Optimized le.Tensor_out returns b <= a for reverse-argument broadcasts

未关闭
  • 哪里不对？
#24,485 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

@SofianElmotiem 已经在做这个了。

开始于 2026年11月3日。

  • #24486 来自 @SofianElmotiem —— 未关闭

评估

难度
3/5
预计耗时
1-2 天
新手友好度
35/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
cpp, pytorch
领域
machine-learning

调研方向

Start with kernels/optimized/cpu/op_le.cpp and the reverse-argument handling described in the issue; then inspect the related tests in kernels/test/op_le_test.cpp. Add regression coverage that exercises the optimized paths with matching input and output dtypes. Done means those paths produce the expected a <= b results for the listed broadcast shapes.

由索引模型根据 Issue 内容生成。

描述

Describe the bug

With the optimized CPU kernels, le.Tensor_out returns b <= a instead of a <= b when the left operand is the one being broadcast and both inputs and the output share a dtype (for example two Bool masks, or a non-Bool output).

Repro (C++ kernel test)
TensorFactory<ScalarType::Bool> tf;
Tensor a = tf.make({4}, {false, true, false, true});
Tensor b = tf.make({2, 4}, {false, false, true, true, true, true, false, false});
Tensor out = tf.zeros({2, 4});
torch::executor::aten::le_outf(context_, a, b, out);
Result
Expected (a <= b, also what the portable kernel returns) {1, 0, 1, 1, 1, 1, 1, 0}
Optimized kernel {1, 1, 0, 1, 0, 1, 1, 1} (that is b <= a)

The same happens for Float and Int inputs with a Float/Int output, on all three reverse-argument broadcast paths: (4,) vs (2, 4), (2, 1) vs (2, 3), and (2, 1, 2) vs (2, 3, 2). Float inputs with a Bool output are not affected, because select_optimized_path returns kNone when the output dtype differs and the portable comparison is used.

Cause

select_optimized_path returns kBroadcast2dBy1dReverseArguments, kBroadcastLastDimReverseArguments or kBroadcastNdByNdReverseArguments when the left operand is the broadcast one, and handle_broadcast_elementwise then calls the lambda with the operands swapped. op_div.cpp and op_add_sub_impl.h account for this by swapping their lambda on those paths, but op_le.cpp always passes x.le(y):

https://github.com/pytorch/executorch/blob/aed328df53/kernels/optimized/cpu/op_le.cpp#L57-L63

The existing BroadcastLastDimReverseTest, BroadcastNdByNdReverseTest and similar tests in kernels/test/op_le_test.cpp use a Bool output with Int inputs, so they take the portable path and never exercise the optimized broadcast code.

I have a fix with regression tests and will open a PR.

Versions

main at aed328df53, built from source with CMake (EXECUTORCH_BUILD_TESTS=ON, EXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON), Ubuntu, x86_64, GCC 15.2.0.

主要语言
Python
星标
5.2k
派生
1.2k
平均合并
1 天 22 小时
30 天内合并 PR
675

环境准备

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

pytorch/executorch 的其他 Issue

查看 pytorch/executorch 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。