generate-dependencies-with-compile in RTX3060 Cuda11.1

[v-nhandt21]

I wonder the cuda implement of SRU work in CUDA11, I found in the forum that to solve the problem, we must use cuda10.2 https://discuss.pytorch.org/t/just-in-time-loading-and-compiling-cuda-kernels-was-unsuccesful/124486

Here is my computer info

Does the speed of SRU decrease if we just use operation of pytorch, not the opt write in c++, is there that version?

If there are someone solve problem run in 3060, can you share!

[taolei87]

hi @v-nhandt21 , the compilation arguments such as "--generate-dependencies-with-compile" are automatically added by ninja / nvcc.

Looking at your first screenshot, ninja/nvcc attempts to build the code using --sm_75 and --compute_75 architecture code. However, the correct code should be sm_86 for your RTX 3060, according to this article:
https://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/#ampere-cuda-11-1-and-later

Maybe you can google search or ask in the Pytorch forum how to fix this arch code issue.