[BugFix] Fix put_along_axis grad kernel crash for integer div-by-zero and 0-size input by wangbingguang2026 · Pull Request #78482 · PaddlePaddle/Paddle

wangbingguang2026 · 2026-03-25T12:17:11Z

PR Category

Operator Mechanism

PR Types

Bug fixes

Description

修复 paddle.put_along_axis 反向传播中的两个 crash bug：

Bug A：reduce="mul" 整数除零 SIGFPE

当 reduce="mul" 且输入为整数类型（uint8/int32/int64）时，如果 x 或 value 中包含零值，反向传播中的梯度计算会触发整数除零（SIGFPE），导致进程崩溃。

根因：梯度公式 grad * out / x 和 self * (out / value) 在除数为 0 时未做保护。浮点类型产生 inf/nan 不会 crash，但整数类型会触发 SIGFPE 信号。

修复：在 CPU 和 GPU 的 4 处除法位置增加零值判断，除数为 0 时梯度直接赋 0。

修改文件：

paddle/phi/kernels/funcs/gather_scatter_functor.cc：cpu_scatter_mul_min_max_input_grad_kernel 和 cpu_scatter_mul_min_max_value_grad_kernel
paddle/phi/kernels/funcs/gather_scatter_functor.cu：ScatterMulInputGradGPUKernel 和 ScatterMulValueGradGPUKernel

Bug B：0-size input CPU 反向传播越界访问

当 input tensor 含 0-size 维度（如 shape [0, 60]）但 index 在该维度非零时，CPU 反向传播会越界写入内存导致崩溃。GPU grad kernel 已有 x.numel() == 0 的 early return 保护，但 CPU 版本遗漏了。

修复：在 put_along_axis_grad_kernel.cc 入口处增加与 GPU 版本相同的 0-size 保护。

修改文件：

paddle/phi/kernels/cpu/put_along_axis_grad_kernel.cc

单测补充

在 test/legacy_test/test_put_along_axis_op.py 中新增两个测试类：

TestPutAlongAxisMulIntegerDivByZero：覆盖 Bug A，测试 uint8/int32/int64 在 CPU/GPU 上 reduce="mul" backward 不 crash
TestPutAlongAxisZeroSizeInputGrad：覆盖 Bug B，测试 input 各维度为 0 时 assign/add/mul 前向不 crash

是否引起精度变化

否。修改仅在除数为零或 numel 为零的边界条件下生效（梯度赋 0），不影响正常数值计算路径。

case 复测结果

0_size_tensor_1_8_1.txt 中 put_along_axis 相关的 case 一共有123个，测试结果如下：

Passed cases：96
Failed cases：0
Skipped cases：27

0_size_tensor_1_8_invalid_1.txt 中 put_along_axis 相关的 case 一共有40个，测试结果如下：

Passed cases：27
Failed cases：10
Skipped cases：3

其中，Failed cases 都是 paddle error，Python 层的 shape 校验正确地拦截了越界输入并抛出 ValueError，这些 case 本身就是非法输入，应该报错。Skipped cases 都是 numpy error，PaddleAPITest 框架将它们标记为 numpy_error 并跳过，这是测试数据本身的边界问题。

… and 0-size input Fix two crash bugs in put_along_axis backward: 1. SIGFPE in reduce="mul" grad when integer divisor is zero (CPU & GPU) 2. Out-of-bounds memory access in CPU grad kernel when input has 0-size dim Add unit tests covering both scenarios. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

paddle-bot · 2026-03-25T12:17:19Z

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.

CLAassistant · 2026-03-25T12:17:19Z

All committers have signed the CLA.

codecov-commenter · 2026-03-25T15:58:16Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (develop@3d475a5). Learn more about missing BASE report.

Additional details and impacted files

@@             Coverage Diff             @@
##             develop    #78482   +/-   ##
===========================================
  Coverage           ?   100.00%           
===========================================
  Files              ?         2           
  Lines              ?        16           
  Branches           ?         0           
===========================================
  Hits               ?        16           
  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.

- clang-format: break long line in gather_scatter_functor.cu - ruff-format: reformat test_put_along_axis_op.py - Fix test_uint8_cpu: integer types do not support autograd (sum_grad kernel not registered for uint8/int32/int64), change to forward-only verification to avoid NotFoundError in CI Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

wanghuancoder

LGTM

wangbingguang2026 · 2026-03-30T03:23:57Z

/re-run all-failed

wangbingguang2026 · 2026-03-30T06:56:35Z

/re-run all-failed

wangbingguang2026 · 2026-03-30T09:15:43Z

/coverage test_put_along_axis_op

From00

LGTM

Add TestPutAlongAxisMulFloat32DivByZeroGrad to exercise the new zero-divisor guards in gather_scatter_functor.cc/.cu: - test_x_grad_with_zero_in_x: covers x==0 branch in input_grad kernel - test_value_grad_with_zero_in_value: covers value==0 branch in value_grad kernel Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

