【UnitTestFix No.1】fix test_activation_op.py by scyyh11 · Pull Request #75553 · PaddlePaddle/Paddle

scyyh11 · 2025-09-26T10:56:39Z

PR Category

Operator Mechanism

PR Types

Bug fixes

Description

修复了test_activation_op.py中不可用的test。
删除了SoftRelu相关的测试，因为2.6版本后已经不支持SoftRelu。

@luotao1 @YqGe585

….gradient" is deprecated since 2.1.0, and will be removed in future versions. Reason: Please use tensor.grad, which returns the tensor value of the gradient."

…tPow_API, TestRelu6APIWarnings) with OldIrGuard plus fresh Program guards. Adjusted TestRelu_NanInput to convert the NaN-count tensor to a host scalar before asserting, sidestepping PIR’s static bool(Tensor) restriction.

- Fix shape comparison in TestSinhAPI and TestCoshAPI by converting shapes to lists - Disable gradient check for TestRelu_NanInput class to handle NaN input cases - Refactor TestSqrtOutAndAlias to use PIR-compatible API with positional arguments - Simplify test execution by removing unnecessary startup program call - Update variable naming and data feeding for better PIR support

- Enable int32 input support for sqrt, tanh, sinh, cosh ops with auto-cast to float32 - Fix shape comparison in TestTanAPI by converting shapes to lists - Refactor TestRelu_NanInput to support both static and dygraph execution modes - Update test comments to reflect new int32 input support capabilities

- Deleted the TestSoftRelu class to streamline activation operation tests. - Updated test creation calls to exclude TestSoftRelu for both FP16 and BF16 classes.

- Added a test_check_output method to override the base class behavior. - Refactored NaN count calculation to use numpy's isnan method for clarity.

- Set check_prim_pir to False in TestSigmoidBF16 and TestPow classes to improve compatibility with PIR. - Adjusted test configurations to ensure consistent behavior across activation operation tests.

The TestPow FP16 test was failing because it incorrectly expected the pow operation to be decomposed in PIR mode (check_prim_pir=True). However, pow is a primitive operation and should not be decomposed. Changed the configuration to check_prim_pir=False to match the primitive nature of the pow operation.

- Remove unnecessary comments and clean up code. - Adjusted assertions in TestPow_API for clarity and consistency.

SigureMo · 2025-09-26T12:30:00Z

            check_prim=False,
            check_pir=True,
-            check_prim_pir=True,
+            check_prim_pir=False,


check_prim 可以关，check_prim_pir 不能关呀，PIR 下有问题应该去修复

您好，我在修复这个部分的时候遇到了如下的问题：

在 primitive.yaml 和 primitive_ops.h 里，sigmoid 和 pow 被定义为 primitive 算子。

但在 composite_rules.py 里，它们又有明确的 分解规则（比如 sigmoid → 1/(1+exp(-x))）。

在测试时，虽然我调用了 decomposition pass，但这两个算子并不会真的被分解，操作列表保持不变，导致断言失败。

所以我现在的问题是，对于pow和sigmoid，是要把他们作为primitive 算子来对待么

组合算子的细节需要节后 @xiaoguoguo626807 确认下

这个理论上这两个算子是组合算子，但是因为拆解之后会引起精度或者性能问题，所以实际使用的时候不会拆解

你描述的问题是旧ir 下遇到的吗，提到的文件都是旧ir的

你描述的问题是旧ir 下遇到的吗，提到的文件都是旧ir的

我是在做旧IR到新IR的迁移，修复部分因为框架更新不能用的test

…ack to True for sigmoid and pow test(need fix).

- skip check_static_comp for prim operator in prim_op_test.py - refactor the prim_op_type for TestPow and TestSigmoidBF16 in test_activation_op.py since they are both primitive operators.

scyyh11 · 2025-10-09T04:46:16Z

@xiaoguoguo626807 您好，可以确认一下 sigmoid 和 pow 目前是否作为primitive算子？

xiaoguoguo626807 · 2025-10-10T02:32:22Z

            core.set_prim_eager_enabled(False)

    def check_static_comp(self):
+        # prim op don't need gradient decomposition


这里不对吧，prim_op 是需要测试反向拆解的

xiaoguoguo626807 · 2025-10-10T02:33:45Z

 class TestSigmoidBF16(OpTest):
    def setUp(self):
        self.op_type = "sigmoid"
-        self.prim_op_type = "comp"


