Use the cudaMallocAsync backend in Torch. Enabling this is generally preferable, but it may cause issues with certain workloads. Try disabling it if you experience intermittent OoM errors. If false, Torch will use the allocator defined by the system environment.
option