[Cpp API Compatibility] Align some compat APIs#78484
[Cpp API Compatibility] Align some compat APIs#78484youge325 wants to merge 22 commits intoPaddlePaddle:developfrom
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
There was a problem hiding this comment.
Pull request overview
This PR updates Paddle’s C++ ATen/c10 compatibility layer to better align API shapes with PyTorch and removes legacy DeepEP-transition interfaces.
Changes:
- Tighten CUDA stream-pool APIs by adding a
getStreamFromPool(int priority, ...)overload and mapping the bool form onto it; also add missing CUDA error checking. - Expose
c10::Streamasat::Stream. - Remove legacy
raw_stream()/cudaStream_tcompatibility overloads for stream/event/record_stream.
Reviewed changes
Copilot reviewed 5 out of 5 changed files in this pull request and generated 4 comments.
Show a summary per file
| File | Description |
|---|---|
| paddle/phi/api/include/compat/c10/cuda/CUDAStream.h | Adds priority-based pool API + adjusts high/low selection semantics; removes legacy raw_stream(); adds error-checking for priority range query. |
| paddle/phi/api/include/compat/c10/core/Stream.h | Re-exports c10::Stream into namespace at. |
| paddle/phi/api/include/compat/c10/core/Event.h | Removes legacy record(cudaStream_t) compatibility overload. |
| paddle/phi/api/include/compat/ATen/ops/record_stream.h | Removes record_stream(at::cuda::CUDAStream) and record_stream(cudaStream_t) overloads, leaving record_stream(at::Stream). |
| paddle/phi/api/include/compat/ATen/core/TensorBody.h | Removes corresponding Tensor::record_stream overload declarations for CUDAStream/cudaStream_t. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
|
|
||
| namespace at { | ||
| using c10::Stream; | ||
| } |
There was a problem hiding this comment.
c10::Stream is now exposed as at::Stream, but (unlike most other c10/core/* compat headers) it is not re-exported into namespace torch. If the goal is that namespace at APIs are also visible under torch (as mentioned in the PR description and as done e.g. in c10/core/Device.h), consider adding a namespace torch { using c10::Stream; } re-export here as well to keep the surface consistent for downstream users.
| } | |
| } | |
| namespace torch { | |
| using c10::Stream; | |
| } |
| } | ||
|
|
||
| void record_stream(at::Stream s) const; |
There was a problem hiding this comment.
After removing the record_stream(at::cuda::CUDAStream) / record_stream(cudaStream_t) overloads, the earlier forward-declaration block for c10::cuda::CUDAStream in this header (and its accompanying comment about the overload) is now unused/misleading. Please remove or update that forward-declaration/comment block to reflect the new API surface.
| #endif | ||
| } | ||
|
|
||
| #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP) | ||
| inline void Tensor::record_stream(at::cuda::CUDAStream s) const { | ||
| record_stream(static_cast<at::Stream>(s)); | ||
| } | ||
|
|
||
| // TODO(youge325): Remove after DeepEP paddle branch is updated to use | ||
| // at::Stream | ||
| inline void Tensor::record_stream(cudaStream_t s) const { | ||
| auto dense_tensor = | ||
| std::dynamic_pointer_cast<phi::DenseTensor>(tensor_.impl()); | ||
| PD_CHECK(dense_tensor != nullptr, | ||
| "record_stream only supports DenseTensor, but got a non-dense " | ||
| "tensor implementation."); | ||
| PD_CHECK(dense_tensor->place().GetType() != phi::AllocationType::CPU, | ||
| "record_stream is not supported for CPU tensors."); | ||
| paddle::memory::RecordStream(dense_tensor->Holder(), s); | ||
| } | ||
| #endif | ||
| } // namespace at |
There was a problem hiding this comment.
With the CUDAStream/cudaStream_t record_stream overloads removed from this header, the conditional #include <c10/cuda/CUDAStream.h> at the top of this file appears to be unused now. Consider dropping that include to reduce unnecessary dependencies/compile time.
| inline CUDAStream getStreamFromPool(const bool isHighPriority = false, | ||
| c10::DeviceIndex device_index = -1); | ||
|
|
There was a problem hiding this comment.
The getStreamFromPool(bool, device_index) forward declaration is only needed to carry default arguments, but it makes the header a bit harder to follow since the actual implementation is below. Consider moving the defaults onto the inline definition (and removing the separate forward declaration), or defining the bool overload before the int priority overload, to avoid a declaration-only prototype.
|
/re-run all-failed |
PaddleFleet 有更新 DeepEP commit 的 PR 么？这个 PR 直接删不行的呀，没有形成「Paddle 更新（保留兼容接口）」->「DeepEP 更新」->「PaddleFleet 更新」->「Paddle 更新（清理兼容接口）」的闭环 另外这个 PR 建议等等 #78389，会解决 CI 里测不到编译问题的问题 |
好的，我等这个CI过了再 rebase 一遍，后面还有一些用来对齐 Device 相关接口的 commits |
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## develop #78484 +/- ##
==========================================
Coverage ? 97.20%
==========================================
Files ? 8
Lines ? 143
Branches ? 0
==========================================
Hits ? 139
Misses ? 4
Partials ? 0 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
|
@ShigureNyako 来 review 下这个 PR |
|
/re-run all-failed |
|
这个 PR 是不是可以 close 了？应该只剩 #78549 了 |
PR Category
Execute Infrastructure
PR Types
Bug fixes
Description
移除 DeepEP 已经迁移完成的旧接口
将
c10::Stream暴露到at::Stream，namespace at下的接口均会暴露到namespace torch，DeepEP有一个接口暂时用的是at::cuda::CUDAStream，要改回torch::Stream修复编译 FlashMLA 遇到的 redefinition 错误，如果要用
at::cuda::is_available()，需要#include <ATen/cuda/CUDAContext.h>，而不是#include <torch/csrc/api/include/torch/cuda.h>修复
at::arange在不指定dtype时创建的 tensor 数据类型不是kLong的错误，用于DeepGEMM的对齐已对齐的接口可以在 PFCCLab/PaddleCppAPITest#58 中查看
是否引起精度变化
否