[API Compatiblity] add compat.softmax to compat with torch.softmax #74874

cyy536 · 2025-08-25T10:57:03Z

PR Category

User Experience

PR Types

New features

Description

add compat.softmax to compat with torch.softmax

paddle-bot · 2025-08-25T10:57:08Z

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.

codecov-commenter · 2025-08-25T23:37:32Z

Codecov Report

❌ Patch coverage is 88.67925% with 6 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@2fd8a7e). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
python/paddle/utils/decorator_utils.py	70.00%	6 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop   #74874   +/-   ##
==========================================
  Coverage           ?   88.67%           
==========================================
  Files              ?        7           
  Lines              ?       53           
  Branches           ?        0           
==========================================
  Hits               ?       47           
  Misses             ?        6           
  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

这里为啥还分了这么多softmax。

只有两个softmax，compat.softmax将dim进行设置直接调用原来的softmax。

cyy536 · 2025-08-26T04:12:14Z

一共写了4个，它们类型不一样
paddle.Tensor.softmax==torch.Tensor.softmax的类型是 (int dim, torch.dtype dtype = None)
paddle.softmax==torch.softmax的类型是 (Tensor input, int dim, torch.dtype dtype = None, *, Tensor out = None)
paddle.compat==torch.nn.functional.softmax的类型是 (input, dim=None, _stacklevel=3, dtype=None)
paddle.nn.functional.softmax的类型是 (input, dim=None, _stacklevel=3, dtype=None)，但默认值不一样

…e.softmax

zhwesky2010 · 2025-08-26T10:01:55Z

python/paddle/tensor/softmax.py

+
+
+@softmax_param_ignore_alias
+def compat_softmax(


这个实现放到对应位置里去吧，tensor/compat.py 里。paddle.compat.softmax 可以直接调 paddle.softmax

zhwesky2010 · 2025-08-26T10:02:39Z

python/paddle/tensor/softmax.py

@@ -159,7 +164,156 @@ def softmax(
              [0.03205860, 0.08714432, 0.23688282, 0.64391426],
              [0.03205860, 0.08714432, 0.23688282, 0.64391426]]])
    """
+    return _softmax_impl(x, axis, dtype, name, out=out)


这个可以直接展开

zhwesky2010 · 2025-08-26T10:08:13Z

python/paddle/tensor/softmax.py

+
+@softmax_param_ignore_alias
+def compat_softmax(
+    x: Tensor,


compat下的签名风格就按 input/dim 来吧，与min/max/sort 这些保持一致。

zhwesky2010 · 2025-08-26T10:09:14Z

python/paddle/compat.py

@@ -19,5 +19,6 @@
    sort,
    split,
 )
+from .tensor.softmax import compat_softmax as softmax


去 .tensor.compat 下实现。

zhwesky2010 · 2025-08-26T10:16:51Z

python/paddle/tensor/softmax.py

+    return _softmax_impl(x, axis, dtype, name, out=out)
+
+
+@softmax_param_ignore_alias


这个能复用当前的一些基建吗，保持风格一致的warning和提示比较好。参考 paddle.compat.min 风格。

zhwesky2010 · 2025-08-27T06:42:36Z

python/paddle/tensor/__init__.py

@@ -494,6 +494,8 @@
 )
 from .to_string import set_printoptions  # noqa: F401

+__all__ = ['softmax']


这个不能放__all__里，这个要放到下面的tensor_method_func，才会bind到paddle.Tensor上

zhwesky2010 · 2025-08-27T06:43:21Z

python/paddle/tensor/compat.py

+    if in_dynamic_or_pir_mode():
+        outs_cast = input if dtype is None else _C_ops.cast(input, dtype)
+        return paddle.assign(_C_ops.softmax(outs_cast, dim), out)
+    else:


下面的这个老静态图的的分支不需要实现

zhwesky2010 · 2025-08-27T06:44:44Z

python/paddle/tensor/compat.py

        Size2,
    )


 from paddle import nn
+from paddle.base.data_feeder import check_dtype, check_variable_and_dtype
+from paddle.base.framework import convert_np_dtype_to_dtype_
+from paddle.base.layer_helper import LayerHelper


老静态图分支不需要实现，移除无需导入的东西

zhwesky2010 · 2025-08-27T06:46:17Z

python/paddle/tensor/compat.py

+def softmax(
+    input: Tensor,
+    dim: int | None = None,
+    _stacklevel: int = 3,


这个能对齐到torch.softmax吗，这个_stacklevel参数

你是不是继承下ForbidKeywordsDecorator，隐藏处理下_stacklevel，然后外面的签名还是按input, dim, dtype, *, out来

zhwesky2010 · 2025-08-27T06:48:51Z

python/paddle/tensor/compat.py

+        dtype = convert_np_dtype_to_dtype_(dtype)
+    if in_dynamic_or_pir_mode():
+        outs_cast = input if dtype is None else _C_ops.cast(input, dtype)
+        return paddle.assign(_C_ops.softmax(outs_cast, dim), out)


这个直接 _C_ops.softmax(outs_cast, dim, out=out)

zhwesky2010 · 2025-08-27T06:50:48Z

python/paddle/tensor/softmax.py

@@ -121,6 +121,7 @@ def softmax(
            :math:`axis + D` . Default is -1.
        dtype (str, optional): The data type of the output tensor, can be bfloat16, float16, float32, float64.
        name (str|None, optional): For details, please refer to :ref:`api_guide_Name`. Generally, no setting is required. Default: None.
+        out (Tensor, optional): The output Tensor.


