Is there performance issue when using GGML_CUDA_GRAPH=ON when do cmake?

[wjkim00]

I have faced the problem to use GGML_CUDA=ON.
If I install this with CMAKE_ARGS="GGML_CUDA=ON", The error is issued at cuda graphs when eval. (like /api/chat)

cudaGraphKernelNodeSetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i])
CUDA error: invalid argument

My environment is that NVIDIA Graphic Driver version is 525.25 and CUDA is 12.0. Graphic card is A100.
I have searched similiar issues or discussions, but the solutions are not in anywhere.
So I use the CMAKE_ARGS+="GGML_CUDA_GRAPH=OFF", There is no error when eval.

I wonder that there is performance issue when use GGML_CUDA_GRAPH=OFF.
Thanks for your reply.