[API Compatibility] add aminmax op-part by Manfredss · Pull Request #78441 · PaddlePaddle/Paddle

Manfredss · 2026-03-23T20:00:44Z

PR Category

User Experience

PR Types

New features

Description

Add paddle.aminmax API that returns both minimum and maximum of tensor elements over the given axis, aligned with
torch.aminmax
Supports axis, keepdim parameters and PyTorch-compatible aliases (input for x, dim for axis)
Backward gradient correctly distributes evenly among duplicate min/max elements (same as amin/amax)

Changes

New files

paddle/phi/kernels/aminmax_kernel.h/.cc — forward kernel (reuses AMinRawKernel + AMaxRawKernel)
paddle/phi/kernels/aminmax_grad_kernel.h/.cc — backward kernel (amax_grad + amin_grad)
test/legacy_test/test_aminmax_op.py — unit tests

Modified files

paddle/phi/ops/yaml/ops.yaml / backward.yaml — op and grad definitions
paddle/phi/ops/yaml/python_api_info.yaml / op_compat.yaml — API info and compat
paddle/phi/infermeta/unary.h/.cc — AMinMaxInferMeta (two-output reduce shape)
.../infer_symbolic_shape/unary_infer_sym.h/.cc — AminmaxOpInferSymbolicShape
python/paddle/__init__.py, python/paddle/tensor/__init__.py — exports
python/paddle/_paddle_docs.py — docstring and signature
test/legacy_test/test_api_compatibility.py — API compatibility test

API Signature

paddle.aminmax(x, axis=None, keepdim=False, name=None) -> tuple[Tensor, Tensor]

Test plan

test_aminmax_op.py — 27 tests (axis variants, keepdim, dtypes, zero-dim, gradient)
test_api_compatibility.py::TestAminmaxAPI — dygraph + static, paddle/torch keyword aliases
Passed all 15 tests on enhanced PaConvert tests

是否引起精度变化

否

paddle-bot · 2026-03-23T20:00:50Z

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.

Manfredss · 2026-03-24T02:13:43Z

/re-run all-failed

Manfredss · 2026-03-24T04:40:27Z

/re-run all-failed

zhwesky2010 · 2026-03-24T04:27:37Z

+        aminmax evenly distributes gradient between these equal values.
+
+    Args:
+        x (Tensor): A tensor, the data type is float32, float64, int32, int64.


Alias写下

zhwesky2010 · 2026-03-24T04:35:49Z

    amin_grad : GetReduceGradExpectedKernelType
  manual_signature : [amin]

+- op : aminmax


这个应该是一些不兼容历史问题的适配，这个需要改吗，不报错可以不改

zhwesky2010 · 2026-03-24T04:40:31Z

+
+    Args:
+        x (Tensor): A tensor, the data type is float32, float64, int32, int64.
+        axis (int|list|tuple|None, optional): The axis along which the minimum and maximum


zhwesky2010 · 2026-03-24T04:40:38Z

+            than the `x` unless :attr:`keepdim` is true, default
+            value is False.
+        name (str|None, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.
+


关键字参数文档

zhwesky2010 · 2026-03-24T04:41:09Z

+class TestAminmaxAPI(unittest.TestCase):
+    def setUp(self):
+        self.init_case()
+        self.place = paddle.CPUPlace()


无需设置place，新增OP还是在每个设备上都测下较好

zhwesky2010 · 2026-03-24T04:42:10Z

+        paddle.disable_static()
+        x = paddle.to_tensor(self.x_np, dtype=self.dtype, stop_gradient=False)
+        min_val, max_val = paddle.aminmax(x, self.axis, self.keepdim)
+        loss = min_val.sum() + max_val.sum()


反向测下值吧

zhwesky2010 · 2026-03-24T04:42:48Z

+import paddle
+
+
+class TestAminmaxAPI(unittest.TestCase):


新增OP可以测全一些，可以继承OPTest基类来测下。

zhwesky2010 · 2026-03-24T04:45:22Z

 """,
 )

