[CINN] Make CINN compilable with arm CPU and add missing nvidia-* dependencies for arm wheel

[SigureMo]

PR Category

Environment Adaptation

PR Types

Devs

Description

使 Paddle 能够在 WITH_CINN 下编译、运行，这包含如下几部分

• 添加 ginac、isl、llvm arm64 预编译包版本，并上传 bos，其中 llvm 使用低版本 gcc8 进行编译，确保只包含 glibc2.17 以保证最好的兼容性
• 更新 CUDA 13 下的 nvidia-* wheel 包依赖：
  • CUDA 13+ 下不再限定仅 x86 下安装，arm64 也会安装（之前会导致依赖里丢掉这些 nvidia 包，用户在没配置好环境的情况下可能跑不起来）
  • 更新 nvidia-cudnn-cu13 到 9.15.1.9、nvidia-cublas 到 13.1.0.3

^{This PR is co-authored with @copilot (gpt-5.4)}

Remaining issues

正是因为之前极度依赖于开发镜像里装了 nvidia 相关依赖，所以 wheel 包缺失 nvidia 依赖才没被发现，但是在没有 wheel 包依赖的情况下，CINN 单测会暴露出默认 nvrtc 编译时找不到系统里 nv 相关头文件的问题，这个问题依然存在，但是对于用户而言，通过装 wheel 包的方式装 nv 相关依赖，该问题就不存在了

后续如果要在 ARM CPU上（也可能是 cu13 的问题，不是 ARM 的问题）跑单测或者开发可以注意下这一点，单独适配下，这个就暂时低优了

是否引起精度变化

是

[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.

[Copilot]

Pull request overview

This PR updates Paddle/CINN build configuration to compile under WITH_CINN on ARM CPUs by removing x86-only assumptions and introducing reusable CMake helpers for architecture-aware paths.

Changes:

• Add cmake/architecture.cmake with helpers to normalize architecture, detect CUDA target directories, and map LLVM native targets.
• Update multiple CUDA/CINN-related CMake scripts to use detected target directories (CUDA targets/<arch> layout) and avoid x86-only flags on non-x86 platforms.
• Tighten x86 feature detection in pir spin lock and adjust CINN LLVM runtime IR generation flags to be conditional on x86_64.

Reviewed changes

Copilot reviewed 14 out of 14 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
paddle/pir/include/core/spin_lock.h	Avoid defining x86-only relax primitive when building for ARM64.
paddle/phi/CMakeLists.txt	Make NVTX3 include path depend on detected CUDA target dir.
paddle/cinn/backends/llvm/CMakeLists.txt	Make LLVM IR generation flags conditional on x86_64; use configurable LLVM tool paths.
paddle/cinn/backends/CMakeLists.txt	Gate x86-only fake lib build behind x86_64 check.
cmake/cupti.cmake	Use normalized arch and detected CUDA target dir for include/lib lookup.
cmake/cudnn.cmake	Add detected CUDA target include/lib directories to cuDNN lookup.
cmake/cinn/nvtx.cmake	Add detected CUDA target lib directories to NVTX lookup.
cmake/cinn/llvm.cmake	Link LLVM target component based on native architecture instead of hardcoded X86.
cmake/cinn/external/llvm.cmake	Select ARM-specific LLVM prebuilt and use native target component; expose LLVM tool executables.
cmake/cinn/external/isl.cmake	Add ARM-specific ISL prebuilt selection and ensure download directory exists.
cmake/cinn/external/ginac.cmake	Add ARM-specific GiNaC prebuilt selection and ensure download directory exists.
cmake/cinn/core.cmake	Apply AVX/FMA flags only on x86_64 to allow ARM builds.
cmake/cinn.cmake	Remove hardcoded system lib include dir; search CUDA libraries in detected target dirs; prefer LLVM include dirs when available.
cmake/architecture.cmake	New shared CMake module for arch normalization and CUDA/LLVM target detection.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 75.00000% with 2 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@894a8e5). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
python/paddle/__init__.py	75.00%	2 Missing ⚠️

❌ Your patch status has failed because the patch coverage (75.00%) is below the target coverage (90.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop   #78712   +/-   ##
==========================================
  Coverage           ?   75.00%           
==========================================
  Files              ?        1           
  Lines              ?        8           
  Branches           ?        0           
==========================================
  Hits               ?        6           
  Misses             ?        2           
  Partials           ?        0           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[risemeup1]

LGTM

[risemeup1]

LGTM

[wanghuancoder]

LGTM

[From00]

LGTM

[lugimzzz]

lgtm

[risemeup1111]

✅ Cherry-pick successful! Created PR: #78730