llama_cpp_sys_4/
common.rs

1//! Manual wrapper for values in llama.cpp/common/common.h
2//!
3//! [`common_sampler_params`] mirrors upstream's `common_params_sampling` (kept in
4//! sync with llama.cpp `b10470`). It is a plain Rust convenience struct, not an
5//! FFI type: `grammar` and `logit_bias` are simplified to owned Rust values
6//! rather than upstream's `common_grammar` / `llama_logit_bias`, and the
7//! server/CLI-only fields (reasoning budget, grammar triggers) are omitted.
8
9use crate::LLAMA_DEFAULT_SEED;
10
11pub const COMMON_SAMPLER_TYPE_NONE: common_sampler_type = 0;
12pub const COMMON_SAMPLER_TYPE_DRY: common_sampler_type = 1;
13pub const COMMON_SAMPLER_TYPE_TOP_K: common_sampler_type = 2;
14pub const COMMON_SAMPLER_TYPE_TOP_P: common_sampler_type = 3;
15pub const COMMON_SAMPLER_TYPE_MIN_P: common_sampler_type = 4;
16// 5 was COMMON_SAMPLER_TYPE_TFS_Z — removed upstream, the slot is left unused.
17pub const COMMON_SAMPLER_TYPE_TYPICAL_P: common_sampler_type = 6;
18pub const COMMON_SAMPLER_TYPE_TEMPERATURE: common_sampler_type = 7;
19pub const COMMON_SAMPLER_TYPE_XTC: common_sampler_type = 8;
20pub const COMMON_SAMPLER_TYPE_INFILL: common_sampler_type = 9;
21pub const COMMON_SAMPLER_TYPE_PENALTIES: common_sampler_type = 10;
22pub const COMMON_SAMPLER_TYPE_TOP_N_SIGMA: common_sampler_type = 11;
23pub const COMMON_SAMPLER_TYPE_ADAPTIVE_P: common_sampler_type = 12;
24pub type common_sampler_type = ::core::ffi::c_uint;
25
26/// common sampler params
27#[repr(C)]
28#[derive(Debug, PartialEq)]
29pub struct common_sampler_params {
30    /// the seed used to initialize `llama_sampler`
31    pub seed: u32,
32    /// number of previous tokens to remember
33    pub n_prev: i32,
34    /// if greater than 0, output the probabilities of top `n_probs` tokens.
35    pub n_probs: i32,
36    /// 0 = disabled, otherwise samplers should return at least `min_keep` tokens
37    pub min_keep: i32,
38    /// <= 0 to use vocab size
39    pub top_k: i32,
40    /// 1.0 = disabled
41    pub top_p: f32,
42    /// 0.0 = disabled
43    pub min_p: f32,
44    /// 0.0 = disabled
45    pub xtc_probability: f32,
46    /// > 0.5 disables XTC
47    pub xtc_threshold: f32,
48    /// typical_p, 1.0 = disabled
49    pub typ_p: f32,
50    /// <= 0.0 to sample greedily, 0.0 to not output probabilities
51    pub temp: f32,
52    /// 0.0 = disabled
53    pub dynatemp_range: f32,
54    /// controls how entropy maps to temperature in dynamic temperature sampler
55    pub dynatemp_exponent: f32,
56    /// last n tokens to penalize (0 = disable penalty)
57    pub penalty_last_n: i32,
58    /// 1.0 = disabled
59    pub penalty_repeat: f32,
60    /// 0.0 = disabled
61    pub penalty_freq: f32,
62    /// 0.0 = disabled
63    pub penalty_present: f32,
64    /// 0.0 = disabled;      DRY repetition penalty for tokens extending repetition:
65    pub dry_multiplier: f32,
66    /// 0.0 = disabled;      multiplier * base ^ (length of sequence before token - allowed length)
67    pub dry_base: f32,
68    /// tokens extending repetitions beyond this receive penalty
69    pub dry_allowed_length: i32,
70    /// how many tokens to scan for repetitions (0 = disable penalty)
71    pub dry_penalty_last_n: i32,
72    /// select tokens near this probability (valid range 0.0 to 1.0; negative = disabled)
73    pub adaptive_target: f32,
74    /// EMA decay for adaptation; history ≈ 1/(1-decay) tokens (0.0 - 0.99)
75    pub adaptive_decay: f32,
76    /// 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
77    pub mirostat: i32,
78    /// -1.0 = disabled
79    pub top_n_sigma: f32,
80    /// target entropy
81    pub mirostat_tau: f32,
82    /// learning rate
83    pub mirostat_eta: f32,
84    pub ignore_eos: bool,
85    /// disable performance metrics
86    pub no_perf: bool,
87    /// report timings per token
88    pub timing_per_token: bool,
89    pub dry_sequence_breakers: Vec<String>,
90    pub samplers: Vec<common_sampler_type>,
91    pub grammar: Vec<String>,
92    pub logit_bias: Vec<(i32, f64)>,
93}
94
95impl Default for common_sampler_params {
96    fn default() -> Self {
97        Self {
98            seed: LLAMA_DEFAULT_SEED, // the seed used to initialize llama_sampler
99            n_prev: 64,               // number of previous tokens to remember
100            n_probs: 0, // if greater than 0, output the probabilities of top n_probs tokens.
