[Environment Adaptation] Add sm_120 to Blackwell CUDA arch selection

[BostonSupremeMantou]

PR Category

Environment Adaptation

PR Types

Improvements

Description

This PR updates CUDA architecture selection for Blackwell GPUs by adding sm_120 alongside sm_100 when the CUDA compiler supports native Blackwell code generation.

It also guards CUDA_ARCH_NAME=Blackwell behind CUDA 12.8+, because older CUDA toolchains such as CUDA 12.4 do not recognize compute_120. For CUDA 12.x releases before 12.8, CUDA_ARCH_NAME=All continues to exclude sm_100/sm_120 to avoid generating unsupported NVCC flags.

This is a build configuration readiness change, not a full claim of RTX 50-series runtime support.

Related context: #79314 handles CUDA 13 sm_121 for GB10 / DGX Spark. This PR is intentionally focused on sm_120 and the CUDA 12.8+ guard.

Tests run:

• git diff --check
• pre-commit run --files cmake/cuda.cmake test/compat/test_cpp_extension_api.py
• CMake probe with CUDA 12.4: CUDA_ARCH_NAME=Blackwell fails with the expected CUDA 12.8+ requirement
• CMake probe with CUDA 12.4: CUDA_ARCH_NAME=All does not emit sm_100/sm_120
• CMake probe with CUDA 12.8: CUDA_ARCH_NAME=Blackwell emits sm_100/sm_120
• python -m unittest discover -s test/compat -p test_cpp_extension_api.py -k blackwell

是否引起精度变化

否

[CLAassistant]

All committers have signed the CLA.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | pr_review | 2026-06-18 11:16:46

📋 Review 摘要

PR 概述：为 Blackwell CUDA 架构选择补充 sm_120，并对 CUDA 12.8 以下工具链增加门禁。
变更范围：cmake/cuda.cmake、python/paddle/utils/cpp_extension/extension_utils.py、test/compat/test_cpp_extension_api.py
影响面 Tag：Build Python API Tests

问题

未发现阻塞性问题。

历史 Findings 修复情况

Finding	问题	状态
F1	PADDLE_CUDA_ARCH_LIST 显式写数值 Blackwell 架构时会绕过 CUDA 12.8 检查	✅ 已修复

📝 PR 规范检查

符合规范。

总体评价

当前 diff 在命名架构展开后统一检查 Blackwell 数值架构，覆盖了 10.0、10.1、12.0+PTX 以及自动探测到的 12.0 路径，历史绕过问题已修复。未发现新的阻塞性问题；本地尝试运行 python -m unittest discover -s test/compat -p test_cpp_extension_api.py -k blackwell 时因当前环境缺少 paddle.base 未能执行测试断言。