Struct ModelSamplingDefaults 

pub struct ModelSamplingDefaults {
    pub temperature: ModelSamplingDefault,
    pub top_p: ModelSamplingDefault,
    pub top_k: ModelSamplingDefault,
    pub min_p: ModelSamplingDefault,
    pub repeat_penalty: ModelSamplingDefault,
}

The sampler defaults one model declares.

Copy, and deliberately so: it rides ModelLaunchSpec into the resident set and out again on every current_model() call, so it is cloned far more often than it is built. Adding a String here would put an allocation on the admission fast path and cost with_model_sampling its const.

Fields

temperature: ModelSamplingDefault

general.sampling.temp.

top_p: ModelSamplingDefault

general.sampling.top_p.

top_k: ModelSamplingDefault

general.sampling.top_k.

min_p: ModelSamplingDefault

general.sampling.min_p.

repeat_penalty: ModelSamplingDefault

general.sampling.penalty_repeat.

Implementations

impl ModelSamplingDefaults

pub fn from_metadata<S: BuildHasher>( metadata: &HashMap<String, String, S>, ) -> Self

Read what a model declares out of its stored GGUF metadata.

Values arrive stringified by GgufValue::to_string(), so a FLOAT32 0.7 is the string "0.7" — Rust’s Display prints the shortest form that round-trips. Parsing it back yields 0.7, while /props reports the same value widened from f32 (0.699999988079071); the ~1.2e-8 gap is well inside the epsilon the comparison uses.

pub fn get(&self, field: &str) -> ModelSamplingDefault

Look one field up by its gglib wire name.

Mirrors SlotParams::get so the baseline check reads both sides the same way. A name with no GGUF key — presence_penalty, dry_multiplier, or anything unknown — is Absent by construction, not because this model happened not to set it.

pub fn gguf_key(field: &str) -> Option<&'static str>

The GGUF key that can move field, if any.

impl ModelSamplingDefaults

pub fn compare_field( &self, field: &str, sending: Option<f64>, ) -> SamplingOverride

Compare one field’s published value against what gglib will send.

sending is what the ladder resolved: None means gglib names the field nowhere and llama.cpp is left to apply the model’s own value.

pub fn compare_all( &self, resolved: impl Fn(&str) -> Option<f64>, ) -> Vec<(&'static str, SamplingOverride)>

Compare every field a model can publish, in MODEL_SAMPLING_KEYS order.

resolved looks each field up by its gglib wire name and returns what the ladder decided to send, so callers hand in whichever config they are explaining rather than this module learning about InferenceConfig.

Trait Implementations

impl Clone for ModelSamplingDefaults

fn clone(&self) -> ModelSamplingDefaults

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ModelSamplingDefaults

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

Formats the value using the given formatter.

impl Default for ModelSamplingDefaults

fn default() -> ModelSamplingDefaults

Returns the “default value” for a type.

impl PartialEq for ModelSamplingDefaults

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

Tests for self and other values to be equal, and is used by ==.

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

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Copy for ModelSamplingDefaults

impl StructuralPartialEq for ModelSamplingDefaults