MIOpen: Failed to launch kernel: invalid configuration argument

[deeptimhe]

Hi,

I am running PyTorch 2.4.1 / ROCm 6.1.2 on MI300X. But got the following error:

MIOpen(HIP): Error [EvaluateInvokers] /long_pathname_so_that_rpms_can_package_the_debug_info/src/extlibs/MLOpen/src/hipoc/hipoc_kernel.cpp:104: Failed to launch kernel: invalid configuration argument

The code can still run after printing the error. Will this error affect the run?

For reproduction, I use the Docker images: deeptimhe/ubuntu22.04-rocm6.1.2-python3.10-pytorch2.4:orig-sing-1010

[junliume]

@deeptimhe could you report back with the log when the following env vars are enabled?

export MIOPEN_ENABLE_LOGGING=1
export MIOPEN_ENABLE_LOGGING_CMD=1

[deeptimhe]

@junliume Hi, thanks for the reply. Actually I can't reproduce the error right now. It seems to happen at random.