From00

LGTM

wanghuancoder

LGTM

wangbingguang2026 · 2026-03-31T02:47:03Z

@lugimzzz @Jiang-Jia-Jun 麻烦两位大佬帮忙审批一下～

lugimzzz

lgtm

wangbingguang2026 · 2026-03-31T03:30:00Z

/re-run all-failed

…erage The TestPutAlongAxisMulFloat32DivByZeroGrad test was not covering gather_scatter_functor.cc because tensors defaulted to GPU on CI machines. Add explicit place=paddle.CPUPlace() to ensure the CPU backward kernel path is exercised. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

wanghuancoder

LGTM

wanghuancoder

LGTM

1. gather_scatter_functor.cc: split multi-line expression into separate statements so gcov can attribute execution counts to each line. 2. test_put_along_axis_op.py: add place=CPUPlace() and loss.backward() to TestPutAlongAxisZeroSizeInputGrad so the CPU grad kernel's numel==0 early-return path is exercised in Coverage CI. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…coverage The existing zero-size tests all had indices.numel()==0, which triggers a Python-level early return and never reaches the C++ grad kernel. Add a test case with x_shape=[2,0,3] (x.numel()==0) but idx_shape=[2,1,3] (indices.numel()!=0) so the C++ kernel's numel==0 early-return guard (lines 38-45) is actually exercised. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

wangbingguang2026 · 2026-04-03T02:29:37Z

/re-run all-failed

wangbingguang2026 · 2026-04-03T07:32:37Z

/re-run all-failed

wanghuancoder

LGTM

wangbingguang2026 · 2026-04-03T07:38:57Z

/re-run all-failed

From00

LGTM

wangbingguang2026 · 2026-04-08T07:00:31Z

/re-run all-failed

wangbingguang2026 · 2026-04-08T13:23:22Z

@sneaxiy 帮忙审核并合入一下

… and 0-size input (PaddlePaddle#78482) * [BugFix] Fix put_along_axis grad kernel crash for integer div-by-zero and 0-size input Fix two crash bugs in put_along_axis backward: 1. SIGFPE in reduce="mul" grad when integer divisor is zero (CPU & GPU) 2. Out-of-bounds memory access in CPU grad kernel when input has 0-size dim Add unit tests covering both scenarios. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * [BugFix] Fix codestyle and test issues in put_along_axis fix - clang-format: break long line in gather_scatter_functor.cu - ruff-format: reformat test_put_along_axis_op.py - Fix test_uint8_cpu: integer types do not support autograd (sum_grad kernel not registered for uint8/int32/int64), change to forward-only verification to avoid NotFoundError in CI Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * [BugFix] Add float32 backward test for put_along_axis div-by-zero guard Add TestPutAlongAxisMulFloat32DivByZeroGrad to exercise the new zero-divisor guards in gather_scatter_functor.cc/.cu: - test_x_grad_with_zero_in_x: covers x==0 branch in input_grad kernel - test_value_grad_with_zero_in_value: covers value==0 branch in value_grad kernel Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [BugFix] Force CPU placement in float32 div-by-zero grad test for coverage The TestPutAlongAxisMulFloat32DivByZeroGrad test was not covering gather_scatter_functor.cc because tensors defaulted to GPU on CI machines. Add explicit place=paddle.CPUPlace() to ensure the CPU backward kernel path is exercised. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * [BugFix] Improve patch coverage for put_along_axis grad fix 1. gather_scatter_functor.cc: split multi-line expression into separate statements so gcov can attribute execution counts to each line. 2. test_put_along_axis_op.py: add place=CPUPlace() and loss.backward() to TestPutAlongAxisZeroSizeInputGrad so the CPU grad kernel's numel==0 early-return path is exercised in Coverage CI. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * [BugFix] Add zero-size x with non-empty indices test for grad kernel coverage The existing zero-size tests all had indices.numel()==0, which triggers a Python-level early return and never reaches the C++ grad kernel. Add a test case with x_shape=[2,0,3] (x.numel()==0) but idx_shape=[2,1,3] (indices.numel()!=0) so the C++ kernel's numel==0 early-return guard (lines 38-45) is actually exercised. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

