Skip to content

[WIP][API-Compat] Add paddle.compat.min/max and new PHI kernel (min/max_with_index) #74512

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Closed
wants to merge 40 commits into from
Closed

[WIP][API-Compat] Add paddle.compat.min/max and new PHI kernel (min/max_with_index) #74512

wants to merge 40 commits into from

Conversation

Enigmatisms
Copy link
Contributor

@Enigmatisms Enigmatisms commented Aug 9, 2025 •
edited
Loading

PR Category

Operator Mechanism

PR Types

New features

Description

本 PR 为 #74495 的 reopen 版本，rebase 到了一个更新的版本，并且解决了与 #74506 的冲突。#74495 当时只用于CI发现问题，本 PR 尝试对其中“共用amin/amax backward op 但amin/amax 不支持某些整数类型“的问题进行了修复，基于SFINAE与python端检查。目前本 PR 在 #74506 未合入前会显得改动过多，实际上是包含了部分前序 PR 的改动，前序 PR merge 后应该可以自动 resolve。

本 PR 尚未完成：缺少对应的单测（进行了测试，见最后的TODO），并且依赖一个前置 PR（前置PR目前没有合入，合入后本PR信息将会修改）： #74446，需要其中的 ForbidKeywordsDecorator 装饰器。

本 PR 新增的 feature:

  • 两个新的 PHI kernel：(min/max)_with_index，底层基于 cub 的 Argmin/Argmax 操作，同时 reduce key/value。此 PHI kernel 仅在 GPU 端进行了支持。评估：
    • CPU 端的 argmin/argmax 操作基于 Eigen::TensorMap 提供的 argmin/argmax 操作，无法对 value 同时进行 reduce。
    • 其他后端暂不考虑支持。
  • 新的 PHI kernel 的反向传播算子：(min/max)_with_index_grad。注意⚠️：此 backward 行为与 torch documentation 不一致！原因：torch 的文档错了，见我给 PyTorch 提的这个 issue：Torch Issue 160273，其新闻根本不是 amin/amax 的行为：只对minimum/maximum index位置的结果传梯度，像是对梯度进行了一个 take_along_axis。
  • 其他与新算子相关的支持（如InferMeta/InferMetaSymbolic）
  • 两个新增的操作：paddle.compat.min, paddle.compat.max，与 torch 的行为进行对齐。

torch.min/torch.max 输入输出关系很复杂（一个API包含了太多功能）：

  • 单输入时是全 reduce：仅输出 tensor
  • 输入 dim/keepdim 时：输出 values/indices
  • 输入 other 时：行为与 minimum/maximum 一致

除上述【情况2】在 CUDA GPU 后端下会调用 (min/max)_with_index，其余情况都是由 python 调用 _C_ops.xxx 获得结果的。其中情况1/2/3 在CUDA GPU后端下应该都具有较好的性能（没有进行组合，调用单算子完成），而情况2在其他后端下使用 argmin/max 与 take_along_axis 组合（并且需要配合 squeeze_ 操作），不是最优性能方案，但应当具有较高的开发性价比。

TODO
  • 本地 25个简单单测（单算子不同的输入组合）
  • PaConvert 库测试（10/14），4个失败的测试由于 out 机制没有进行开发导致。此问题暂时不由本人解决。
  • PHI kernel 相关的单测（比如静态图 OpPass 单测等等）
  • test_compat_minmax.py: 能达到单测覆盖率要求的算子单测。
  • 算子python函数内英文文档

Pcard-89620

@paddle-bot Paddle-bot
Copy link

paddle-bot bot commented Aug 9, 2025

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.

@Enigmatisms
Copy link
Contributor Author

/re-run all-failed

@Enigmatisms
[API-Compat] paddle.compat.split is added and tested
255f8f1
@Enigmatisms
[API-Compat] paddle.compat.split is rigorously tested
95aedbe
@Enigmatisms
[API-Compat] Fixed erroneous func help doc
6166b8d
@Enigmatisms
[API-Compat] Make the forbid_keywords decorator transparent
41a0775
@Enigmatisms
[API-Compat] Fixed decorator str input
c8930f8
@Enigmatisms
[API-Compat] Fixed type annotation and removed legacy graph branch
d66c5f1
@Enigmatisms
[API-Compat] More unittest & static graph check & updated decorator
b40feed
@Enigmatisms
[API-Compat] Force update (local and not reproduce the bug)
4eb4925
@Enigmatisms
[API-Compat] Removed unittest that paddle.split will also fail
48f5bb0
@Enigmatisms
[API-Compat] Add paddle.compat.min/max and new PHI kernel (min/max_wi…
933c7c0 
…th_index)
@Enigmatisms
[API-Compat] Add compat.min/max EN doc
93d7c86 
Attempting to fix integral type gradient computation (rejection)
@Enigmatisms
[WIP][API-Compat] Add dyna-graph unittests for min/max
8ebe825
@Enigmatisms
[WIP][API-Compat] Fixed CPU failure
3d42943
@Enigmatisms
[API-Compat] Correct min/max_with index gradient behavior
ff38ddb
@Enigmatisms
[API-Compat] XPU fix (attempt)
09aeb0d
Enigmatisms and others added 11 commits August 11, 2025 02:35
@Enigmatisms
[API-Compat] Updated ForbidKeywordsDecorator
2f77d94
@co63oc @Enigmatisms
Rename ctx to dev_ctx in paddle/phi/kernels/ [fluid_ops] (PaddlePaddl…
e05a82d 
…e#74479)
@ooooo-create @Enigmatisms
refine some error message to avoid linking words together part7 (Padd…
5ed7519 
…lePaddle#74519)
@ooooo-create @Enigmatisms
refine some error message to avoid linking words together part6 (Padd…
73b235f 
…lePaddle#74520)
@zty-king @Enigmatisms
[AutoParallel] fix the grad_clip logic of auto_hybrid_pp (PaddlePaddl…
580d4a9 
…e#74409)