+add_doc_and_signature(


文档统一按这个格式规范来吧：

英文-python装饰器
https://github.com/PaddlePaddle/docs/blob/develop/docs/dev_guides/coding_agent/api_compatibility/python-decorator.md#step-4-%E6%9B%B4%E6%96%B0%E5%87%BD%E6%95%B0%E6%96%87%E6%A1%A3%E5%AD%97%E7%AC%A6%E4%B8%B2

英文-cpp下沉
https://github.com/PaddlePaddle/docs/blob/develop/docs/dev_guides/coding_agent/api_compatibility/cpp-sink.mdr#L66

中文
https://github.com/PaddlePaddle/docs/blob/develop/docs/dev_guides/coding_agent/api_compatibility/api-docs-updater.mdr

Manfredss · 2026-03-25T00:54:46Z

/re-run all-failed

codecov-commenter · 2026-03-25T00:57:42Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (develop@4174bb8). Learn more about missing BASE report.

Additional details and impacted files

@@             Coverage Diff             @@
##             develop    #78441   +/-   ##
===========================================
  Coverage           ?   100.00%           
===========================================
  Files              ?         5           
  Lines              ?        50           
  Branches           ?         0           
===========================================
  Hits               ?        50           
  Misses             ?         0           
  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.

zhwesky2010 · 2026-03-25T03:21:38Z

覆盖率没过

Manfredss · 2026-03-26T09:52:20Z

/re-run all-failed

zhwesky2010 · 2026-03-26T12:28:45Z

    amin_grad : GetReduceGradExpectedKernelType
  manual_signature : [amin]

+- op : aminmax


这个是否可以不改？

这里的定义用于映射旧框架大写命名（X、Min、Max）到新 phi 内核小写命名（x、min、max）。OpTest 类依赖此映射来构建算子，移除后 check_output 和 check_grad 会报 Missing x as input 或 out_dtype not found 错误，已通过验证不能去掉

zhwesky2010 · 2026-03-26T12:31:17Z

+        # 4. Mixed arguments (positional x + keyword axis)
+        min4, max4 = paddle.aminmax(x, axis=0)
+        # 5. PyTorch keyword arguments (dim alias)
+        min5, max5 = paddle.aminmax(x, dim=0)


测一下out参数

zhwesky2010 · 2026-03-26T12:31:53Z

+    x: Tensor,
+    axis: int | Sequence[int] | None = None,
+    keepdim: bool = False,
+    name: str | None = None,


默认是支持out为关键字参数的，所有op会默认解析kwargs: out

Manfredss · 2026-03-27T22:58:59Z

/re-run all-failed

Manfredss · 2026-03-29T21:28:13Z

/re-run all-failed

zhwesky2010 · 2026-03-31T12:38:40Z

+  outputs :
+    {min : Min, max : Max}
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_onednn = false]


这些extra应该是没有的

zhwesky2010 · 2026-03-31T12:39:39Z

+    x: Tensor,
+    axis: int | Sequence[int] | None = None,
+    keepdim: bool = False,
+    name: str | None = None,


新增API统一不支持name

zhwesky2010 · 2026-03-31T12:41:28Z

+    def init_shape(self):
+        self.shape = [10, 10]
+
+    def test_check_grad(self):


这个是什么问题

float32 精度下（约7位有效数字），0.005 的扰动会改变 min/max 落在哪个元素上，导致数值梯度和解析梯度完全不一致，最大误差达到 1.0。float64 精度足够，元素间的数值差距远大于扰动量，不会出现这个问题

zhwesky2010 · 2026-03-31T12:43:33Z

+            [[0.2, 0.3, 0.5, 0.9], [0.1, 0.2, 0.6, 0.7]], dtype='float64'
+        )
+
+    def test_dygraph_Compatibility(self):


