[API compatibility] add new API paddle.tensor and save paddle.tensor module #74540

zeroRains · 2025-08-11T10:15:20Z

PR Category

User Experience

PR Types

New features

Description

维持paddle.tensor作为module的原本功能的同时，新增API，paddle.tensor()实现tensor的创建。

在PaConvert下自测，paddle.tensor()与torch.tensor()对齐。

pcard-71500

paddle-bot · 2025-08-11T10:15:26Z

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.

zhwesky2010 · 2025-08-11T12:05:45Z

tools/windows/run_unittests.sh

@@ -335,6 +335,7 @@ disable_win_inference_test="^trt_quant_int8_yolov3_r50_test$|\
 ^trt_split_converter_test$|\
 ^paddle_infer_api_copy_tensor_tester$|\
 ^test_eager_tensor$|\
+^test_new_eager_tensor$|\


CI这些应该不用改吧

zhwesky2010 · 2025-08-11T12:11:06Z

python/paddle/tensor/creation.py

+                _warned_in_to_tensor = True
+        tensor = _to_tensor_non_static(data, dtype, place, stop_gradient)
+        if pin_memory:
+            tensor = tensor.pin_memory()


最终仅返回一次tensor，将pin_memory这些统一处理

zhwesky2010 · 2025-08-11T12:12:01Z

python/paddle/tensor/creation.py

@@ -957,6 +1075,9 @@ def to_tensor(
            [[(1+1j), (2+0j)],
             [(3+2j), (4+0j)]])
    """
+    warnings.warn(


参考下_warned_in_to_tensor，这里仅第一次打印warning，避免刷屏

zhwesky2010 · 2025-08-11T12:20:23Z

python/paddle/__init__.py

+
+
+tensor = _TensorMethodOrModule()  # noqa: F811
+
 # CINN has to set a flag to include a lib


_TensorMethodOrModule 这个class的其他一些方法也需要实现下，结合python原生方法，使体验上更好。比如type、print、repr应该看起来是一个方法，而dir() 看起来是模块

zhwesky2010 · 2025-08-11T12:24:24Z

test/legacy_test/test_zero_dim_sundry_dygraph_api.py

@@ -2061,6 +2061,20 @@ def test_to_tensor(self):
        self.assertEqual(out2.shape, [])
        self.assertEqual(out2, 2.5)

+    def test_tensor(self):


这里测试位置不对，建议这样测试吧，to_tensor直接alias到tensor，仅在前面加一个warning

这样历史所有的 paddle.to_tensor 都切换到 paddle.tensor

是指在to_tensor的方法中删除所有的实现，然后改成warning+调用paddle.tensor(xxx)的方式么？

诶，这样修改的话，是不是就不用新增单测了，直接利用to_tensor的单测测试tensor

zhwesky2010 · 2025-08-11T12:25:10Z

python/paddle/__init__.py

+
+
+tensor = _TensorMethodOrModule()  # noqa: F811
+
 # CINN has to set a flag to include a lib


new_tensor_api -> tensor_api

codecov-commenter · 2025-08-12T04:46:15Z

Codecov Report

❌ Patch coverage is 88.00000% with 6 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@d434c9e). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
python/paddle/tensor/creation.py	81.81%	6 Missing ⚠️

❌ Your patch status has failed because the patch coverage (88.00%) 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   #74540   +/-   ##
==========================================
  Coverage           ?   88.00%           
==========================================
  Files              ?        2           
  Lines              ?       50           
  Branches           ?        0           
==========================================
  Hits               ?       44           
  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 · 2025-08-14T09:47:34Z

python/paddle/tensor/creation.py

@@ -65,6 +65,7 @@
 __all__ = []

 _warned_in_to_tensor = False
+_warned_in_use_to_tensor = False


warning可以优化下名字，

_warned_in_tensor = False _warned_in_to_tensor = False

zhwesky2010 · 2025-08-14T09:49:55Z

python/paddle/tensor/creation.py

+                tensor = tensor.pin_memory()
+            return tensor
+
+
 @ParamAliasDecorator({"place": ["device"]})


这个装饰器可以去掉了，这个是为了兼容 paddle.to_tensor 调用device的情况，但现在有了严格对齐的 paddle.tensor，这个装饰器就不需要了，直接用paddle.tensor 就行，单测如果挂了也一并修改过来

zhwesky2010 · 2025-08-14T09:51:05Z

python/paddle/tensor/creation.py

+        place_str = re.findall(re_exp, str(place))[0]
+        with paddle.static.device_guard(place_str):
+            tensor = _to_tensor_static(data, dtype, stop_gradient)
+            if pin_memory:


pin_memory直接的实现是不是能优化下，根据pin_memory，修改place为 CudaPinnedPlace，避免二次拷贝的成本

除了Cuda能pin_memory外，XPU也可以，所以这里多加了一些判断，另外是有个分支通过tensor_from_cuda_array_interface创建tensor，这个接口没有place参数，所以这里显式调用了一次pin_memory

… tensor

zhwesky2010 · 2025-08-14T11:55:23Z

python/paddle/tensor/creation.py

+    place = _get_paddle_place(device)
+    if place is None:
+        place = _current_expected_place_()
+    if pin_memory and not isinstance(


这里是not_instance，下面是isinstance？

paddle.tensor写单测测试一下吧，虽然 to_tensor会调用 tensor ，但这两个API还是有些不同，比如 pin_memory、device这些参数

这里not isinstance是想确定place是不是已经是PinnedPlace了，是的话就没必要继续判断了，下面的isinstance是确定当前的Place是GPU或XPU。

好的我补充一下单测

zhwesky2010

这个API benchmark挂了，可能是to_tensor加的warning装饰器影响到了其他API，你需要把框架其他API用到to_tensor的地方改成新的 tensor 试试能不能过

