Can't run torch conv2d with AMD Instinct MI250X

projects
1

I am on AMD 250X GPU after installed rocm5.6 with pytorch version 2.2 following commend

 pip install torch==2.2.0 torchvision==0.17.0 torchaudio==2.2.0 --index-url https://download.pytorch.org/whl/rocm5.6

I am not able to run the test code:

import torch

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
x = torch.rand(1, 3, 64, 64, device=device)  # Random tensor
conv = torch.nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1).to(device)
y = conv(x)
print(y.shape)

error:

MIOpen(HIP): Warning [FindWinogradSolutions] /MIOpen/src/sqlite_db.cpp:220: Internal error while accessing SQLite database: disk I/O error
MIOpen(HIP): Warning [FindDataGemmSolutions] /MIOpen/src/sqlite_db.cpp:220: Internal error while accessing SQLite database: disk I/O error
MIOpen(HIP): Warning [FindDataDirectSolutions] /MIOpen/src/sqlite_db.cpp:220: Internal error while accessing SQLite database: disk I/O error
MIOpen(HIP): Warning [FindDataImplicitGemmSolutions] /MIOpen/src/sqlite_db.cpp:220: Internal error while accessing SQLite database: disk I/O error
MIOpen(HIP): Warning [FindFftSolutions] /MIOpen/src/sqlite_db.cpp:220: Internal error while accessing SQLite database: disk I/O error
MIOpen Error: /MIOpen/src/ocl/convolutionocl.cpp:517: Forward Convolution cannot be executed due to incorrect params
Traceback (most recent call last):
....
return F.conv2d(input, weight, bias, self.stride,
RuntimeError: miopenStatusUnknownError

However I don’t have issue with AMD 210X GPU.