prism-inference 0.1.15

Bounded Rust CPU inference and tokenization for Prism-compatible GGUF models
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

//! Exact integer reductions over one decoded block shared by a prompt tile.

use super::super::{MatvecBackend, q8_native_lanes};

#[cfg(all(
    feature = "experimental-neon",
    target_arch = "aarch64",
    target_os = "windows",
    target_env = "msvc",
    target_feature = "neon"
))]
#[path = "prefill_neon.rs"]
mod neon;

/// Exactly one block, scoped to the packed row loop; no heap allocation.
pub(super) enum Block {
    Scalar([i16; 128]),
    #[cfg(all(
        feature = "experimental-neon",
        target_arch = "aarch64",
        target_os = "windows",
        target_env = "msvc",
        target_feature = "neon"
    ))]
    Neon(neon::Block),
}

impl Block {
    #[allow(
        clippy::missing_const_for_fn,
        reason = "Only the scalar fallback is constant; retain one signature across targets and feature sets."
    )]
    pub(super) fn new(weights: [i16; 128], backend: MatvecBackend) -> Self {
        #[cfg(all(
            feature = "experimental-neon",
            target_arch = "aarch64",
            target_os = "windows",
            target_env = "msvc",
            target_feature = "neon"
        ))]
        if backend != MatvecBackend::Scalar
            && let Some(block) = neon::Block::new(&weights)
        {
            return Self::Neon(block);
        }
        // Unsupported builds retain the scalar result for every request policy.
        let _ = backend;
        Self::Scalar(weights)
    }

    #[inline]
    pub(super) fn dot128(&self, values: &[i8; 128]) -> i32 {
        match self {
            Self::Scalar(weights) => scalar(weights, values),
            #[cfg(all(
                feature = "experimental-neon",
                target_arch = "aarch64",
                target_os = "windows",
                target_env = "msvc",
                target_feature = "neon"
            ))]
            Self::Neon(block) => block.dot128(values),
        }
    }

    #[inline]
    pub(super) fn dot32(&self, group: usize, values: &[i8; 32]) -> i32 {
        match self {
            Self::Scalar(weights) => scalar(&weights[group * 32..group * 32 + 32], values),
            #[cfg(all(
                feature = "experimental-neon",
                target_arch = "aarch64",
                target_os = "windows",
                target_env = "msvc",
                target_feature = "neon"
            ))]
            Self::Neon(block) => block.dot32(group, values),
        }
    }

    #[inline]
    pub(super) fn lanes32(&self, group: usize, values: &[i8; 32]) -> [i32; 4] {
        match self {
            Self::Scalar(weights) => q8_native_lanes(&weights[group * 32..group * 32 + 32], values),
            #[cfg(all(
                feature = "experimental-neon",
                target_arch = "aarch64",
                target_os = "windows",
                target_env = "msvc",
                target_feature = "neon"
            ))]
            Self::Neon(block) => block.lanes32(group, values),
        }
    }
}

fn scalar(weights: &[i16], values: &[i8]) -> i32 {
    weights
        .iter()
        .zip(values)
        .map(|(&weight, &value)| i32::from(weight) * i32::from(value))
        .sum()
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn tiled_integer_reductions_preserve_extreme_values_and_native_lanes() {
        for seed in 0_u8..=255 {
            let weights = std::array::from_fn(|index: usize| {
                i16::from(seed.wrapping_add(index.to_le_bytes()[0]) & 3) - 1
            });
            let values = std::array::from_fn(|index: usize| {
                i8::from_ne_bytes([seed.wrapping_mul(197).wrapping_add(index.to_le_bytes()[0])])
            });
            for backend in [
                MatvecBackend::Scalar,
                MatvecBackend::Auto,
                MatvecBackend::ExperimentalNeon,
            ] {
                let block = Block::new(weights, backend);
                if backend == MatvecBackend::Scalar {
                    assert!(matches!(block, Block::Scalar(_)));
                }
                #[cfg(all(
                    feature = "experimental-neon",
                    target_arch = "aarch64",
                    target_os = "windows",
                    target_env = "msvc",
                    target_feature = "neon"
                ))]
                if backend != MatvecBackend::Scalar {
                    assert!(matches!(block, Block::Neon(_)));
                }
                assert_eq!(block.dot128(&values), scalar(&weights, &values));
                for (group, values) in values.as_chunks::<32>().0.iter().enumerate() {
                    let weights = &weights[group * 32..group * 32 + 32];
                    assert_eq!(block.dot32(group, values), scalar(weights, values));
                    assert_eq!(
                        block.lanes32(group, values),
                        q8_native_lanes(weights, values)
                    );
                }
            }
        }
    }
}