[Cpp API Compatibility] Fix MaybeResetHolder

[youge325]

PR Category

Execute Infrastructure

PR Types

Bug fixes

Description

本次修复了外部扩展场景下 TensorBase::storage() 未能同步 compat storage holder 的问题。

问题原因

在 paddle.enable_compat() 编译外部扩展时，phi::DenseTensor::ResetHolder() 不可见，原来的 MaybeResetHolder() 会退化到空操作，导致 DenseTensor 的 holder 仍然是原始 Allocation，没有被替换成 StorageHolderView。因此后续 resize_() 扩容虽然更新了 compat StorageImpl，但 tensor 再次读取 storage().nbytes() 时仍可能从旧 holder 得到旧容量。

修复内容

• 保留 ResetHolder() 可见时的直接调用路径
• 为 custom-kernel/外部扩展场景补充 fallback：
  • 通过公开 Holder() 返回的 live holder 引用替换为 compat holder。
  • 复刻 ResetHolder() 的关键语义：空 tensor 重置 offset；连续 tensor 校验 holder 容量。
• 未修改 phi/core，改动限定在 compat 层。

验证

• ninja -C build test/cpp/c10_storage_test test/cpp/ATen_resize_test 通过
• ctest -R 'c10_storage_test|ATen_resize_test' --output-on-failure 通过
• -DPADDLE_WITH_CUSTOM_KERNEL 语法验证通过
• 外部扩展复现通过：resize_({4}) 后 storage().nbytes() 从 8 更新为 16

上面的说明是 AI 生成的，大概意思是，paddle/phi/api/include/compat/ATen/core/TensorBase.h:416-422 中定义的MaybeResetHolder 模板实例化的过程失败了，因为在启用 PADDLE_WITH_CUSTOM_KERNEL 的情况下，paddle/phi/core/dense_tensor.h 不会 #include "paddle/phi/core/dense_tensor.inl" （详见 paddle/phi/core/dense_tensor.h:342-344 ），导致在 paddle/phi/core/dense_tensor.inl 中声明的 phi::DenseTensor::ResetHolder() 没有符号导出，调用 ResetHolder 函数的 MaybeResetHolder 函数自然就实例化失败了，现在就直接显式实例化一份 MaybeResetHolder，在里面将 ResetHolder 函数展开

这个问题比较隐蔽，在 C++ 模板编程的 SFINAE（Substitution Failure Is Not An Error）机制中，编译不会报错，导致 resize_ 函数调用了一个空的 MaybeResetHolder ，现已修复

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.

[Copilot]

Pull request overview

Fixes an external-extension compatibility issue where TensorBase::storage() could observe stale capacity because the compat StorageHolderView was not synced back into phi::DenseTensor when ResetHolder() is not available (e.g., custom-kernel / external builds).

Changes:

• Replaces the MaybeResetHolder(...) no-op fallback with a phi::DenseTensor*-specific implementation that updates the live holder via Holder().
• Mirrors key ResetHolder() semantics in the fallback: empty-tensor offset reset and contiguous-tensor capacity validation.

---

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

[Copilot]

In the numel() == 0 branch, updating offset via dense->set_meta(meta) can throw when meta.dims.size() == -1 and meta.strides.size() == -1, because set_meta will call DenseTensorMeta::calc_strides(meta_.dims) and calc_strides does not handle dims.size() == -1. ResetHolder() doesn’t have this failure mode. Consider resetting the offset without going through set_meta (e.g., mutate meta().offset via a guarded const_cast, or ensure strides/dims are set to a supported rank before calling set_meta).

Suggested change

	auto meta = dense->meta();
	meta.offset = 0;
	dense->set_meta(meta);
	auto& meta =
	const_cast<phi::DenseTensorMeta&>(dense->meta());
	meta.offset = 0;

[Copilot]

The null-holder check uses the same message as the size check ("The size of Holder is not enough..."). If holder is actually null, this error is misleading. Suggest using a distinct message for the null case (e.g., "Holder must not be null") and keep the size-related message for the required_bytes <= holder->size() check.

Suggested change

	TORCH_CHECK(holder != nullptr,
	"The size of Holder is not enough to store the Tensor.");
	TORCH_CHECK(holder != nullptr, "Holder must not be null");