* fix the grad clip performance

* add test

* empty commit to rerun CI

* modify the note

* Simplify code logic
@co63oc @Enigmatisms
test/cpp rename use_mkldnn (PaddlePaddle#74501)
9076968
@co63oc @Enigmatisms
test/ directory modify use_mkldnn [fluid_ops] - part (PaddlePaddle#74487
fc98858 
)

* Fix

* Fix

* Fix
@co63oc @Enigmatisms
test/deprecated/cpp modify use_mkldnn [fluid_ops] (PaddlePaddle#74502)
e237e01
@co63oc @Enigmatisms
rename ctx to dev_ctx,xpu_ctx (PaddlePaddle#74513)
4a9c975
@co63oc @Enigmatisms
is_test_pass_tester.cc modify use_mkldnn [fluid_ops] (PaddlePaddle#74518
6b71a4b 
)
@co63oc @Enigmatisms
create_inference_config modify use_mkldnn [fluid_ops] (PaddlePaddle#7…
f962b15 
…4516)

* Fix

* Fix

* Fix
ooooo-create and others added 14 commits August 11, 2025 12:57
@ooooo-create @Enigmatisms
fix typos (PaddlePaddle#74497)
157bbb2
@zhupengyang @Enigmatisms
[xpu] fix compile (PaddlePaddle#74492)
c55f40b
@YuhanXu @Enigmatisms
CINN Add more Simplify scenario: Select2MinMax, BoundSimplify, PowerL…
026dd5f 
…og2BitOp. (PaddlePaddle#74292)
@wanghuancoder @Enigmatisms
dygraph support input a out Tensor (PaddlePaddle#74484)
c8f5638 
* dygraph support input a out Tensor

* refine

* refine

* refine

* refine

* refine

* refine

* refine

* refine
@enkilee @Enigmatisms
[Fix Issue] paddle.unique Exhibits Inconsistent Behavior with NaN Val…
07368c2 
…ues Between CPU and GPU (PaddlePaddle#74222)

* fix issue 73692

* fix error
@wangyuwen1999 @Enigmatisms
correct copysign backward (PaddlePaddle#74322)
cd23a02 
* correct copysign backward

* correct codestyle
@A-nnonymous @Enigmatisms
Enhancing fused_transpose_split_quant with fp8 capability. (PaddlePad…
c91fe43 
…dle#74471)

* Enhanced fused_transpose_split_quant with fp8 capability.

* optimize performance.

* Clean comment

* clean miscs

* Fix example
@co63oc @Enigmatisms
test_mkldnn_matmul_v2_elementwise_add_fuse_pass.py modify use_mkldnn …
fbd9f59 
…[fluid_ops] (PaddlePaddle#74507)

* Fix

* Fix
@zhwesky2010 @Enigmatisms
some create api support more usage (PaddlePaddle#74494)
17b20dd
@Enigmatisms
[API-Compat] More unittest & static graph check & updated decorator
761cd99
@Enigmatisms
[API-Compat] Updated ForbidKeywordsDecorator
bf0b293
@Enigmatisms
[API-Compat] Static Graph and CPU end debug
1514511
@Enigmatisms
[API-Compat] Revert erroneous rebase
c55dc29
@Enigmatisms
[API-Compat] Removed one split unittest, since the former PR is not m…
9f7d036 
…erged.
@Enigmatisms Enigmatisms deleted the min_max_api branch August 29, 2025 05:04
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@SigureMo SigureMo Awaiting requested review from SigureMo SigureMo is a code owner

@DrRyanHuang DrRyanHuang Awaiting requested review from DrRyanHuang DrRyanHuang is a code owner

@zrr1999 zrr1999 Awaiting requested review from zrr1999 zrr1999 is a code owner

@gouzil gouzil Awaiting requested review from gouzil gouzil is a code owner

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.