prism-inference 0.1.17

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
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274

//! Safe baseline NEON for the exact pinned GDN tree and checked row update.

use fearless_simd::{Neon, f32x4, prelude::*};

use crate::Result;

/// Geometry is checked by the parent before any load. The fixed four-register
/// tree is the scalar reference's tree, including adjacent horizontal pairs.
#[inline]
#[allow(
    clippy::suboptimal_flops,
    reason = "Pinned scalar remainder retains separate multiply/add rounding."
)]
pub(super) fn dot(first: &[f32], second: &[f32]) -> Option<f32> {
    let neon = fearless_simd::Level::baseline().as_neon()?;
    let vectorized = first.len() / 16 * 16;
    let mut s0 = f32x4::splat(neon, 0.0);
    let mut s1 = f32x4::splat(neon, 0.0);
    let mut s2 = f32x4::splat(neon, 0.0);
    let mut s3 = f32x4::splat(neon, 0.0);
    for (first, second) in first[..vectorized]
        .as_chunks::<16>()
        .0
        .iter()
        .zip(second[..vectorized].as_chunks::<16>().0)
    {
        s0 = f32x4::from_slice(neon, &first[0..4])
            .mul_add_precise(f32x4::from_slice(neon, &second[0..4]), s0);
        s1 = f32x4::from_slice(neon, &first[4..8])
            .mul_add_precise(f32x4::from_slice(neon, &second[4..8]), s1);
        s2 = f32x4::from_slice(neon, &first[8..12])
            .mul_add_precise(f32x4::from_slice(neon, &second[8..12]), s2);
        s3 = f32x4::from_slice(neon, &first[12..16])
            .mul_add_precise(f32x4::from_slice(neon, &second[12..16]), s3);
    }
    let mut result = ((s0 + s2) + (s1 + s3)).reduce_sum();
    for (&first, &second) in first[vectorized..].iter().zip(&second[vectorized..]) {
        result += first * second;
    }
    Some(result)
}

/// Scales `state` by `decay` and accumulates its four-register FMA dot product with `key` in one pass.
#[cfg(test)]
#[inline]
#[allow(
    clippy::suboptimal_flops,
    reason = "Pinned scalar remainder retains separate multiply/add rounding."
)]
pub(super) fn decay_and_dot(state: &mut [f32], key: &[f32], decay: f32) -> Option<f32> {
    let neon = fearless_simd::Level::baseline().as_neon()?;
    let vectorized = state.len() / 16 * 16;
    let decays = f32x4::splat(neon, decay);
    let mut s0 = f32x4::splat(neon, 0.0);
    let mut s1 = f32x4::splat(neon, 0.0);
    let mut s2 = f32x4::splat(neon, 0.0);
    let mut s3 = f32x4::splat(neon, 0.0);
    for (state, key) in state[..vectorized]
        .as_chunks_mut::<16>()
        .0
        .iter_mut()
        .zip(key[..vectorized].as_chunks::<16>().0)
    {
        let d0 = f32x4::from_slice(neon, &state[0..4]) * decays;
        let d1 = f32x4::from_slice(neon, &state[4..8]) * decays;
        let d2 = f32x4::from_slice(neon, &state[8..12]) * decays;
        let d3 = f32x4::from_slice(neon, &state[12..16]) * decays;
        d0.store_slice(&mut state[0..4]);
        d1.store_slice(&mut state[4..8]);
        d2.store_slice(&mut state[8..12]);
        d3.store_slice(&mut state[12..16]);
        s0 = d0.mul_add_precise(f32x4::from_slice(neon, &key[0..4]), s0);
        s1 = d1.mul_add_precise(f32x4::from_slice(neon, &key[4..8]), s1);
        s2 = d2.mul_add_precise(f32x4::from_slice(neon, &key[8..12]), s2);
        s3 = d3.mul_add_precise(f32x4::from_slice(neon, &key[12..16]), s3);
    }
    let mut result = ((s0 + s2) + (s1 + s3)).reduce_sum();
    for (state, &key) in state[vectorized..].iter_mut().zip(&key[vectorized..]) {
        *state *= decay;
        result += *state * key;
    }
    Some(result)
}