… and 0-size input (#78482) (#78621) * [BugFix] Fix put_along_axis grad kernel crash for integer div-by-zero and 0-size input Fix two crash bugs in put_along_axis backward: 1. SIGFPE in reduce="mul" grad when integer divisor is zero (CPU & GPU) 2. Out-of-bounds memory access in CPU grad kernel when input has 0-size dim Add unit tests covering both scenarios. * [BugFix] Fix codestyle and test issues in put_along_axis fix - clang-format: break long line in gather_scatter_functor.cu - ruff-format: reformat test_put_along_axis_op.py - Fix test_uint8_cpu: integer types do not support autograd (sum_grad kernel not registered for uint8/int32/int64), change to forward-only verification to avoid NotFoundError in CI * [BugFix] Add float32 backward test for put_along_axis div-by-zero guard Add TestPutAlongAxisMulFloat32DivByZeroGrad to exercise the new zero-divisor guards in gather_scatter_functor.cc/.cu: - test_x_grad_with_zero_in_x: covers x==0 branch in input_grad kernel - test_value_grad_with_zero_in_value: covers value==0 branch in value_grad kernel * [BugFix] Force CPU placement in float32 div-by-zero grad test for coverage The TestPutAlongAxisMulFloat32DivByZeroGrad test was not covering gather_scatter_functor.cc because tensors defaulted to GPU on CI machines. Add explicit place=paddle.CPUPlace() to ensure the CPU backward kernel path is exercised. * [BugFix] Improve patch coverage for put_along_axis grad fix 1. gather_scatter_functor.cc: split multi-line expression into separate statements so gcov can attribute execution counts to each line. 2. test_put_along_axis_op.py: add place=CPUPlace() and loss.backward() to TestPutAlongAxisZeroSizeInputGrad so the CPU grad kernel's numel==0 early-return path is exercised in Coverage CI. * [BugFix] Add zero-size x with non-empty indices test for grad kernel coverage The existing zero-size tests all had indices.numel()==0, which triggers a Python-level early return and never reaches the C++ grad kernel. Add a test case with x_shape=[2,0,3] (x.numel()==0) but idx_shape=[2,1,3] (indices.numel()!=0) so the C++ kernel's numel==0 early-return guard (lines 38-45) is actually exercised. --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

paddle-bot Bot added the contributor External developers label Mar 25, 2026

wanghuancoder previously approved these changes Mar 27, 2026

View reviewed changes

From00 previously approved these changes Mar 30, 2026

View reviewed changes

wangbingguang2026 dismissed stale reviews from From00 and wanghuancoder via 413ee3c March 30, 2026 12:48

From00 previously approved these changes Mar 30, 2026

View reviewed changes

wanghuancoder previously approved these changes Mar 31, 2026

View reviewed changes

Jiang-Jia-Jun approved these changes Mar 31, 2026

View reviewed changes

lugimzzz approved these changes Mar 31, 2026

View reviewed changes

wangbingguang2026 dismissed stale reviews from wanghuancoder and From00 via 6cba785 March 31, 2026 10:47

wanghuancoder previously approved these changes Mar 31, 2026

View reviewed changes

wangbingguang2026 dismissed wanghuancoder’s stale review via 100bb7c April 1, 2026 07:28

wanghuancoder previously approved these changes Apr 1, 2026

View reviewed changes

wangbingguang2026 dismissed wanghuancoder’s stale review via b1d52bc April 1, 2026 08:41

wangbingguang2026 force-pushed the fix/put_along_axis_grad_crash branch from 100bb7c to b1d52bc Compare April 1, 2026 08:41

wanghuancoder approved these changes Apr 3, 2026

View reviewed changes

From00 approved these changes Apr 3, 2026

View reviewed changes

wangbingguang2026 closed this Apr 7, 2026

wangbingguang2026 reopened this Apr 7, 2026

wanghuancoder merged commit 5ae373f into PaddlePaddle:develop Apr 9, 2026
244 of 256 checks passed

wangbingguang2026 mentioned this pull request Apr 9, 2026

[cherry-pick] Fix put_along_axis grad kernel crash for integer div-by-zero… #78621

Merged

Conversation

wangbingguang2026 commented Mar 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR Category

PR Types

Description

Bug A：reduce="mul" 整数除零 SIGFPE

Bug B：0-size input CPU 反向传播越界访问

单测补充

是否引起精度变化

case 复测结果

Uh oh!

paddle-bot Bot commented Mar 25, 2026

Uh oh!

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

Uh oh!

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!

wanghuancoder left a comment

Choose a reason for hiding this comment

Uh oh!

wangbingguang2026 commented Mar 30, 2026

Uh oh!

wangbingguang2026 commented Mar 30, 2026

Uh oh!

wangbingguang2026 commented Mar 30, 2026

Uh oh!

From00 left a comment

Choose a reason for hiding this comment

Uh oh!

From00 left a comment

Choose a reason for hiding this comment

Uh oh!

wanghuancoder left a comment

Choose a reason for hiding this comment

Uh oh!

wangbingguang2026 commented Mar 31, 2026

Uh oh!

lugimzzz left a comment

Choose a reason for hiding this comment

Uh oh!

wangbingguang2026 commented Mar 31, 2026

Uh oh!

wanghuancoder left a comment

Choose a reason for hiding this comment

Uh oh!

wanghuancoder left a comment

Choose a reason for hiding this comment

Uh oh!

wangbingguang2026 commented Apr 3, 2026

Uh oh!

wangbingguang2026 commented Apr 3, 2026

Uh oh!

wanghuancoder left a comment

Choose a reason for hiding this comment

Uh oh!

wangbingguang2026 commented Apr 3, 2026

Uh oh!

From00 left a comment

Choose a reason for hiding this comment

Uh oh!

wangbingguang2026 commented Apr 8, 2026

Uh oh!

wangbingguang2026 commented Apr 8, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

wangbingguang2026 commented Mar 25, 2026 •

edited

Loading

CLAassistant commented Mar 25, 2026 •

edited

Loading

codecov-commenter commented Mar 25, 2026 •

edited

Loading