1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171

//! Safe wrapper around `llama_context`.

use std::fmt::{Debug, Formatter};
use std::num::NonZeroI32;

use crate::llama_backend::LlamaBackend;
use crate::llama_batch::LlamaBatch;
use crate::model::LlamaModel;
use crate::timing::LlamaTimings;
use crate::token::data::LlamaTokenData;
use crate::token::LlamaToken;
use crate::{DecodeError, LlamaContextLoadError};
use params::LlamaContextParams;
use std::os::raw::c_int;
use std::ptr::NonNull;
use std::slice;

pub mod kv_cache;
pub mod params;
pub mod sample;

/// Safe wrapper around `llama_context`.
#[allow(clippy::module_name_repetitions)]
pub struct LlamaContext<'a> {
    pub(crate) context: NonNull<llama_cpp_sys_2::llama_context>,
    /// a reference to the contexts model.
    pub model: &'a LlamaModel,
    initialized_logits: Vec<i32>,
}

impl Debug for LlamaContext<'_> {
    fn fmt(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("LlamaContext")
            .field("context", &self.context)
            .finish()
    }
}

impl<'model> LlamaContext<'model> {
    pub(crate) fn new(
        llama_model: &'model LlamaModel,
        llama_context: NonNull<llama_cpp_sys_2::llama_context>,
    ) -> Self {
        Self {
            context: llama_context,
            model: llama_model,
            initialized_logits: Vec::new(),
        }
    }

    /// Gets the size of the context.
    #[must_use]
    pub fn n_ctx(&self) -> c_int {
        unsafe { llama_cpp_sys_2::llama_n_ctx(self.context.as_ptr()) }
    }

    /// Decodes the batch.
    ///
    /// # Errors
    ///
    /// - `DecodeError` if the decoding failed.
    ///
    /// # Panics
    ///
    /// - the returned [`c_int`] from llama-cpp does not fit into a i32 (this should never happen on most systems)
    pub fn decode(&mut self, batch: &mut LlamaBatch) -> Result<(), DecodeError> {
        let result =
            unsafe { llama_cpp_sys_2::llama_decode(self.context.as_ptr(), batch.llama_batch) };

        match NonZeroI32::new(result as i32) {
            None => {
                self.initialized_logits = batch.initialized_logits.clone();
                Ok(())
            }
            Some(error) => Err(DecodeError::from(error)),
        }
    }

    /// Get the logits for the ith token in the context.
    ///
    /// # Panics
    ///
    /// - logit `i` is not initialized.
    pub fn candidates_ith(&self, i: i32) -> impl Iterator<Item = LlamaTokenData> + '_ {
        assert!(
            self.initialized_logits.contains(&i),
            "logit {i} is not initialized. only {:?} is",
            self.initialized_logits
        );
        (0_i32..).zip(self.get_logits_ith(i)).map(|(i, logit)| {
            let token = LlamaToken::new(i);
            LlamaTokenData::new(token, *logit, 0_f32)
        })
    }

    /// Get the logits for the ith token in the context.
    ///
    /// # Panics
    ///
    /// - `i` is greater than `n_ctx`
    /// - `n_vocab` does not fit into a usize
    #[must_use]
    pub fn get_logits_ith(&self, i: i32) -> &[f32] {
        assert!(
            self.n_ctx() > i,
            "n_ctx ({}) must be greater than i ({})",
            self.n_ctx(),
            i
        );

        let data = unsafe { llama_cpp_sys_2::llama_get_logits_ith(self.context.as_ptr(), i) };
        let len = usize::try_from(self.model.n_vocab()).expect("n_vocab does not fit into a usize");

        unsafe { slice::from_raw_parts(data, len) }
    }

    /// Reset the timings for the context.
    pub fn reset_timings(&mut self) {
        unsafe { llama_cpp_sys_2::llama_reset_timings(self.context.as_ptr()) }
    }

    /// Rows: `n_tokens`
    /// Cols: `n_vocab`
    ///
    /// # Panics
    ///
    /// - `n_vocab` does not fit into a usize
    pub fn logits_mut(&mut self, n_tokens: usize) -> &mut [f32] {
        let logits_ptr = unsafe { llama_cpp_sys_2::llama_get_logits(self.context.as_ptr()) };

        let n_vocab = usize::try_from(self.model.n_vocab()).expect("n_vocab should be positive");
        unsafe { slice::from_raw_parts_mut(logits_ptr, n_vocab * n_tokens) }
    }

    /// get the logits
    ///
    /// # Panics
    ///
    /// - `n_vocab` does not fit into a usize
    #[deprecated]
    #[must_use]
    pub fn logits(&self, n_tokens: usize) -> &[f32] {
        let n_vocab = usize::try_from(self.model.n_vocab()).expect("n_vocab should be positive");
        let logits_ptr = unsafe { llama_cpp_sys_2::llama_get_logits(self.context.as_ptr()) };

        unsafe { slice::from_raw_parts(logits_ptr, n_vocab * n_tokens) }
    }

    /// Returns the timings for the context.
    pub fn timings(&mut self) -> LlamaTimings {
        let timings = unsafe { llama_cpp_sys_2::llama_get_timings(self.context.as_ptr()) };
        LlamaTimings { timings }
    }

    /// Create a new `LlamaContext` from a model.
    #[deprecated(note = "use `Model::new_context` instead")]
    #[tracing::instrument(skip_all)]
    pub fn new_with_model(
        backend: &LlamaBackend,
        model: &'model mut LlamaModel,
        context_params: &LlamaContextParams,
    ) -> Result<Self, LlamaContextLoadError> {
        model.new_context(backend, context_params)
    }
}

impl Drop for LlamaContext<'_> {
    fn drop(&mut self) {
        unsafe { llama_cpp_sys_2::llama_free(self.context.as_ptr()) }
    }
}