[Cpp API Compatibility] Complement ScalarType by youge325 · Pull Request #78581 · PaddlePaddle/Paddle

youge325 · 2026-04-03T12:59:41Z

PR Category

Execute Infrastructure

PR Types

Improvements

Description

完善 ScalarType 类型，提供占位符，以对齐枚举类型

是否引起精度变化

否

paddle-bot · 2026-04-03T12:59:49Z

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 enhances the C10 compatibility layer’s ScalarType support by adding missing placeholder scalar wrapper types (quantized, bit-packed, and additional float formats) and wiring them into the scalar-type dispatch/mapping macros to better align with upstream enum coverage.

Changes:

Add placeholder c10 scalar wrapper headers for qint8/qint32, quint8/quint4x2/quint2x4, bits*, and several Float* low-precision formats.
Extend AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS to include these types (and ComplexHalf) so toString, elementSize, and type mapping specializations are generated consistently from the macro.
Remove now-redundant manual toString/elementSize switch cases and scalar-type constants that are covered by the macro generation.

Reviewed changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
paddle/phi/api/include/compat/c10/util/quint8.h	Adds placeholder `c10::quint8` wrapper type and exports to `at`/`torch`.
paddle/phi/api/include/compat/c10/util/quint4x2.h	Adds placeholder `c10::quint4x2` wrapper type and exports to `at`/`torch`.
paddle/phi/api/include/compat/c10/util/quint2x4.h	Adds placeholder `c10::quint2x4` wrapper type and exports to `at`/`torch`.
paddle/phi/api/include/compat/c10/util/qint8.h	Adds placeholder `c10::qint8` wrapper type and exports to `at`/`torch`.
paddle/phi/api/include/compat/c10/util/qint32.h	Adds placeholder `c10::qint32` wrapper type and exports to `at`/`torch`.
paddle/phi/api/include/compat/c10/util/Float8_e8m0fnu.h	Adds placeholder `c10::Float8_e8m0fnu` wrapper type and exports to `at`/`torch`.
paddle/phi/api/include/compat/c10/util/Float8_e5m2fnuz.h	Adds placeholder `c10::Float8_e5m2fnuz` wrapper type and exports to `at`/`torch`.
paddle/phi/api/include/compat/c10/util/Float8_e4m3fnuz.h	Adds placeholder `c10::Float8_e4m3fnuz` wrapper type and exports to `at`/`torch`.
paddle/phi/api/include/compat/c10/util/Float4_e2m1fn_x2.h	Adds placeholder `c10::Float4_e2m1fn_x2` wrapper type and exports to `at`/`torch`.
paddle/phi/api/include/compat/c10/util/bits.h	Adds placeholder `c10::bits*` wrapper types and exports to `at`/`torch`.
paddle/phi/api/include/compat/c10/core/ScalarType.h	Updates scalar-type macro lists and derives `toString`/`elementSize`/type mappings from the expanded macro coverage.

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

youge325 · 2026-04-04T02:36:23Z

/re-run all-failed

ShigureNyako

这版把 ScalarType 的枚举覆盖、toString/elementSize 宏化整理到一起，方向是对的；我也看了当前唯一失败的 IXUCA，日志里是 Paddle-iluvatar patch apply 失败，看起来和这次 ScalarType 改动没有直接关系。

不过我这边先不 approve，主要是 public C++ surface 还有两处没完全对齐 upstream：

quantized wrapper 目前缺少 upstream 的 using underlying = ...，下游如果走量化 dispatch，source level 还是会断。
Float4_e2m1fn_x2 的公开形状也和 upstream 不一致，至少字段名与比较运算还没补齐。

这两个点都属于兼容层的“半迁移”问题。建议先把我行内评论里的差异补齐，再继续推进；否则会出现 ScalarType/类型名已经暴露出来了，但真实 C++ API 还不能按 PyTorch 方式使用的情况。

ShigureNyako · 2026-04-04T05:33:02Z

paddle/phi/api/include/compat/c10/util/qint8.h

+
+namespace c10 {
+
+struct qint8 {


这里建议把 quantized wrapper 的 public surface 再向 upstream 对齐一下。PyTorch 的 torch/headeronly/util/qint8.h / qint32.h / quint8.h / quint4x2.h / quint2x4.h 都定义了 using underlying = ...，而且 AT_DISPATCH_CASE_QINT 这类宏会直接依赖 typename scalar_t::underlying。

当前实现只有 val_，下游如果拿这些类型走 quantized dispatch，source level 还是会再次卡住。建议至少把 underlying 别名补上；如果想完全贴近 upstream，alignas(...) 也最好一起保持一致。

建议把 quantized wrapper 的 public surface 向 upstream 对齐，补上 using underlying = ...，并保持 alignas(...) 一致。

已修复。最新 commit 00f14cbd55d2 已在 qint/quint wrapper 中补齐 using underlying = ...，并把 alignas(...) 调整为与 upstream 一致；我复查后认为这个兼容性问题已经消除。

ShigureNyako · 2026-04-04T05:34:00Z

paddle/phi/api/include/compat/c10/util/Float4_e2m1fn_x2.h

+
+namespace c10 {
+
+struct Float4_e2m1fn_x2 {


这个占位类型最好也直接复刻 upstream 的最小 public surface。PyTorch 的 torch/headeronly/util/Float4_e2m1fn_x2.h 使用的是 uint8_t val_，并提供了 operator==/!=。

