[Cpp API Compatibility] Align event api by youge325 · Pull Request #78553 · PaddlePaddle/Paddle

youge325 · 2026-04-01T11:17:13Z

PR Category

Execute Infrastructure

PR Types

Improvements

Description

拆分自 #78484

Align Event related APIs，将 c10::Event 实现为完整的跨平台兼容类。

变更详情

问题背景

历史上 c10::Event 存在以下问题：

被 #ifdef PADDLE_WITH_CUDA 包裹，非 CUDA 构建时不可用
缺少 EventFlag 枚举和对应构造函数
缺少移动语义和属性读取接口
CPU 路径异常行为不一致

变更内容 (c10/core/Event.h)

重构 c10::Event 为完整的跨平台兼容类：

// EventFlag 枚举
enum class EventFlag { PYTORCH_DEFAULT, BACKEND_DEFAULT, INVALID };

// 完整构造函数
Event(const DeviceType device_type,
      const EventFlag flag = EventFlag::PYTORCH_DEFAULT);

// 禁止拷贝、允许移动
Event(const Event&) = delete;
Event& operator=(const Event&) = delete;
Event(Event&&) = default;
Event& operator=(Event&&) = default;

// 属性读取
Device device() const noexcept;
DeviceType device_type() const noexcept;
DeviceIndex device_index() const noexcept;
EventFlag flag() const noexcept;
bool was_marked_for_recording() const noexcept;

// 记录方法
void record(const Stream& stream);
void recordOnce(const Stream& stream);

// 同步/查询
void block(const Stream& stream) const;
bool query() const;
void synchronize() const;
double elapsedTime(const Event& event) const;
void* eventId() const;

跨平台支持

将 c10::Event 移出 #ifdef PADDLE_WITH_CUDA 条件编译
CUDA 构建时包含完整的 CUDA event 功能
非 CUDA 构建时提供占位实现，抛出 Backend doesn't support events 异常

对齐效果

测试项	当前 Paddle	PyTorch	结论
`EventCompatTest.EventDefault` / `EventWithFlag` / `EventRecordThrows` / `EventRecordOnceThrows` / `EventMove` / `EventDevice`	已进入常规回归	一致	✅ 已纳入回归

是否引起精度变化

否

paddle-bot · 2026-04-01T11:17:25Z

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 the C++ compat layer to better match PyTorch’s stream/event APIs, including stream-pool signature parity and a more feature-complete c10::Event surface.

Changes:

Update c10::cuda::getStreamFromPool to support the (int priority, DeviceIndex) overload and map negative priorities to the high-priority pool.
Expand c10::Event to track device metadata, add PyTorch-like methods (e.g., recordOnce, query, eventId, synchronize), and tighten stream/device-type checks.
Remove deprecated/raw stream record_stream overloads and legacy raw_stream() exposure.

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 PyTorch-shaped `getStreamFromPool(int priority, ...)` and updates priority selection logic.
paddle/phi/api/include/compat/c10/core/Stream.h	Exposes `c10::Stream` into `namespace at` for API compatibility.
paddle/phi/api/include/compat/c10/core/Event.h	Refactors/extends `c10::Event` interface and behavior to resemble PyTorch.
paddle/phi/api/include/compat/ATen/ops/record_stream.h	Removes deprecated `record_stream` overloads (CUDAStream / cudaStream_t) in favor of `at::Stream`.
paddle/phi/api/include/compat/ATen/core/TensorBody.h	Removes the corresponding `record_stream` overload declarations.

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

paddle/phi/api/include/compat/c10/core/Stream.h

paddle/phi/api/include/compat/c10/core/Event.h

Copilot · 2026-04-01T11:24:48Z

paddle/phi/api/include/compat/c10/core/Event.h

+#include <c10/core/Stream.h>
+
+#ifdef PADDLE_WITH_CUDA
 #include <c10/cuda/CUDAStream.h>


This header uses std::mutex/std::unique_lock in EventPool, but doesn't include <mutex> directly (it currently relies on transitive includes from other headers). Adding an explicit #include <mutex> under the PADDLE_WITH_CUDA guard would make this file self-contained and less fragile to include-order changes.

Suggested change

#include <c10/cuda/CUDAStream.h>

#include <c10/cuda/CUDAStream.h>

#include <mutex>

Copilot · 2026-04-01T11:24:48Z

paddle/phi/api/include/compat/ATen/core/TensorBody.h

  }

  void record_stream(at::Stream s) const;


Now that the record_stream(at::cuda::CUDAStream) overload was removed, the forward-declaration block for at::cuda::CUDAStream near the top of this file (and its comment) appears unused and misleading (the file no longer declares any overload needing it). Consider removing that forward declaration/comment to avoid stale documentation and unnecessary namespace pollution.

ShigureNyako

这轮我先给 request changes。

主要有两类 blocker：

当前 diff 仍然包含 #78549 那组 legacy API 清理（例如 CUDAStream::raw_stream()、Event::record(cudaStream_t)、record_stream(at::cuda::CUDAStream/cudaStream_t)）。这不是纯粹的 Event 对齐，而是对外兼容层删除；结合 #78484 里已经讨论过的 DeepEP / PaddleFleet 迁移闭环，这里属于隐藏的 BREAKING CHANGE。
Event 对齐本身还没有对齐到 upstream PyTorch 语义：
- cudaEvent_t 在构造时就从全局 pool 创建，绑定的是“构造时当前 device”，不是第一次 record() 的 stream.device()；多卡场景有兼容风险。
- EventFlag 目前没有参与 event 创建，elapsedTime() 直接返回 0.0，会把 timing 语义静默改坏。

