Struct SamplingParams 

pub struct SamplingParams {

    pub max_new_tokens: Option<i64>,
    pub stop: Option<StrOrList>,
    pub stop_token_ids: Option<Vec<i64>>,
    pub temperature: f64,
    pub top_p: f64,
    pub top_k: i64,
    pub min_p: f64,
    pub frequency_penalty: f64,
    pub presence_penalty: f64,
    pub repetition_penalty: f64,
    pub min_new_tokens: i64,
    pub json_schema: Option<String>,
    pub regex: Option<String>,
    pub ebnf: Option<String>,
    pub structural_tag: Option<String>,
    pub ignore_eos: bool,
    pub skip_special_tokens: bool,
    pub spaces_between_special_tokens: bool,
    pub no_stop_trim: bool,
    pub thinking_budget: Option<i64>,
    pub custom_params: Option<Dict<String, Value>>,
    pub stream_interval: Option<i64>,
    pub logit_bias: Option<Dict<String, f64>>,
    pub seed: Option<i64>,
    pub logprobs: Option<i64>,
    pub n: i64,
    pub stop_strs: Option<StrOrList>,
    pub stop_str_max_len: i64,
    pub is_normalized: bool,
}

SamplingParams, an untagged array-layout struct nested in a request.

This is the data only. The Python class also normalizes on construction and decode (__post_init__: greedy for a near-zero temperature, stop copied to stop_strs, …) unless is_normalized is set; decoding here applies none of that.

Fields

max_new_tokens: Option<i64>

Generation length limit; None leaves it to the engine.

stop: Option<StrOrList>

API input for stop strings; normalization copies it to stop_strs.

stop_token_ids: Option<Vec<i64>>

Token ids that end generation (a Python set).

temperature: f64

Sampling temperature.

top_p: f64

Nucleus sampling mass.

top_k: i64

Top-k cutoff; -1 disables it before normalization.

min_p: f64

Minimum probability relative to the most likely token.

frequency_penalty: f64

Frequency penalty.

presence_penalty: f64

Presence penalty.

repetition_penalty: f64

Repetition penalty.

min_new_tokens: i64

Tokens generated before stop conditions apply.

json_schema: Option<String>

JSON schema constraint.

regex: Option<String>

Regex constraint.

ebnf: Option<String>

EBNF grammar constraint.

structural_tag: Option<String>

Structural tag constraint.

ignore_eos: bool

Keep generating past end-of-sequence tokens.

skip_special_tokens: bool

Detokenization option, carried for the frontend.

spaces_between_special_tokens: bool

Detokenization option, carried for the frontend.

no_stop_trim: bool

Keep the matched stop string in the output.

thinking_budget: Option<i64>

Reasoning-token budget.

custom_params: Option<Dict<String, Value>>

Free-form parameters for custom logit processors.

stream_interval: Option<i64>

Streaming chunk interval in tokens.

logit_bias: Option<Dict<String, f64>>

Per-token logit offsets, keyed by token id as a string.

seed: Option<i64>

Sampling seed.

logprobs: Option<i64>

Output logprobs: None off, 0 the sampled token’s logprob.

n: i64

OpenAI n; the engine accepts only 1.

stop_strs: Option<StrOrList>

Stop strings after normalization.

stop_str_max_len: i64

Longest stop string, set by normalization.

is_normalized: bool

Whether normalization already ran on the sender.

Trait Implementations

impl Clone for SamplingParams

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for SamplingParams

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

Formats the value using the given formatter.

impl Default for SamplingParams

fn default() -> Self

Returns the “default value” for a type.

impl PartialEq for SamplingParams

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for SamplingParams

impl Wire for SamplingParams

fn encode(&self, encoder: &mut Encoder)

Append this value to a message.

fn decode(decoder: &mut Decoder<'_>) -> Result<Self>

Read one value of this type.