Not run:

• Full Paddle/CUDA build: not feasible locally in this environment.
• Docker BuildKit --check: attempted, but Docker daemon is not running (Cannot connect to the Docker daemon at unix:///var/run/docker.sock).

是否引起精度变化

否

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[ShigureNyako]

自查完成，backport diff 已核对。

• GitHub PR diff 与本地 release/3.4 backport commit 的 stable patch-id 一致：db9e43689b5334b198557a15265035efdf67bbbf。
• 本地验证通过：git diff HEAD --check、bash -n tools/dockerfile/manylinux/common/install_cuda.sh、python3 -m py_compile setup.py、CUDA 13.2 相关 CMake/Python/Dockerfile/shell marker 一致性检查。
• Commit hook 已通过 ruff、clang-format、cmake-format、CMake lint 等仓库检查。
• 未跑完整 Paddle/CUDA 构建；本机 Docker daemon 未运行，Docker BuildKit --check 无法连接 daemon。

@SigureMo 请 review 这个 release/3.4 backport PR。

[ShigureNyako]

/re-run all-failed

[SigureMo]

@ShigureNyako 这个 PR 是 cherry-pick，要和 develop 完全一致，你给我把刚刚的 commit revert 了！！！

[ShigureNyako]

@SigureMo 已撤销刚刚的 review-fix commit，并强推回原来的单 commit cherry-pick 状态：363ad9abe2d0aac09f73132862990171032594f1。

我重新核对了 patch-id，当前 backport diff 与 develop 源 PR #78720 的 diff 一致：db9e43689b5334b198557a15265035efdf67bbbf。PR 描述也已同步改回只描述原始 cherry-pick/backport。

[ShigureNyako]

/re-run ixuca mac

[ShigureNyako]

/re-run approval

[ShigureNyako]

Check approval 已在 SigureMo approve 后重跑，当前仍显示缺少 required RD approval：

• setup.py / python/setup.py.in：需要 risemeup1 或 zhangbo9674 任一 approval（当前已有 zhangbo9674 review request）。
• paddle/phi/*：需要 wanghuancoder / zrr1999 / DanielSun11 任一 approval。

@zhangbo9674 @wanghuancoder 麻烦有空帮忙看下这个 release/3.4 cherry-pick 的 approval gate。当前 PR 仍保持与源 PR #78720 diff 一致。

[risemeup1]

LGTM

[wanghuancoder]

LGTM

[sneaxiy]

LGTM