这个算子比较特殊，可以测comp , 测试前向拆解正确性，也可以测prim, 测反向拆解正确性，严格来说都需要测一下

…rectness of sigmoid under FP32

scyyh11 · 2025-10-10T09:31:29Z

@xiaoguoguo626807 您好，我为sigmoid算子添加了TestSigmoidFp32_Comp用于测试前向拆解正确性。对于pow算子，我保留了self.prim_op_type = "prim"。请问这样的修改可以么？

xiaoguoguo626807 · 2025-10-11T03:00:30Z

@xiaoguoguo626807 您好，我为sigmoid算子添加了TestSigmoidFp32_Comp用于测试前向拆解正确性。对于pow算子，我保留了self.prim_op_type = "prim"。请问这样的修改可以么？

pow 也需要这样处理下，辛苦加一下吧

… and gradient checks for pow operation in FP64 precision

scyyh11 · 2025-10-11T07:00:40Z

@xiaoguoguo626807 您好，我为sigmoid算子添加了TestSigmoidFp32_Comp用于测试前向拆解正确性。对于pow算子，我保留了self.prim_op_type = "prim"。请问这样的修改可以么？

pow 也需要这样处理下，辛苦加一下吧

您好，现在已经加上了，麻烦您review一下

SigureMo · 2025-10-11T15:47:35Z

 )
 create_test_act_fp16_class(TestBRelu, check_pir=True)
 create_test_act_fp16_class(TestRelu6)
-create_test_act_fp16_class(TestSoftRelu, check_dygraph=False)


TestSoftRelu 删掉的理由是什么？3090313 提到 Deleted the TestSoftRelu class to streamline activation operation tests. 是说 TestSoftRelu 和别的有重复么？相关改动原因最好在 PR 描述里说明下

您好，删除SoftRelu相关测试的原因是在2.6版本之后SoftRelu在官方文档里已经搜不到相关内容，保留会遇到SoftRelu算子未注册的报错，我和@YqGe585 沟通确认后删除了相关的测试。

scyyh11 · 2025-10-11T18:00:07Z

/re-run all-failed

xiaoguoguo626807

LGTM for decomp test

YqGe585

LGTM

* fix: eliminate warning "API "paddle.base.dygraph.tensor_patch_methods.gradient" is deprecated since 2.1.0, and will be removed in future versions. Reason: Please use tensor.grad, which returns the tensor value of the gradient." * fix: skip unsupported integer gradient checks for ceil/floor prim tests * fix: Added paddle.pir_utils import and wrapped legacy-only tests (TestPow_API, TestRelu6APIWarnings) with OldIrGuard plus fresh Program guards. Adjusted TestRelu_NanInput to convert the NaN-count tensor to a host scalar before asserting, sidestepping PIR’s static bool(Tensor) restriction. * fix: improve activation tests for PIR compatibility and shape handling - Fix shape comparison in TestSinhAPI and TestCoshAPI by converting shapes to lists - Disable gradient check for TestRelu_NanInput class to handle NaN input cases - Refactor TestSqrtOutAndAlias to use PIR-compatible API with positional arguments - Simplify test execution by removing unnecessary startup program call - Update variable naming and data feeding for better PIR support * fix: improve activation op tests for type compatibility and PIR support - Enable int32 input support for sqrt, tanh, sinh, cosh ops with auto-cast to float32 - Fix shape comparison in TestTanAPI by converting shapes to lists - Refactor TestRelu_NanInput to support both static and dygraph execution modes - Update test comments to reflect new int32 input support capabilities * refactor: remove TestSoftRelu class from activation tests - Deleted the TestSoftRelu class to streamline activation operation tests. - Updated test creation calls to exclude TestSoftRelu for both FP16 and BF16 classes. * fix: update TestRelu_NanInput to prevent base class method call - Added a test_check_output method to override the base class behavior. - Refactored NaN count calculation to use numpy's isnan method for clarity. * fix: update activation op tests to disable check_prim_pir - Set check_prim_pir to False in TestSigmoidBF16 and TestPow classes to improve compatibility with PIR. - Adjusted test configurations to ensure consistent behavior across activation operation tests. * fix: correct TestPow FP16 prim checker configuration The TestPow FP16 test was failing because it incorrectly expected the pow operation to be decomposed in PIR mode (check_prim_pir=True). However, pow is a primitive operation and should not be decomposed. Changed the configuration to check_prim_pir=False to match the primitive nature of the pow operation. * fix: - Remove unnecessary comments and clean up code. - Adjusted assertions in TestPow_API for clarity and consistency. * refactor: optimize import * fix: remove OldIr related test case. revert modified check_prim_pir back to True for sigmoid and pow test(need fix). * fix: - skip check_static_comp for prim operator in prim_op_test.py - refactor the prim_op_type for TestPow and TestSigmoidBF16 in test_activation_op.py since they are both primitive operators. * fix: reset check_static_comp * fix: add new TestSigmoidFp32_Comp to verify forward decomposition correctness of sigmoid under FP32 * fix: add TestPowFp64_Comp to verify forward decomposition correctness and gradient checks for pow operation in FP64 precision

scyyh11 and others added 14 commits September 18, 2025 23:53

fix: eliminate warning "API "paddle.base.dygraph.tensor_patch_methods…

13c16e6

….gradient" is deprecated since 2.1.0, and will be removed in future versions. Reason: Please use tensor.grad, which returns the tensor value of the gradient."