这里按照测试规范来填条目吧，不用自由发挥：

https://github.com/PaddlePaddle/docs/blob/develop/docs/dev_guides/coding_agent/api_compatibility/python-decorator.md#step-5-%E6%B7%BB%E5%8A%A0%E6%B5%8B%E8%AF%95%E7%94%A8%E4%BE%8B

zhwesky2010 · 2026-03-31T12:44:19Z

+                np.testing.assert_allclose(fetches[i + 1], ref_max)
+
+
+if __name__ == '__main__':


还需要加一下 paddle.jit.to_static(backend='CINN') 的测试，覆盖AminmaxOpInferSymbolicShape的测试

…e extra in op_compat.yaml

Manfredss · 2026-04-01T03:18:26Z

/re-run all-failed

zhwesky2010

test_api_compatibility_part3.py的TestConv2dTransposeAPI有点问题，导致现在CI全部无法通过了，这个PR注释下case或者修复

zhwesky2010 · 2026-04-01T11:45:19Z

+        self.init_shape = [None, None, 6]
+        self.x = np.random.random(self.shape).astype(self.dtype)
+        self.net = aminmax_net
+        self.enable_cinn = False


你这没开cinn，导致还是没覆盖到

zhwesky2010 · 2026-04-01T11:47:00Z

+    return paddle.aminmax(x, axis=1)
+
+
+def apply_to_static(net, use_cinn, input_spec=None):


这一层建议不包了，包了太多层了

zhwesky2010 · 2026-04-01T11:47:32Z

+
+    def base_net(self, flag=None):
+        x = paddle.to_tensor(self.x)
+        if flag == "static":


if "static"

elif "dygraph"

…f branch

Manfredss · 2026-04-05T10:16:12Z

/re-run all-failed

Manfredss · 2026-04-05T20:35:04Z

/re-run all-failed

…file

Manfredss · 2026-04-07T10:06:34Z

/re-run all-failed

Manfredss · 2026-04-10T21:58:34Z

@zhwesky2010 加了打印信息在跑 pyhton test/legacy_test/test_aminmax_op.py 的时候会有打印，这是覆盖到了吧？

zhwesky2010 · 2026-04-17T03:38:49Z

@Manfredss 看起来是已经正常覆盖了

zhwesky2010

LGTM

zhwesky2010 · 2026-04-17T03:39:50Z

@Manfredss 示例代码还需要先关一下

…check in docs

Manfredss · 2026-04-19T05:24:09Z

/re-run all-failed

Manfredss · 2026-04-21T23:32:16Z

/re-run all-failed

zhwesky2010 · 2026-04-22T07:39:34Z

+            Tensor(shape=[4], dtype=float64, place=Place(cpu), stop_gradient=False,
+            [0.90000000, 0.90000000, 0.90000000, 0.90000000])
+
+            >>> min_val, max_val = paddle.aminmax(x, axis=1, keepdim=True)


这些示例代码跳过失效了，这个可能只是跳过对输出结果的检查。不跳过示例代码本身的执行？

… ApiEnhance_aminmax

Manfredss · 2026-04-22T19:43:50Z

/re-run all-failed

Manfredss · 2026-04-23T03:30:39Z

/re-run all-failed

zhwesky2010 · 2026-04-23T07:16:03Z

需要看一下CI结果，这个不是随机挂，rerun只针对随机挂才有用

Manfredss · 2026-04-23T20:00:38Z

/re-run all-failed

Manfredss · 2026-04-23T23:00:28Z

/re-run all-failed

zhwesky2010

LGTM

wanghuancoder

LGTM

develop paddle.aminmax to align paddle with torch.Tensor.aminmax

3fc934e

paddle-bot Bot added the contributor External developers label Mar 23, 2026

resolve conflict

bf2fd62

Manfredss mentioned this pull request Mar 23, 2026

[API Compatibility] enhance torch.aminmax tests -part PaddlePaddle/PaConvert#859

Open

zhwesky2010 added the API Compatibility label Mar 24, 2026

zhwesky2010 changed the title ~~[API Compatibility] develop paddle.aminmax to align paddle with torch.Tensor.aminmax -part~~ [API Compatibility] add aminmax to align with torch -part Mar 24, 2026

zhwesky2010 changed the title ~~[API Compatibility] add aminmax to align with torch -part~~ [API Compatibility] add aminmax -part Mar 24, 2026

zhwesky2010 reviewed Mar 24, 2026

View reviewed changes

Manfredss added 2 commits March 24, 2026 17:17

refine tests, improve en docs

e62826d

fix

e735b93

try fix coverage

d780164

zhwesky2010 reviewed Mar 26, 2026

View reviewed changes

add out test and add out in docs.

