Description
🐛 Describe the bug
This issue was discovered in #134184 and left as follow-up work.
There's a correctness issue in nn.ReplicationPad1d and nn.ReplicationPad2d with certain shapes as input. The reproducer below is for nn.ReplicationPad1d, and the first shape does not show any correctness issues, but the second shape does.
import torch
shapes = ([2, 65736, 4], [65736, 2, 4])
pl, pr = 3, 4
for shape in shapes:
x_cpu = torch.randn(shape, device='cpu', requires_grad=True)
x_mps = x_cpu.clone().detach().to('mps').requires_grad_(True)
model = torch.nn.ReplicationPad1d((pl, pr))
# forward
out_cpu = model(x_cpu)
out_mps = model(x_mps)
print(f"{((x_cpu - x_mps.cpu()).abs() > 1e-5).sum() = }")
# backward
g_cpu = torch.randn_like(out_cpu)
g_mps = g_cpu.clone().detach().to('mps').requires_grad_(True)
print(f"{((g_cpu - g_mps.cpu()).abs() > 1e-5).sum() = }")
out_cpu.backward(g_cpu)
out_mps.backward(g_mps)
print(f"{((x_cpu.grad - x_mps.grad.cpu()).abs() > 1e-5).sum() = }")
print()
# Output:
# ((x_cpu - x_mps.cpu()).abs() > 1e-5).sum() = tensor(0)
# ((g_cpu - g_mps.cpu()).abs() > 1e-5).sum() = tensor(0)
# ((x_cpu.grad - x_mps.grad.cpu()).abs() > 1e-5).sum() = tensor(0)
#
# ((x_cpu - x_mps.cpu()).abs() > 1e-5).sum() = tensor(0)
# ((g_cpu - g_mps.cpu()).abs() > 1e-5).sum() = tensor(0)
# ((x_cpu.grad - x_mps.grad.cpu()).abs() > 1e-5).sum() = tensor(524283)A reproducer for nn.ReplicationPad2d can be found in
Lines 8612 to 8656 in defb515
Versions
PyTorch version: 2.5.0a0+git042f2f7
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A
OS: macOS 14.6.1 (arm64)
GCC version: Could not collect
Clang version: 15.0.0 (clang-1500.3.9.4)
CMake version: version 3.30.1
Libc version: N/A
Python version: 3.12.4 | packaged by conda-forge | (main, Jun 17 2024, 10:13:44) [Clang 16.0.6 ] (64-bit runtime)
Python platform: macOS-14.6.1-arm64-arm-64bit
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
CPU:
Apple M3 Max
Versions of relevant libraries:
[pip3] flake8==6.1.0
[pip3] flake8-bugbear==23.3.23
[pip3] flake8-comprehensions==3.15.0
[pip3] flake8-executable==2.1.3
[pip3] flake8-logging-format==0.9.0
[pip3] flake8-pyi==23.3.1
[pip3] flake8-simplify==0.19.3
[pip3] mypy==1.10.0
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.26.0
[pip3] optree==0.12.1
[pip3] torch==2.5.0a0+git042f2f7
[pip3] torch-tb-profiler==0.4.3
[pip3] torchvision==0.20.0a0+0d80848
[pip3] triton==3.0.0
[conda] numpy 1.26.0 pypi_0 pypi
[conda] optree 0.12.1 pypi_0 pypi
[conda] torch 2.5.0a0+git042f2f7 dev_0
[conda] torch-tb-profiler 0.4.3 pypi_0 pypi
[conda] torchfix 0.4.0 pypi_0 pypi
[conda] torchvision 0.20.0a0+0d80848 dev_0
[conda] triton 3.0.0 pypi_0 pypi
cc @ezyang @albanD @gqchen @pearu @nikitaved @soulitzer @Varal7 @xmfan @kulinseth @malfet @DenisVieriu97 @jhavukainen