fix: skip unsupported integer gradient checks for ceil/floor prim tests

241528f

refactor: remove TestSoftRelu class from activation tests

3090313

- Deleted the TestSoftRelu class to streamline activation operation tests. - Updated test creation calls to exclude TestSoftRelu for both FP16 and BF16 classes.

fix: update TestRelu_NanInput to prevent base class method call

d9fc668

- Added a test_check_output method to override the base class behavior. - Refactored NaN count calculation to use numpy's isnan method for clarity.

fix: update activation op tests to disable check_prim_pir

576672a

- Set check_prim_pir to False in TestSigmoidBF16 and TestPow classes to improve compatibility with PIR. - Adjusted test configurations to ensure consistent behavior across activation operation tests.

Merge branch 'PaddlePaddle:develop' into fix/test_activation_op

4c9f26c

Merge branch 'PaddlePaddle:develop' into fix/test_activation_op

163786e

Merge branch 'PaddlePaddle:develop' into fix/test_activation_op

73a3b3a

fix:

8fb63ab

- Remove unnecessary comments and clean up code. - Adjusted assertions in TestPow_API for clarity and consistency.

refactor: optimize import

f9de1fb

scyyh11 marked this pull request as ready for review September 26, 2025 10:59

YqGe585 suggested changes Sep 26, 2025

View reviewed changes

Comment thread test/legacy_test/test_activation_op.py Outdated

luotao1 mentioned this pull request Sep 26, 2025

【启航计划】PaddlePaddle GPU单测修复 #75208

Closed

YqGe585 requested a review from SigureMo September 26, 2025 12:23

SigureMo reviewed Sep 26, 2025

View reviewed changes

paddle-bot Bot added the contributor External developers label Sep 26, 2025

scyyh11 added 2 commits September 26, 2025 13:13

fix: remove OldIr related test case. revert modified check_prim_pir b…

e1eaef0

…ack to True for sigmoid and pow test(need fix).

fix:

76254fd

- skip check_static_comp for prim operator in prim_op_test.py - refactor the prim_op_type for TestPow and TestSigmoidBF16 in test_activation_op.py since they are both primitive operators.

scyyh11 requested review from SigureMo and YqGe585 September 26, 2025 23:51

luotao1 added the HappyOpenSource 快乐开源活动issue与PR label Sep 27, 2025

luotao1 assigned luotao1 and YqGe585 Sep 27, 2025

YqGe585 requested a review from xiaoguoguo626807 September 28, 2025 12:19

scyyh11 mentioned this pull request Oct 9, 2025

[WeeklyReports] 2025.09.29~2025.10.10 第一次周报收集 PFCCLab/Starter#672

Closed

42 tasks

xiaoguoguo626807 reviewed Oct 10, 2025

View reviewed changes

fix: reset check_static_comp

7e1a8ac

scyyh11 marked this pull request as draft October 10, 2025 08:35

fix: add new TestSigmoidFp32_Comp to verify forward decomposition cor…

4dbe344

…rectness of sigmoid under FP32

scyyh11 marked this pull request as ready for review October 10, 2025 09:22

fix: add TestPowFp64_Comp to verify forward decomposition correctness…

935355e

… and gradient checks for pow operation in FP64 precision

scyyh11 requested a review from xiaoguoguo626807 October 11, 2025 07:20

SigureMo reviewed Oct 11, 2025

View reviewed changes

scyyh11 requested a review from SigureMo October 12, 2025 00:54

xiaoguoguo626807 approved these changes Oct 13, 2025

View reviewed changes

YqGe585 approved these changes Oct 13, 2025

View reviewed changes

luotao1 merged commit 169e64c into PaddlePaddle:develop Oct 13, 2025
68 of 69 checks passed

scyyh11 deleted the fix/test_activation_op branch October 13, 2025 10:13

Conversation

scyyh11 commented Sep 26, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR Category

PR Types

Description

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

scyyh11 Sep 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!

scyyh11 commented Oct 9, 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!

scyyh11 commented Oct 10, 2025

Uh oh!

xiaoguoguo626807 commented Oct 11, 2025

Uh oh!

scyyh11 commented Oct 11, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

scyyh11 Oct 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!

scyyh11 commented Oct 11, 2025

Uh oh!

xiaoguoguo626807 left a comment

Choose a reason for hiding this comment

Uh oh!

YqGe585 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

scyyh11 commented Sep 26, 2025 •

edited

Loading

scyyh11 Sep 26, 2025 •

edited

Loading

scyyh11 Oct 11, 2025 •

edited

Loading