【UnitTestFix No.14】fix test_matmul_v2_op.py

[scyyh11]

PR Category

Operator Mechanism

PR Types

Bug fixes

Description

This PR fixes a critical issue where the matmul_v2 gradient operation produces inconsistent gradient shapes between eager mode and static compilation mode when the inputs are 1-D tensors.

Problem

• Eager mode correctly returns 1-D gradients (n,) for 1-D inputs.
• Static compilation mode (with check_prim_pir=True) incorrectly returns 2-D gradients (n, 1) or (1, n) for 1-D inputs.
• This inconsistency causes test failures with shape mismatch errors like:

  AssertionError: Not equal to tolerance rtol=1e-15, atol=1e-15
  Check static comp grad out failed. Mismatch between static comp and eager on Place(gpu:0)
  static comp grad out tensor: [[-0.01653409 -0.33412735 ...]]  # shape (1, 100)
  eager grad out tensor: [-0.01653409 -0.33412735 ...]        # shape (100,)

Root Cause

The shape inconsistency occurs in two places:

1. Composite implementation (paddle/fluid/primitive/decomp_rule/decomp_vjp/details.h):
   When check_prim_pir=True, the composite matmul_grad function unsqueezes 1-D inputs to 2-D for computation but does not convert the resulting 2-D gradients back to 1-D.

2. Kernel implementation (paddle/phi/kernels/impl/matmul_grad_kernel_impl.h):
   Internal matrix operations produce 2-D gradients that are not reshaped back to 1-D for 1-D inputs.

Fix

• In composite implementation, added shape handling logic to squeeze 2-D gradients back to 1-D if the original input was 1-D.
• In kernel implementation, applied the same logic for consistency.
• Both paths now check whether the original input was 1-D and squeeze the appropriate dimension to ensure consistent gradient shapes.

Changes

1. Modified matmul_grad in details.h to handle 1-D gradient shapes.
2. Modified MatmulGradKernel in matmul_grad_kernel_impl.h for consistency.

This fix ensures consistent behavior between eager and static compilation modes, resolving gradient shape mismatch errors in matmul_v2 backpropagation.

---

问题描述

• 动态图模式：对 1 维输入，正确返回 1 维梯度 (n,)。
• 静态编译模式（check_prim_pir=True）：错误地返回 2 维梯度 (n,1) 或 (1,n)。
• 这种不一致会导致测试失败，报出形状不匹配错误，例如：

  AssertionError: Not equal to tolerance rtol=1e-15, atol=1e-15
  Check static comp grad out failed. Mismatch between static comp and eager on Place(gpu:0)
  static comp grad out tensor: [[-0.01653409 -0.33412735 ...]]  # shape (1, 100)
  eager grad out tensor: [-0.01653409 -0.33412735 ...]        # shape (100,)

根本原因

问题出现在 两个地方：

1. 复合实现（paddle/fluid/primitive/decomp_rule/decomp_vjp/details.h）：
   当 check_prim_pir=True 时，matmul_grad 复合函数会将 1 维输入 unsqueeze 为 2 维 来进行矩阵运算，但 没有在梯度返回时再 squeeze 回 1 维。

2. 内核实现（paddle/phi/kernels/impl/matmul_grad_kernel_impl.h）：
   内部矩阵计算同样会生成 2 维梯度，未根据原始输入形状进行还原。

解决方案

• 复合实现：添加形状处理逻辑，如果原始输入是 1 维，则将 2 维梯度 squeeze 回 1 维。
• 内核实现：添加相同的逻辑，保证一致性。
• 现在两个路径都会检查原始输入维度，并在必要时对梯度进行 squeeze 处理。

修改内容

1. 修改 details.h 中的 matmul_grad 复合函数，支持 1 维梯度形状处理。
2. 修改 matmul_grad_kernel_impl.h 中的 MatmulGradKernel，保证与复合逻辑一致。

此修复确保了 eager 与静态编译模式下梯度计算行为一致，彻底解决了 matmul_v2 反向传播中的形状不匹配问题。

@luotao1 @YqGe585

[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.

[scyyh11]

/re-run all-failed

[A-nnonymous]

LGTM in squeezing logics