| 52 | bool uses_host_graph_plan(ggml_backend_t backend); |
| 53 | bool requested_backend_uses_host_graph_plan(const BackendConfig & config); |
| 54 | // Drop the CUDA/HIP context's cached (idle) pool memory back to the driver. |
| 55 | // No-op on other backends. For use on allocation-failure paths before a retry. |
| 56 | void trim_backend_pools(ggml_backend_t backend); |
| 57 |