Enum LlamaLazyMode 

#[repr(u32)]
pub enum LlamaLazyMode {
    Off = 0,
    Auto = 1,
    On = 2,
}

Whether tensors the architecture marks as lazy are read on demand rather than up front.

Only tensors the model architecture flags carry this at all — today that is Gemma-4’s per-layer token embedding and Qwen4Exp’s PLE rows — so on every other architecture the setting has no effect. Lazy reading always needs mmap; without it llama.cpp warns and loads the tensor in full regardless.

llama_lazy_mode is an unsigned enum upstream (all discriminants are non-negative), unlike the signed LlamaLoadMode.

Variants

Off = 0

Never read lazily — always pull the whole tensor up front.

Auto = 1

Read lazily only for marked tensors larger than 4 GiB. llama.cpp’s default, and downgraded to LlamaLazyMode::Off at load time if any backend device lacks mmap support (iGPUs, for instance).

On = 2

Read every marked tensor’s rows on demand, whatever its size. Trades I/O for resident memory; the 4 GiB floor exists because the per-read overhead is not worth it on small tensors.

Trait Implementations

impl Clone for LlamaLazyMode

fn clone(&self) -> LlamaLazyMode

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for LlamaLazyMode

impl Debug for LlamaLazyMode

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Eq for LlamaLazyMode

impl PartialEq for LlamaLazyMode

fn eq(&self, other: &LlamaLazyMode) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for LlamaLazyMode