101            min_keep: 0, // 0 = disabled, otherwise samplers should return at least min_keep tokens
102            top_k: 40,  // <= 0 to use vocab size
103            top_p: 0.95, // 1.0 = disabled
104            min_p: 0.05, // 0.0 = disabled
105            xtc_probability: 0.00, // 0.0 = disabled
106            xtc_threshold: 0.10, // > 0.5 disables XTC
107            typ_p: 1.00, // typical_p, 1.0 = disabled
108            temp: 0.80, // <= 0.0 to sample greedily, 0.0 to not output probabilities
109            dynatemp_range: 0.00, // 0.0 = disabled
110            dynatemp_exponent: 1.00, // controls how entropy maps to temperature in dynamic temperature sampler
111            penalty_last_n: 64,      // last n tokens to penalize (0 = disable penalty)
112            penalty_repeat: 1.00,    // 1.0 = disabled
113            penalty_freq: 0.00,      // 0.0 = disabled
114            penalty_present: 0.00,   // 0.0 = disabled
115            dry_multiplier: 0.0, // 0.0 = disabled;      DRY repetition penalty for tokens extending repetition:
116            dry_base: 1.75, // 0.0 = disabled;      multiplier * base ^ (length of sequence before token - allowed length)
117            dry_allowed_length: 2, // tokens extending repetitions beyond this receive penalty
118            dry_penalty_last_n: 64, // how many tokens to scan for repetitions (0 = disable penalty)
119            adaptive_target: -1.0, // select tokens near this probability (negative = disabled)
120            adaptive_decay: 0.90, // EMA decay for adaptation; history ≈ 1/(1-decay) tokens
121            mirostat: 0,    // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
122            top_n_sigma: -1.00, // -1.0 = disabled
123            mirostat_tau: 5.00, // target entropy
124            mirostat_eta: 0.10, // learning rate
125            ignore_eos: false,
126            no_perf: false,          // disable performance metrics
127            timing_per_token: false, // report timings per token
128
129            dry_sequence_breakers: vec!["\n".into(), ":".into(), "\"".into(), "*".into()], // default sequence breakers for DRY
130
131            samplers: vec![
132                COMMON_SAMPLER_TYPE_PENALTIES,
133                COMMON_SAMPLER_TYPE_DRY,
134                COMMON_SAMPLER_TYPE_TOP_N_SIGMA,
135                COMMON_SAMPLER_TYPE_TOP_K,
136                COMMON_SAMPLER_TYPE_TYPICAL_P,
137                COMMON_SAMPLER_TYPE_TOP_P,
138                COMMON_SAMPLER_TYPE_MIN_P,
139                COMMON_SAMPLER_TYPE_XTC,
140                COMMON_SAMPLER_TYPE_TEMPERATURE,
141            ],
142
143            grammar: vec![], // optional BNF-like grammar to constrain sampling
144
145            logit_bias: vec![], // logit biases to apply
146        }
147    }
148}
llama_cpp_sys_4/common.rs

llama_cpp_sys_4/
common.rs
