[CI] Disable auto CUDA arch injection to avoid duplicate gencode flags

[EmmonsCurse]

Motivation

Recent changes in Paddle #78704 modified the behavior of CUDAExtension, introduced automatic CUDA architecture flag injection via PADDLE_CUDA_ARCH_LIST, even when custom -gencode flags are already specified.

This results in duplicated CUDA arch flags during compilation, increasing binary size and potentially causing linker errors such as:

• relocation truncated to fit

To maintain stable builds and avoid unnecessary code generation, a workaround is required.

Modifications

• Patched extension_utils._get_cuda_arch_flags to return an empty list when user-defined -gencode flags are detected, preventing Paddle from auto-injecting CUDA arch flags.
• Added a secondary safeguard by overriding CUDAExtension._add_cuda_arch_flags to ensure no additional arch flags are appended internally.
• Explicitly controlled CUDA architecture flags via get_gencode_flags, avoiding reliance on PADDLE_CUDA_ARCH_LIST.
• Effectively disabled Paddle’s automatic CUDA arch injection mechanism to prevent duplicated -gencode entries.
• Ensured correct generation of arch=compute_xxa,code=sm_xxa pairs (e.g., 90a, 100a) and avoided incomplete flags like arch=compute_90a.
• Reduced the risk of compilation and linking issues (e.g., relocation overflow) caused by conflicting or duplicated CUDA arch flags.

Usage or Command

N/A

Accuracy Tests

N/A

Checklist

• Add at least a tag in the PR title.
  • Tag list: [[FDConfig],[APIServer],[Engine], [Scheduler], [PD Disaggregation], [Executor], [Graph Optimization], [Speculative Decoding], [RL], [Models], [Quantization], [Loader], [OP], [KVCache], [DataProcessor], [BugFix], [Docs], [CI], [Optimization], [Feature], [Benchmark], [Others], [XPU], [HPU], [GCU], [DCU], [Iluvatar], [Metax]]
  • You can add new tags based on the PR content, but the semantics must be clear.
• Format your code, run pre-commit before commit.
• Add unit tests. Please write the reason in this PR if no unit tests.
• Provide accuracy results.
• If the current PR is submitting to the release branch, make sure the PR has been submitted to the develop branch, then cherry-pick it to the release branch with the [Cherry-Pick] PR tag.

[paddle-bot]

Thanks for your contribution!

[EmmonsCurse]

/skip-ci ci_iluvatar
/skip-ci ci_hpu
/skip-ci stable_test
/skip-ci base_test
/skip-ci pre_ce_test

[PaddlePaddle-bot]

🤖 AI Code Review | 2026-04-20 17:10:08

📋 Review 摘要

PR 概述：通过 monkey-patch Paddle 的 _get_cuda_arch_flags 函数，防止 Paddle PR #78704 引入的自动 CUDA arch flag 注入导致重复 gencode 标志和链接错误。
变更范围：custom_ops/setup_ops.py（构建配置）
影响面 Tag：CI OP

📝 PR 规范检查

PR 描述中 Modifications 部分与实际实现不完全一致：描述称"Overrode PADDLE_CUDA_ARCH_LIST by setting it to an empty string in get_gencode_flags"，但实际实现是 monkey-patch extension_utils._get_cuda_arch_flags，未修改 get_gencode_flags 函数本身，也未设置环境变量。建议更新描述以准确反映实际实现方式。

问题

级别	文件	概述
🟡 建议	custom_ops/setup_ops.py:63	"第二道防线" safeguard 大概率为死代码
🟡 建议	custom_ops/setup_ops.py:52	flag 子串匹配存在误匹配风险

总体评价

该 PR 通过 monkey-patch 方式解决了 Paddle 上游变更导致的重复 gencode 标志问题，核心 patch 逻辑（_patched_get_cuda_arch_flags）思路正确，注释详尽。两个小建议：1) "第二道防线" 代码块可能无法生效，建议确认或移除；2) flag 匹配条件可以更精确以避免潜在误判。

[PaddlePaddle-bot]

🟡 建议 extension_utils 模块上不太可能存在 CUDAExtension 类属性。

CUDAExtension 是从 paddle.utils.cpp_extension 导入的函数/类，而非 extension_utils 模块的属性。因此 hasattr(extension_utils, "CUDAExtension") 大概率为 False，这段 "第二道防线" 实际上是死代码，无法起到防护作用。

建议：

1. 确认 Paddle 对应版本中 extension_utils 模块是否确实有 CUDAExtension 属性，如果没有则移除这段代码以避免误导；
2. 如果确需额外防护，可考虑直接 patch paddle.utils.cpp_extension.CUDAExtension 本身。

[PaddlePaddle-bot]

🟡 建议 flag 检测逻辑存在误匹配风险。

"compute_" in flag 和 "sm_" in flag 使用子串匹配，可能误匹配非 gencode 标志（例如包含路径中含 sm_ 或 compute_ 的 -I include 路径）。虽然目前场景下概率较低，但作为通用 patch 函数可以更精确。

建议收紧匹配条件，例如仅检查 -gencode 和 -arch 前缀：

if isinstance(flag, str) and (flag.startswith("-gencode") or flag.startswith("-arch")):
    return []