When running step3, the error "CUDA error: misaligned address"?

[EircYangQiXin]

When running step3, the error "CUDA error: misaligned address
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions." occurred. Is this caused by insufficient graphics memory?

[SmartmediaAI]

I met the same problem. Does anyone know the reasons?