[Release/3.3] Fix out-of-bounds memory access in SetKernel for 0-size tensor

[hushenwei2000]

PR Category

Operator Mechanism

PR Types

Bug fixes

Description

devPR:#78486

修复 0-Size 报错问题

paddle.Tensor.set_	accuracy	CPU	精度不对	paddle.Tensor.set_(Tensor([20],"complex64"), Tensor([0, 3],"complex64"), list[20,], list[2,], 0, ) paddle.Tensor.set_(Tensor([20],"float32"), Tensor([0, 3],"float32"), list[20,], list[2,], 0, )

Summary

• Fix GPU out-of-bounds memory access (detected by compute-sanitizer) when calling Tensor.set_(source, shape, stride, offset) with a 0-size source tensor and non-zero target shape.
• When source.numel() == 0, the output tensor now inherits the source's 0-size dims/strides instead of using user-specified shape/stride, preventing invalid memory access.
• Updated existing TestSet_API_ZeroSize and added 5 new test cases covering various 0-size scenarios.

Root Cause

In SetKernel (paddle/phi/kernels/set_kernel.cc), the conditional logic for handling 0-size tensors had a missing branch: when source.numel() == 0 and x.numel() != 0, no branch was executed. The output tensor retained its original data holder, but SetInferMeta had already set its meta to the user-specified shape/stride (e.g., shape=[20], stride=[2]). When ContiguousKernel later attempted to read 20 elements via stride=2 (requiring storage for indices 0–38), the underlying storage was empty (nullptr), causing CUDA illegal memory access.

Before fix (missing branch):

if source.numel() != 0:    # False (source is 0-size)
    ...
elif x.numel() == 0:        # False (x has 20 elements)
    ...

Neither branch executes → out keeps stale holder with mismatched meta

Fix

When source.numel() == 0, force the output tensor's dims/strides to match the source's 0-size shape, and rebind the holder to the source's (empty) storage. This ensures numel == 0, so downstream kernels (e.g., ContiguousKernel) skip safely via their numel <= 0 early-return guards.

Test Plan

• Verified fix with compute-sanitizer: ERROR SUMMARY: 0 errors (was 11 errors before fix)
  • paddle.Tensor.set_(Tensor([20],"float32"), Tensor([0, 3],"float32"), list[20,], list[2,], 0, )
• Added 5 new test cases in TestSet_API_ZeroSize:
  • test_zero_size_source_with_nonzero_shape — 0-size source + explicit non-zero shape
  • test_zero_size_source_default_args — 0-size source with default shape/stride
  • test_zero_size_x_nonzero_source — 0-size x with non-zero source
  • test_both_zero_size — both x and source are 0-size
  • test_zero_size_source_no_crash_on_contiguous — no crash on .contiguous() after set_
• APITest 中现有 set_ 全部测试

paddle.Tensor.set_(Tensor([20],"float64"), Tensor([0, 3],"float64"), list[20,], list[2,], 0, )
paddle.Tensor.set_(Tensor([20],"float64"), Tensor([15, 0],"float64"), list[20,], list[2,], 0, )
paddle.Tensor.set_(Tensor([3, 0],"float16"), Tensor([6, 0],"float16"), list[3,8,], list[2,2,], 0, )
paddle.Tensor.set_(Tensor([3, 8],"float16"), Tensor([0, 3],"float16"), list[3,8,], list[2,2,], 0, )
paddle.Tensor.set_(Tensor([3, 8],"float16"), Tensor([6, 0],"float16"), list[3,8,], list[2,2,], 0, )
paddle.Tensor.set_(Tensor([20],"complex64"), Tensor([0, 3],"complex64"), list[20,], list[2,], 0, )
paddle.Tensor.set_(Tensor([20],"float32"), Tensor([0, 3],"float32"), list[20,], list[2,], 0, )
paddle.Tensor.set_(Tensor([20],"complex64"), Tensor([15, 0],"complex64"), list[20,], list[2,], 0, )
paddle.Tensor.set_(Tensor([20],"float32"), Tensor([15, 0],"float32"), list[20,], list[2,], 0, )

是否引起精度变化

否

[paddle-bot]

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]

Codecov Report

❌ Patch coverage is 94.59459% with 2 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (release/3.3@88e1968). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
paddle/phi/kernels/set_kernel.cc	94.59%	2 Missing ⚠️

Additional details and impacted files

@@              Coverage Diff               @@
##             release/3.3   #78536   +/-   ##
==============================================
  Coverage               ?   94.59%           
==============================================
  Files                  ?        1           
  Lines                  ?       37           
  Branches               ?        0           
==============================================
  Hits                   ?       35           
  Misses                 ?        2           
  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.

[hushenwei2000]

/re-run all-failed

[hushenwei2000]

/re-run all-failed

[wanghuancoder]

LGTM