… tensor

zeroRains · 2025-08-16T07:33:34Z

这个API benchmark挂了，可能是to_tensor加的warning装饰器影响到了其他API，你需要把框架其他API用到to_tensor的地方改成新的 tensor 试试能不能过

这个改动是不是太大了（搜了一下大概涉及1409个文件，10152个项），替换除了把paddle.to_tensor改成paddle.tensor之外，因为paddle.tensor和paddle.to_tensor的参数名（place->device, stop_gradient->requires_grad）不一样，有的情况还要一并修改。现在paddle.to_tensor是直接访问相对地址到API，然后在调用同级目录下的tensor方法。但改成paddle.tensor之后还需要走那个转发类，我感觉两个性能差距应该差不多。 ~~我看了一下CI里会导致下降的API不多，可以先对会导致性能下降的API进行替换，看看是否能解决。~~

以下为pr74540引入之后，api调度性能相对于baseline的变化。worse表示性能下降超过30%的api，doubt表示性能下降为15%~30%之间的api
{'worse': [{'paddle.chunk': -1.343596238482728}], 'doubt': [{'paddle.arange': -1.1972296709668069}, {'paddle.fft.rfftfreq': -1.1738153512819482}]}

这几个API也没用到paddle.to_tensor啊，这合理吗？

SigureMo · 2025-08-16T17:53:05Z

我看大多数 PR 的 API benchmark 里 paddle.arange 和 paddle.fft.rfftfreq 基本都是挂的，大概率是别的 PR 导致的

zhwesky2010 · 2025-08-18T07:04:09Z

python/paddle/tensor/creation.py

+            return tensor
+
+
+@deprecated(


如果影响到了性能，这个装饰器也可以不加，长时间应该还会是 paddle.tensor与paddle.to_tensor并存

这个性能影响应该不大，现在API Benchmark的CI更新后API Benchmark是pass的

… tensor

zhwesky2010

LGTM

zeroRains · 2025-08-20T02:49:57Z

/re-run all-failed

…module (PaddlePaddle#74540) * use paddle.tensor to instead paddle.to_tensor * update pin_memory * fix bug * fix the test * add a unit test * update test * add deprecated * update target api * add test * remove the deprecated * update test case * remove deprecated

zeroRains requested review from zhwesky2010, wanghuancoder, SigureMo, DrRyanHuang, zrr1999 and gouzil as code owners August 11, 2025 10:15

zhwesky2010 added the API Compatibility label Aug 11, 2025

zhwesky2010 reviewed Aug 11, 2025

View reviewed changes

use paddle.tensor to instead paddle.to_tensor

ec05c8e

zeroRains force-pushed the tensor branch 2 times, most recently from aa0b964 to ec05c8e Compare August 11, 2025 15:47

zhwesky2010 reviewed Aug 14, 2025

View reviewed changes

zeroRains added 2 commits August 14, 2025 19:46

update pin_memory

2a46d38

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

718ef0a

… tensor

zeroRains force-pushed the tensor branch from 1f90301 to 718ef0a Compare August 14, 2025 11:48

fix bug

97d3365

zhwesky2010 reviewed Aug 14, 2025

View reviewed changes

zeroRains added 5 commits August 14, 2025 20:13

fix the test

3efaab1

add a unit test

ab25099

update test

903e758

add deprecated

2c5405c

update target api

f53fbf4

zhwesky2010 reviewed Aug 15, 2025

View reviewed changes

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

16f7ef0

… tensor

zeroRains force-pushed the tensor branch from 9fc3cf9 to 16f7ef0 Compare August 16, 2025 07:24

zhwesky2010 reviewed Aug 18, 2025

View reviewed changes

zeroRains added 5 commits August 18, 2025 20:33

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

df58a7d

… tensor

add test

97f78c1

remove the deprecated

8e8c1e6

update test case

98fab20

remove deprecated

78f13af

zeroRains force-pushed the tensor branch from de8e492 to 78f13af Compare August 19, 2025 05:17

zhwesky2010 approved these changes Aug 19, 2025

View reviewed changes

SigureMo approved these changes Aug 19, 2025

View reviewed changes

tianshuo78520a added the skip-ci: coverage label Aug 19, 2025

zhwesky2010 merged commit 4e77392 into PaddlePaddle:develop Aug 20, 2025
126 of 134 checks passed

zeroRains deleted the tensor branch August 20, 2025 05:32



		tensor = _TensorMethodOrModule() # noqa: F811

		# CINN has to set a flag to include a lib

[API compatibility] add new API paddle.tensor and save paddle.tensor module #74540

[API compatibility] add new API paddle.tensor and save paddle.tensor module #74540

Conversation

zeroRains commented Aug 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR Category

PR Types

Description

Uh oh!

paddle-bot bot commented Aug 11, 2025

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!

zeroRains Aug 11, 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!

codecov-commenter commented Aug 12, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

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!

zeroRains Aug 14, 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!

zhwesky2010 left a comment • 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!

zeroRains commented Aug 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

SigureMo commented Aug 16, 2025

Uh oh!

zhwesky2010 Aug 18, 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!

zhwesky2010 left a comment

Choose a reason for hiding this comment

Uh oh!

zeroRains commented Aug 20, 2025

zeroRains commented Aug 11, 2025 •

edited

Loading

zeroRains Aug 11, 2025 •

edited

Loading

codecov-commenter commented Aug 12, 2025 •

edited

Loading

zeroRains Aug 14, 2025 •

edited

Loading

zhwesky2010 left a comment •

edited

Loading

zeroRains commented Aug 16, 2025 •

edited

Loading

zhwesky2010 Aug 18, 2025 •

edited

Loading