这个能用 _C_ops.softmax(outs_cast, axis, out=out) 来实现吗

zhwesky2010 · 2025-08-27T13:47:41Z

python/paddle/tensor/__init__.py

@@ -31,6 +31,7 @@
    real,
    shape,
 )
+from .compat_softmax import softmax as softmax


这个得加到tensor_method_func，才会bind到paddle.Tensor上

tensor_method_func一直都有加

zhwesky2010 · 2025-08-27T13:51:54Z

python/paddle/utils/decorator_utils.py

@@ -403,6 +396,70 @@ def process(
        return args, kwargs


+class ForbidKeywordsIgnoreOneParamDecorator(DecoratorBase):


这个可以继承ForbidKeywordsDecorator，只抽出它的ignore逻辑，其他采用 super().init(*args, **kwargs) ，这样可以最大化复用代码。

cyy536 · 2025-08-28T03:06:49Z

/re-run all-failed

zhwesky2010

LGTM

XiaoguangHu01

LGTM

XiaoguangHu01

LGTM

cyy536 · 2025-08-28T08:08:31Z

/re-run coverage build

add compat.softmax to compat with torch.softmax

3385f1c

cyy536 and others added 5 commits August 25, 2025 18:59

Merge branch 'develop' into add_compat_softmax

35e25ef

fix pre-commit

080d63a

fix paddle.softmax,paddle.Tensor.softmax,add compat.softmax

4a008b1

fix import

a38f873

fix paddle.softmax(...,out)

a8533ed

zhwesky2010 added the API Compatibility label Aug 26, 2025

zhwesky2010 reviewed Aug 26, 2025

View reviewed changes

zhwesky2010 changed the title ~~add compat.softmax to compat with torch.softmax~~ [API Compatiblity] add compat.softmax to compat with torch.softmax Aug 26, 2025

paddle.compat.softmax==torch.nn.functional.softmax,left all use paddl…

53be5ed

…e.softmax

zhwesky2010 reviewed Aug 26, 2025

View reviewed changes

cyy536 added 3 commits August 26, 2025 19:09

change file path

7b3f4c9

fix test

f740870

assign in static graph has bug

9b3b8e1

zhwesky2010 reviewed Aug 27, 2025

View reviewed changes

cyy536 and others added 2 commits August 27, 2025 21:30

fix

41841c1

Merge branch 'develop' into add_compat_softmax

68af999

zhwesky2010 reviewed Aug 27, 2025

View reviewed changes

cyy536 added 2 commits August 27, 2025 22:17

fix inherit

c8e632f

fix pre-commit

c1e6436

zhwesky2010 approved these changes Aug 28, 2025

View reviewed changes

zhwesky2010 requested review from SigureMo and XiaoguangHu01 August 28, 2025 03:24

SigureMo approved these changes Aug 28, 2025

View reviewed changes

XiaoguangHu01 approved these changes Aug 28, 2025

View reviewed changes

risemeup1 added the skip-ci: coverage label Aug 28, 2025

swgu98 added the skip-ci: all label Aug 28, 2025

zhwesky2010 merged commit cd64b23 into PaddlePaddle:develop Aug 28, 2025
153 of 162 checks passed

		return _softmax_impl(x, axis, dtype, name, out=out)


		@softmax_param_ignore_alias

		@@ -403,6 +396,70 @@ def process(
		return args, kwargs


		class ForbidKeywordsIgnoreOneParamDecorator(DecoratorBase):

[API Compatiblity] add compat.softmax to compat with torch.softmax #74874

[API Compatiblity] add compat.softmax to compat with torch.softmax #74874

Uh oh!

Conversation

cyy536 commented Aug 25, 2025

PR Category

PR Types

Description

Uh oh!

paddle-bot bot commented Aug 25, 2025

Uh oh!

codecov-commenter commented Aug 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

zhwesky2010 left a comment

Choose a reason for hiding this comment

Uh oh!

cyy536 commented Aug 26, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

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!

zhwesky2010 Aug 26, 2025 • 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!

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!

cyy536 commented Aug 28, 2025

Uh oh!

zhwesky2010 left a comment

Choose a reason for hiding this comment

Uh oh!

XiaoguangHu01 left a comment

Choose a reason for hiding this comment

Uh oh!

XiaoguangHu01 left a comment

Choose a reason for hiding this comment

Uh oh!

cyy536 commented Aug 28, 2025

Uh oh!

Uh oh!

Uh oh!

codecov-commenter commented Aug 25, 2025 •

edited

Loading

cyy536 commented Aug 26, 2025 •

edited

Loading

zhwesky2010 Aug 26, 2025 •

edited

Loading