native_neural_network_std 0.2.1

Ergonomic std wrapper for the `native_neural_network` crate (no_std) — std-friendly re-exports and utilities.
Documentation 
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

pub struct KvCacheStd {
    pub key: Vec<f32>,
    pub value: Vec<f32>,
    pub max_tokens: usize,
    pub head_dim: usize,
    pub num_heads: usize,
    pub used_tokens: usize,
}

#[derive(Debug)]
pub enum KvCacheError {
    ShapeMismatch,
    CapacityExceeded,
}

impl core::fmt::Display for KvCacheError {
    fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
        match self {
            KvCacheError::ShapeMismatch => write!(f, "shape mismatch"),
            KvCacheError::CapacityExceeded => write!(f, "capacity exceeded"),
        }
    }
}

impl std::error::Error for KvCacheError {}

impl KvCacheStd {
    pub fn new(max_tokens: usize, head_dim: usize, num_heads: usize) -> Self {
        let stride = head_dim
            .checked_mul(num_heads)
            .expect("kv_cache: head_dim * num_heads overflow");
        let total = max_tokens
            .checked_mul(stride)
            .expect("kv_cache: max_tokens * stride overflow");
        Self {
            key: vec![0f32; total],
            value: vec![0f32; total],
            max_tokens,
            head_dim,
            num_heads,
            used_tokens: 0,
        }
    }

    fn token_stride(&self) -> Option<usize> {
        self.num_heads.checked_mul(self.head_dim)
    }

    pub fn append_token(
        &mut self,
        key_token: &[f32],
        value_token: &[f32],
    ) -> Result<(), KvCacheError> {
        let mut view = native_neural_network::kv_cache::KvCacheView {
            key: self.key.as_mut_slice(),
            value: self.value.as_mut_slice(),
            max_tokens: self.max_tokens,
            head_dim: self.head_dim,
            num_heads: self.num_heads,
            used_tokens: self.used_tokens,
        };
        match view.append_token(key_token, value_token) {
            Ok(()) => {
                self.used_tokens = view.used_tokens;
                Ok(())
            }
            Err(native_neural_network::kv_cache::KvCacheError::ShapeMismatch) => {
                Err(KvCacheError::ShapeMismatch)
            }
            Err(native_neural_network::kv_cache::KvCacheError::CapacityExceeded) => {
                Err(KvCacheError::CapacityExceeded)
            }
        }
    }

    pub fn len_tokens(&self) -> usize {
        self.used_tokens
    }

    pub fn is_empty(&self) -> bool {
        self.used_tokens == 0
    }

    pub fn is_full(&self) -> bool {
        self.used_tokens >= self.max_tokens
    }

    pub fn remaining_tokens(&self) -> usize {
        self.max_tokens.saturating_sub(self.used_tokens)
    }

    pub fn clear(&mut self) {
        self.used_tokens = 0;
    }

    pub fn token_slices(&self, token_index: usize) -> Result<(&[f32], &[f32]), KvCacheError> {
        let stride = self.token_stride().ok_or(KvCacheError::ShapeMismatch)?;
        if token_index >= self.used_tokens {
            return Err(KvCacheError::CapacityExceeded);
        }
        let off = token_index
            .checked_mul(stride)
            .ok_or(KvCacheError::ShapeMismatch)?;
        let end = off.checked_add(stride).ok_or(KvCacheError::ShapeMismatch)?;
        if end > self.key.len() || end > self.value.len() {
            return Err(KvCacheError::ShapeMismatch);
        }
        Ok((&self.key[off..end], &self.value[off..end]))
    }
}

impl core::fmt::Debug for KvCacheStd {
    fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
        f.debug_struct("KvCacheStd")
            .field("max_tokens", &self.max_tokens)
            .field("used_tokens", &self.used_tokens)
            .finish()
    }
}

pub struct KvCacheViewStd<'a> {
    pub key: &'a mut [f32],
    pub value: &'a mut [f32],
    pub max_tokens: usize,
    pub head_dim: usize,
    pub num_heads: usize,
    pub used_tokens: usize,
}

impl<'a> KvCacheViewStd<'a> {
    pub fn token_slices(&self, token_index: usize) -> Result<(&[f32], &[f32]), KvCacheError> {
        if token_index >= self.used_tokens {
            return Err(KvCacheError::CapacityExceeded);
        }
        let stride = self
            .num_heads
            .checked_mul(self.head_dim)
            .ok_or(KvCacheError::ShapeMismatch)?;
        let off = token_index
            .checked_mul(stride)
            .ok_or(KvCacheError::ShapeMismatch)?;
        let end = off.checked_add(stride).ok_or(KvCacheError::ShapeMismatch)?;
        if end > self.key.len() || end > self.value.len() {
            return Err(KvCacheError::ShapeMismatch);
        }
        Ok((&self.key[off..end], &self.value[off..end]))
    }
}

pub type KvCacheView<'a> = KvCacheViewStd<'a>;

fn kv_cache_init_hook() -> Result<(), crate::InitError> {
    Ok(())
}

pub fn register_module_init_hooks() -> Result<(), crate::RegisterError> {
    crate::register_init_hook(
        "kv_cache",
        crate::InitSubsystem::KvCache,
        kv_cache_init_hook,
    )
}