Currently CUDA Graph are disable when using batched PDLP. We should either find the root cause of the problem or change the batch PDLP implementation.