/// Scales `previous` by `decay` into `state` and accumulates its four-register FMA dot product with `key` in one pass.
#[inline]
#[allow(
    clippy::suboptimal_flops,
    reason = "Pinned scalar remainder retains separate multiply/add rounding."
)]
pub(super) fn decay_and_dot_from(
    previous: &[f32],
    state: &mut [f32],
    key: &[f32],
    decay: f32,
) -> Option<f32> {
    let neon = fearless_simd::Level::baseline().as_neon()?;
    let vectorized = state.len() / 16 * 16;
    let decays = f32x4::splat(neon, decay);
    let mut s0 = f32x4::splat(neon, 0.0);
    let mut s1 = f32x4::splat(neon, 0.0);
    let mut s2 = f32x4::splat(neon, 0.0);
    let mut s3 = f32x4::splat(neon, 0.0);
    for ((prev, state), key) in previous[..vectorized]
        .as_chunks::<16>()
        .0
        .iter()
        .zip(state[..vectorized].as_chunks_mut::<16>().0)
        .zip(key[..vectorized].as_chunks::<16>().0)
    {
        let d0 = f32x4::from_slice(neon, &prev[0..4]) * decays;
        let d1 = f32x4::from_slice(neon, &prev[4..8]) * decays;
        let d2 = f32x4::from_slice(neon, &prev[8..12]) * decays;
        let d3 = f32x4::from_slice(neon, &prev[12..16]) * decays;
        d0.store_slice(&mut state[0..4]);
        d1.store_slice(&mut state[4..8]);
        d2.store_slice(&mut state[8..12]);
        d3.store_slice(&mut state[12..16]);
        s0 = d0.mul_add_precise(f32x4::from_slice(neon, &key[0..4]), s0);
        s1 = d1.mul_add_precise(f32x4::from_slice(neon, &key[4..8]), s1);
        s2 = d2.mul_add_precise(f32x4::from_slice(neon, &key[8..12]), s2);
        s3 = d3.mul_add_precise(f32x4::from_slice(neon, &key[12..16]), s3);
    }
    let mut result = ((s0 + s2) + (s1 + s3)).reduce_sum();
    for ((&prev, state), &key) in previous[vectorized..]
        .iter()
        .zip(&mut state[vectorized..])
        .zip(&key[vectorized..])
    {
        let decayed = prev * decay;
        *state = decayed;
        result += decayed * key;
    }
    Some(result)
}

/// The parent checks lengths; every safe load/store receives exactly four lanes.
/// A nonfinite vector is replayed in scalar lane order before any vector store,
/// preserving the original finite-prefix mutation and untouched failing lane.
#[cfg(test)]
#[inline]
pub(super) fn add_scaled(target: &mut [f32], input: &[f32], scale: f32) -> Option<Result<()>> {
    let neon = fearless_simd::Level::baseline().as_neon()?;
    Some(add_scaled_with_neon(neon, target, input, scale))
}

/// Updates `target` with `scale * key` and accumulates its dot product with `query` in one pass.
#[inline]
pub(super) fn update_and_dot(
    target: &mut [f32],
    key: &[f32],
    scale: f32,
    query: &[f32],
) -> Option<Result<f32>> {
    let neon = fearless_simd::Level::baseline().as_neon()?;
    Some(update_and_dot_with_neon(neon, target, key, scale, query))
}

