Some devices (i.e. GPUs) may free device memory prior to its actual use being completed on the assumption that subsequent allocations can only be used serially with respect to pending uses. If this function returns a non-zero value it is the value of a device-specific counter such that any device memory tagged with an earlier freed-at count is really unencumbered by pending uses. For this to be
| 267 | // by pending uses. For this to be useful the device memory allocator must |
| 268 | // be tagging deallocated memory chunks using the same counter. |
| 269 | virtual uint64 SafeAllocFrontier(uint64 old_value) { return 0; } |
| 270 | |
| 271 | // Copies `input_tensor` to `output_tensor`, where both tensors are on this |
| 272 | // device. This function assumes that `output_tensor` has already been |