Skip to main content

ClikaRT::io::LoadPolicy

struct

Header: ClikaRT/io/tensors_container.h

How a policy-taking loader materializes tensors.

Data members​

device​

StreamOrDevice device = {}

Where materialized tensors live: a device (uploads ride the calling thread's async transfer lane on it; a bare device states placement, never a lane) or a caller stream the uploads ride. Unset ⇒ CPU.

Declared in ClikaRT/io/tensors_container.h, line 42

sticky​

bool sticky = false

Keep the lineage on disk: op outputs derived from these tensors are backed by disk-spill storage (at/above the runtime's spill threshold), so they can be evicted and re-faulted like the weights themselves. Opt-in per load.

Declared in ClikaRT/io/tensors_container.h, line 47

map_payload​

std::optional<bool> map_payload

Read the checkpoint through a file mapping (true), or read each tensor from the file only when it is asked for (false): with the mapping off, the header is read once, no tensor is read ahead, and one tensor's bytes are resident at a time, the form for a checkpoint larger than the memory it loads into. Unset, the loader decides at the open from the checkpoint's size against the available memory.

Declared in ClikaRT/io/tensors_container.h, line 54