#[cfg(test)]
#[inline]
#[allow(
    clippy::suboptimal_flops,
    reason = "Pinned scalar remainder retains separate multiply/add rounding."
)]
fn add_scaled_with_neon(neon: Neon, target: &mut [f32], input: &[f32], scale: f32) -> Result<()> {
    let vectorized = target.len() / 16 * 16;
    let scales = f32x4::splat(neon, scale);
    for (target_block, input_block) in target[..vectorized]
        .as_chunks_mut::<16>()
        .0
        .iter_mut()
        .zip(input[..vectorized].as_chunks::<16>().0)
    {
        let u0 = scales.mul_add_precise(
            f32x4::from_slice(neon, &input_block[0..4]),
            f32x4::from_slice(neon, &target_block[0..4]),
        );
        let u1 = scales.mul_add_precise(
            f32x4::from_slice(neon, &input_block[4..8]),
            f32x4::from_slice(neon, &target_block[4..8]),
        );
        let u2 = scales.mul_add_precise(
            f32x4::from_slice(neon, &input_block[8..12]),
            f32x4::from_slice(neon, &target_block[8..12]),
        );
        let u3 = scales.mul_add_precise(
            f32x4::from_slice(neon, &input_block[12..16]),
            f32x4::from_slice(neon, &target_block[12..16]),
        );
        if (u0.is_finite() & u1.is_finite() & u2.is_finite() & u3.is_finite()).all_true() {
            u0.store_slice(&mut target_block[0..4]);
            u1.store_slice(&mut target_block[4..8]);
            u2.store_slice(&mut target_block[8..12]);
            u3.store_slice(&mut target_block[12..16]);
        } else {
            for (target, &input) in target_block.iter_mut().zip(input_block) {
                *target = super::super::finite(scale.mul_add(input, *target))?;
            }
        }
    }
    for (target, &input) in target[vectorized..].iter_mut().zip(&input[vectorized..]) {
        *target = super::super::finite(*target + scale * input)?;
    }
    Ok(())
}

#[inline]
#[allow(
    clippy::suboptimal_flops,
    reason = "Pinned scalar remainder retains separate multiply/add rounding."
)]
fn update_and_dot_with_neon(
    neon: Neon,
    target: &mut [f32],
    key: &[f32],
    scale: f32,
    query: &[f32],
) -> Result<f32> {
    let vectorized = target.len() / 16 * 16;
    let scales = f32x4::splat(neon, scale);
    let mut s0 = f32x4::splat(neon, 0.0);
    let mut s1 = f32x4::splat(neon, 0.0);
    let mut s2 = f32x4::splat(neon, 0.0);
    let mut s3 = f32x4::splat(neon, 0.0);
    for ((target_block, key_block), query_block) in target[..vectorized]
        .as_chunks_mut::<16>()
        .0
        .iter_mut()
        .zip(key[..vectorized].as_chunks::<16>().0)
        .zip(query[..vectorized].as_chunks::<16>().0)
    {
        let u0 = scales.mul_add_precise(
            f32x4::from_slice(neon, &key_block[0..4]),
            f32x4::from_slice(neon, &target_block[0..4]),
        );
        let u1 = scales.mul_add_precise(
            f32x4::from_slice(neon, &key_block[4..8]),
            f32x4::from_slice(neon, &target_block[4..8]),
        );
        let u2 = scales.mul_add_precise(
            f32x4::from_slice(neon, &key_block[8..12]),
            f32x4::from_slice(neon, &target_block[8..12]),
        );
        let u3 = scales.mul_add_precise(
            f32x4::from_slice(neon, &key_block[12..16]),
            f32x4::from_slice(neon, &target_block[12..16]),
        );
        if (u0.is_finite() & u1.is_finite() & u2.is_finite() & u3.is_finite()).all_true() {
            u0.store_slice(&mut target_block[0..4]);
            u1.store_slice(&mut target_block[4..8]);
            u2.store_slice(&mut target_block[8..12]);
            u3.store_slice(&mut target_block[12..16]);
            s0 = u0.mul_add_precise(f32x4::from_slice(neon, &query_block[0..4]), s0);
            s1 = u1.mul_add_precise(f32x4::from_slice(neon, &query_block[4..8]), s1);
            s2 = u2.mul_add_precise(f32x4::from_slice(neon, &query_block[8..12]), s2);
            s3 = u3.mul_add_precise(f32x4::from_slice(neon, &query_block[12..16]), s3);
        } else {
            for (target_elem, &key_elem) in target_block.iter_mut().zip(key_block) {
                *target_elem = super::super::finite(scale.mul_add(key_elem, *target_elem))?;
            }
        }
    }
    let mut result = ((s0 + s2) + (s1 + s3)).reduce_sum();
    for ((target_elem, &key_elem), &query_elem) in target[vectorized..]
        .iter_mut()
        .zip(&key[vectorized..])
        .zip(&query[vectorized..])
    {
        let updated = super::super::finite(*target_elem + scale * key_elem)?;
        *target_elem = updated;
        result += updated * query_elem;
    }
    super::super::finite(result)
}