oxinfer 0.0.1

Local LLM inference in pure Rust (Gemma 4, Qwen 3.5/3.8): SIMD CPU kernels everywhere, GPUs detected at run time, and a WebAssembly + WebGPU browser build.
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

//! Token sampling: greedy, or temperature + top-k + top-p.

pub struct Sampler {
    pub temperature: f32,
    pub top_k: usize,
    pub top_p: f32,
    rng: u64,
}

impl Sampler {
    pub fn new(temperature: f32, top_k: usize, top_p: f32, seed: u64) -> Self {
        Self {
            temperature,
            top_k,
            top_p,
            rng: seed.max(1),
        }
    }

    fn next_f32(&mut self) -> f32 {
        // xorshift64*
        let mut x = self.rng;
        x ^= x >> 12;
        x ^= x << 25;
        x ^= x >> 27;
        self.rng = x;
        let r = x.wrapping_mul(0x2545_F491_4F6C_DD1D);
        (r >> 40) as f32 / (1u64 << 24) as f32
    }

    pub fn sample(&mut self, logits: &[f32]) -> u32 {
        if self.temperature <= 0.0 {
            return argmax(logits);
        }
        let k = if self.top_k == 0 {
            logits.len()
        } else {
            self.top_k.min(logits.len())
        };
        // Partial selection of the top-k indices.
        let mut idx: Vec<u32> = (0..logits.len() as u32).collect();
        idx.select_nth_unstable_by(k - 1, |&a, &b| {
            logits[b as usize]
                .partial_cmp(&logits[a as usize])
                .unwrap_or(std::cmp::Ordering::Equal)
        });
        idx.truncate(k);
        idx.sort_unstable_by(|&a, &b| {
            logits[b as usize]
                .partial_cmp(&logits[a as usize])
                .unwrap_or(std::cmp::Ordering::Equal)
        });
        let inv_t = 1.0 / self.temperature;
        let max = logits[idx[0] as usize];
        let mut probs: Vec<f32> = idx.iter().map(|&i| ((logits[i as usize] - max) * inv_t).exp()).collect();
        let sum: f32 = probs.iter().sum();
        for p in probs.iter_mut() {
            *p /= sum;
        }
        // top-p
        let mut cut = probs.len();
        if self.top_p < 1.0 {
            let mut acc = 0f32;
            for (i, &p) in probs.iter().enumerate() {
                acc += p;
                if acc >= self.top_p {
                    cut = i + 1;
                    break;
                }
            }
        }
        let total: f32 = probs[..cut].iter().sum();
        let mut r = self.next_f32() * total;
        for i in 0..cut {
            r -= probs[i];
            if r <= 0.0 {
                return idx[i];
            }
        }
        idx[cut - 1]
    }
}

pub fn argmax(x: &[f32]) -> u32 {
    let mut best = 0usize;
    let mut bv = f32::NEG_INFINITY;
    for (i, &v) in x.iter().enumerate() {
        if v > bv {
            bv = v;
            best = i;
        }
    }
    best as u32
}