建议先把 legacy API 清理从这个 PR 里拆掉/基于 #78549 之后重新整理 diff；同时把 Event 的 lazy-create、device 约束和 timing 语义补齐后再继续。

paddle/phi/api/include/compat/c10/core/Event.h

ShigureNyako · 2026-04-02T05:04:28Z

paddle/phi/api/include/compat/ATen/ops/record_stream.h

 #endif
 }

-#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)


这一段删除不是单纯的 Event 对齐，而是对外可见的兼容层清理：record_stream(at::cuda::CUDAStream) / record_stream(cudaStream_t) 被一起拿掉了。

但从 #78484 的讨论看，这类 DeepEP 迁移接口的清理之前已经被指出需要先完成下游闭环，不能在 split PR 里顺手删掉。现在 #78553 的描述只说对齐 Event API，实际 diff 仍然把 #78549 里的 legacy API 删除一并带进来了，这里属于隐藏的 BREAKING CHANGE。

建议把兼容层清理从这个 PR 里拆掉，或者至少等下游迁移闭环完成后再删；否则这个 split PR 的 scope 和风险都不对。

ShigureNyako

我重新看了一遍这次增量：

之前我提的两类主要问题已经修正：legacy stream API 删除已回退；Event 也改成了 lazy create，并补上了 EventFlag / elapsedTime() 语义和对应测试。
但目前还有一个剩余兼容风险：Event 实现仍然只覆盖 PADDLE_WITH_CUDA，而同层 compat 接口和新增测试却按 PADDLE_WITH_CUDA || PADDLE_WITH_HIP 暴露。这会让 HIP 路径处于“接口看起来存在，但行为/重载并未真正补齐”的半对齐状态。

这一点如果本 PR 不处理，我这边还不能转 approve。

paddle/phi/api/include/compat/c10/core/Event.h

youge325 · 2026-04-02T11:12:52Z

/re-run all-failed

codecov-commenter · 2026-04-02T12:29:38Z

Codecov Report

❌ Patch coverage is 89.28571% with 9 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@638c7c3). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
paddle/phi/api/include/compat/c10/core/Event.h	89.28%	9 Missing ⚠️

❌ Your patch status has failed because the patch coverage (89.28%) 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   #78553   +/-   ##
==========================================
  Coverage           ?   89.28%           
==========================================
  Files              ?        1           
  Lines              ?       84           
  Branches           ?        0           
==========================================
  Hits               ?       75           
  Misses             ?        9           
  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.

ShigureNyako

我重新核对了最新一版：

上一轮提的两类主要问题都已经修正：legacy stream API 删除已回退；Event 的 lazy create / EventFlag / elapsedTime() 语义也已经补齐。
我补提的 HIP 路径半对齐问题，这次也一起补上了：Event 的 backend event / stream 处理已经扩展到 PADDLE_WITH_CUDA || PADDLE_WITH_HIP，测试侧也相应收窄/补齐了宏范围。

就这轮代码 review 来看，我这边没有新的 blocker 了，先转 approve。

当前剩余的是 CI / coverage 侧的问题（包括 coverage 阈值未过），请按 CI 结果继续处理；这部分不属于我这轮 code review 的阻塞项。

SigureMo · 2026-04-02T13:25:06Z

/skip-reason: 未覆盖分支均为正常逻辑不可达分支，为确保与 PyTorch 逻辑对齐而添加的完善逻辑

SigureMo

LGTMeow

youge325 added 6 commits April 1, 2026 19:03

[Cpp API Compatibility] remove CUDAStream::raw_stream()

2f67c07

remove some other APIs

2e43d1f

remove useless annotation

e6d3638

using alias for c10::Stream

91428a6

fix compiling error

dc7610a

align Event related APIs

7b70cb0

Copilot AI review requested due to automatic review settings April 1, 2026 11:17

paddle-bot bot added the contributor External developers label Apr 1, 2026

Copilot started reviewing on behalf of youge325 April 1, 2026 11:18 View session

Copilot AI reviewed Apr 1, 2026

View reviewed changes

ShigureNyako suggested changes Apr 2, 2026

View reviewed changes

revert deleted stream APIs and fix Event

ce69b67

ShigureNyako suggested changes Apr 2, 2026

View reviewed changes

paddle/phi/api/include/compat/c10/core/Event.h Outdated Show resolved Hide resolved

complement macro

852d525

ShigureNyako approved these changes Apr 2, 2026

View reviewed changes

SigureMo added the skip-ci: coverage label Apr 2, 2026

SigureMo approved these changes Apr 2, 2026

View reviewed changes

SigureMo merged commit 0a5a78c into PaddlePaddle:develop Apr 2, 2026
148 of 157 checks passed

youge325 deleted the align-event-api branch April 2, 2026 13:31

ShigureNyako mentioned this pull request Apr 3, 2026

[Cpp API Compatibility] Align some compat APIs #78484

Closed

	#include <c10/cuda/CUDAStream.h>
	#include <c10/cuda/CUDAStream.h>
	#include <mutex>

Conversation

youge325 commented Apr 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR Category

PR Types

Description

变更详情

对齐效果

相关文档

是否引起精度变化

Uh oh!

paddle-bot bot commented Apr 1, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Copilot AI Apr 1, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 1, 2026

Choose a reason for hiding this comment

Uh oh!

ShigureNyako left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

ShigureNyako Apr 2, 2026

Choose a reason for hiding this comment

Uh oh!

ShigureNyako left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

youge325 commented Apr 2, 2026

Uh oh!

codecov-commenter commented Apr 2, 2026

Codecov Report

Uh oh!

ShigureNyako left a comment

Choose a reason for hiding this comment

Uh oh!

SigureMo commented Apr 2, 2026

Uh oh!

SigureMo left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

youge325 commented Apr 1, 2026 •

edited

Loading