96b10be

Manfredss closed this Mar 29, 2026

Manfredss reopened this Mar 29, 2026

zhwesky2010 reviewed Mar 31, 2026

View reviewed changes

add CINN test, follow the test routine, remove name in docs and remov…

f2ae4ab

…e extra in op_compat.yaml

zhwesky2010 changed the title ~~[API Compatibility] add aminmax -part~~ [API Compatibility] add aminmax op-part Apr 1, 2026

zhwesky2010 reviewed Apr 1, 2026

View reviewed changes

enable CINN, remove the redundant wrapper, and restructure the if-eli…

04e7223

…f branch

add AminmaxOpInferSymbolicShape coverage test to CINN sym shape test …

9883a5c

…file

ooooo-create mentioned this pull request Apr 13, 2026

[CINN] Clean unused pool2d symbolic infer path #78658

Merged

Manfredss mentioned this pull request Apr 16, 2026

[WeeklyReports] 2026.3.23~2026.4.12 周报汇总 PFCCLab/Camp#566

Open

9 tasks

zhwesky2010 previously approved these changes Apr 17, 2026

View reviewed changes

enable all device to run test for aminmax, disabled the example code …

7a184e2

…check in docs

Manfredss dismissed zhwesky2010’s stale review via 7a184e2 April 17, 2026 21:38

zhwesky2010 reviewed Apr 22, 2026

View reviewed changes

Manfredss added 2 commits April 23, 2026 00:23

add SKIP per example code

f28ebcd

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

2d9488d

… ApiEnhance_aminmax

delete example code

9a11de4

zhwesky2010 approved these changes Apr 24, 2026

View reviewed changes

zhwesky2010 requested review from SigureMo and wanghuancoder April 24, 2026 08:16

wanghuancoder approved these changes Apr 24, 2026

View reviewed changes

zhwesky2010 requested review from SigureMo and removed request for SigureMo April 24, 2026 09:52

SigureMo approved these changes Apr 24, 2026

View reviewed changes

		np.testing.assert_allclose(fetches[i + 1], ref_max)


		if __name__ == '__main__':

		return paddle.aminmax(x, axis=1)


		def apply_to_static(net, use_cinn, input_spec=None):

		import paddle


		class TestAminmaxAPI(unittest.TestCase):

Conversation

Manfredss commented Mar 23, 2026

PR Category

PR Types

Description

Changes

New files

Modified files

API Signature

Test plan

是否引起精度变化

Uh oh!

paddle-bot Bot commented Mar 23, 2026

Uh oh!

Manfredss commented Mar 24, 2026

Uh oh!

Manfredss commented Mar 24, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

zhwesky2010 Mar 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

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!

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!

Manfredss commented Mar 25, 2026

Uh oh!

codecov-commenter commented Mar 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

zhwesky2010 commented Mar 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Manfredss commented Mar 26, 2026

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!

Manfredss commented Mar 27, 2026

Uh oh!

Manfredss commented Mar 29, 2026

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!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Manfredss commented Apr 1, 2026

Uh oh!

zhwesky2010 left a comment

Choose a reason for hiding this comment

Uh oh!

zhwesky2010 Apr 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

zhwesky2010 Mar 24, 2026 •

edited

Loading

codecov-commenter commented Mar 25, 2026 •

edited

Loading

zhwesky2010 commented Mar 25, 2026 •

edited

Loading

zhwesky2010 Apr 1, 2026 •

edited

Loading

Manfredss commented Apr 10, 2026 •

edited

Loading

zhwesky2010 commented Apr 23, 2026 •

edited

Loading