这里改成 x 且没有比较运算，会让依赖这些公开成员/比较接口的代码继续不兼容。既然这个 PR 的目标是 C++ API compatibility，建议至少把字段名和比较运算补齐到 upstream 的形状。

建议复刻 upstream 的最小 public surface：字段名和 operator==/!= 需要补齐。

已修复。最新 commit 00f14cbd55d2 已把 Float4_e2m1fn_x2 调整为 upstream 一致的 val_ 公开形状，并补齐 operator==/!=；我复查后认为这个 source-level compatibility 问题已经解决。

ShigureNyako

我复查了作者新增的 commit 00f14cbd55d2。

此前提出的两个阻塞点都已被回应：

qint/quint wrapper 已补齐 using underlying = ...，并把 alignas(...) 调整为与 upstream 一致。
Float4_e2m1fn_x2 已改成 upstream 一致的 val_ public surface，并补齐 operator==/!=。

基于这次增量修改，我这边没有再看到新的 ABI / 调用方式 / 额外 dispatch 风险，代码 review 先通过；剩余 CI 按流程继续跑即可。

SigureMo · 2026-04-04T10:18:23Z

paddle/phi/api/include/compat/c10/util/bits.h

+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.


这些是不是得加 PyTorch 参考声明

好的，我加下

这些是不是得加 PyTorch 参考声明

已修复。我复查了作者最新 commit e75db41d17f0，现在已在 bits.h 以及这一批新增 compat wrapper / ScalarType.h 中补上 PyTorch 参考声明。这个增量只涉及声明注释补齐，不改变前面已经确认过的类型布局、ABI 或调用语义。

ShigureNyako

LGTMeow

我复查了作者新增的 commit
e75db41。

这次变更主要是响应 @SigureMo 关于“补充 PyTorch 参考声明”的意见：

已在、各个新增 compat wrapper 以及中补上 adapted-from-PyTorch 的声明。
本次增量不涉及类型布局、枚举值、dispatch 行为或调用方式变化。

因此我这边对前一轮的兼容性结论保持不变：当前实现没有新增 ABI / 调用方式 / 额外 dispatch 风险，review 通过；剩余 CI 继续按流程检查即可。

SigureMo · 2026-04-04T13:26:40Z

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

+// The file has been adapted from pytorch project
+// Licensed under BSD-style license -
 // https://github.com/pytorch/pytorch/blob/main/LICENSE


@ShigureNyako 单独提一个 PR 把 repo 里所有这样的声明都参考这里的改动改一下吧

@ShigureNyako 单独提一个 PR 把 repo 里所有这样的声明都参考这里的改动改一下吧

收到。我会把这类 adapted from PyTorch 参考声明的补齐整理成单独 PR 跟进，避免把 repo-wide 清理和这次 ScalarType 兼容性改动混在一起。当前这个 PR 我这边保持通过，后续会单独开任务处理。

@ShigureNyako 单独提一个 PR 把 repo 里所有这样的声明都参考这里的改动改一下吧

已处理，单独提了 repo-wide 清理 PR #78590，按这次 ScalarType.h / 新增 compat wrapper 的声明格式统一了现有 malformed adapted-from-PyTorch reference blocks。

这个跟进 PR 只改参考声明注释，不涉及 ABI、dispatch 行为或调用方式变化。

[Cpp API Compatibility] Complement ScalarType

14f01e4

Copilot AI review requested due to automatic review settings April 3, 2026 12:59

paddle-bot bot added the contributor External developers label Apr 3, 2026

Copilot started reviewing on behalf of youge325 April 3, 2026 13:00 View session

Copilot AI reviewed Apr 3, 2026

View reviewed changes

ShigureNyako suggested changes Apr 4, 2026

View reviewed changes

fix

00f14cb

ShigureNyako approved these changes Apr 4, 2026

View reviewed changes

SigureMo reviewed Apr 4, 2026

View reviewed changes

fix declaration

e75db41

ShigureNyako approved these changes Apr 4, 2026

View reviewed changes

SigureMo approved these changes Apr 4, 2026

View reviewed changes

SigureMo reviewed Apr 4, 2026

View reviewed changes

ShigureNyako mentioned this pull request Apr 4, 2026

[Cpp API Compatibility] Normalize adapted-from-PyTorch declarations #78590

Merged

SigureMo merged commit d3ea546 into PaddlePaddle:develop Apr 4, 2026
82 of 83 checks passed

youge325 deleted the cScalarType branch April 4, 2026 15:35

liuhao2638 pushed a commit to liuhao2638/Paddle that referenced this pull request Apr 7, 2026

[Cpp API Compatibility] Complement ScalarType (PaddlePaddle#78581)

1aa477f


		namespace c10 {

		struct qint8 {


		namespace c10 {

		struct Float4_e2m1fn_x2 {

Conversation

youge325 commented Apr 3, 2026

PR Category

PR Types

Description

是否引起精度变化

Uh oh!

paddle-bot bot commented Apr 3, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

youge325 commented Apr 4, 2026

Uh oh!

ShigureNyako left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ShigureNyako left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ShigureNyako left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants