oxinfer 0.0.5

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
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

//! The image-token decoder of Janus (LlamaGen's VQ-16, `janus/models/vq_model.py`): codebook
//! lookup (16384 l2-normalized 8-dim codes), a 1x1 `post_quant_conv`, then a taming-style
//! convolutional decoder with 16x upsampling (24x24 tokens -> 384x384 RGB). Built from the Stable
//! Diffusion blocks (`crate::sd::ops`), so its convolutions run on the GPU backends too.

use crate::sd::ops::{Conv, GroupNorm, Linear, Pre, Tensors, add_inplace, attention, upsample2x};
use crate::sd::unet::Resnet;
use anyhow::{Result, bail};

/// A single-head spatial self-attention block (`AttnBlock`) with its residual.
struct AttnBlock {
    norm: GroupNorm,
    q: Linear,
    k: Linear,
    v: Linear,
    proj: Linear,
}

impl AttnBlock {
    fn load(t: &Tensors, p: &str) -> Result<Self> {
        Ok(Self {
            norm: GroupNorm::load(t, &format!("{p}.norm"), 32, 1e-6)?,
            q: Linear::load(t, &format!("{p}.q"))?,
            k: Linear::load(t, &format!("{p}.k"))?,
            v: Linear::load(t, &format!("{p}.v"))?,
            proj: Linear::load(t, &format!("{p}.proj_out"))?,
        })
    }

    fn forward(&self, x: &mut [f32], n: usize) {
        let pre = Pre::Group(&self.norm, false);
        let (q, k, v) = if self.q.on_device(n) {
            (
                self.q.forward_pre(x, n, pre),
                self.k.forward_pre(x, n, pre),
                self.v.forward_pre(x, n, pre),
            )
        } else {
            let y = pre.apply(x, n);
            (self.q.forward(&y, n), self.k.forward(&y, n), self.v.forward(&y, n))
        };
        let a = attention(&q, &k, &v, n, n, 1, self.q.n_out, false);
        add_inplace(x, &self.proj.forward(&a, n));
    }
}

struct Level {
    res: Vec<Resnet>,
    /// One per resnet at the lowest resolution, none elsewhere.
    attn: Vec<AttnBlock>,
    up: Option<Conv>,
}

pub struct VqDecoder {
    /// l2-normalized codebook `[codes, dim]`.
    codebook: Vec<f32>,
    pub codes: usize,
    pub dim: usize,
    post_quant: Linear,
    conv_in: Conv,
    mid: (Resnet, AttnBlock, Resnet),
    levels: Vec<Level>,
    norm_out: GroupNorm,
    conv_out: Conv,
}

impl VqDecoder {
    /// The decoder under `p` (`gen_vision_model`) of `t`.
    pub fn load(t: &Tensors, p: &str) -> Result<Self> {
        let cb_name = format!("{p}.quantize.embedding.weight");
        let shape = t.shape(&cb_name)?;
        let [codes, dim] = shape[..] else {
            bail!("{cb_name}: expected a matrix, got {shape:?}")
        };
        let mut codebook = t.f32(&cb_name)?;
        // F.normalize(p=2, dim=-1): x / max(|x|, 1e-12).
        for r in codebook.chunks_mut(dim) {
            let n = r.iter().map(|v| v * v).sum::<f32>().sqrt().max(1e-12);
            r.iter_mut().for_each(|v| *v /= n);
        }
        let d = format!("{p}.decoder");
        let mut levels = Vec::new();
        while t.has(&format!("{d}.conv_blocks.{}.res.0.conv1.weight", levels.len())) {
            let b = format!("{d}.conv_blocks.{}", levels.len());
            let mut res = Vec::new();
            while t.has(&format!("{b}.res.{}.conv1.weight", res.len())) {
                res.push(Resnet::load(t, &format!("{b}.res.{}", res.len()), 1e-6)?);
            }
            let mut attn = Vec::new();
            while t.has(&format!("{b}.attn.{}.q.weight", attn.len())) {
                attn.push(AttnBlock::load(t, &format!("{b}.attn.{}", attn.len()))?);
            }
            if !attn.is_empty() && attn.len() != res.len() {
                bail!("{b}: {} attention blocks for {} resnets", attn.len(), res.len());
            }
            let up = if t.has(&format!("{b}.upsample.conv.weight")) {
                Some(Conv::load(t, &format!("{b}.upsample.conv"))?)
            } else {
                None
            };
            levels.push(Level { res, attn, up });
        }
        if levels.is_empty() {
            bail!("{d}: no conv_blocks");
        }
        let post_quant = Linear::load(t, &format!("{p}.post_quant_conv"))?;
        if post_quant.n_in != dim {
            bail!("{p}.post_quant_conv takes {} channels, the codebook has {dim}", post_quant.n_in);
        }
        Ok(Self {
            codebook,
            codes,
            dim,
            post_quant,
            conv_in: Conv::load(t, &format!("{d}.conv_in"))?,
            mid: (
                Resnet::load(t, &format!("{d}.mid.0"), 1e-6)?,
                AttnBlock::load(t, &format!("{d}.mid.1"))?,
                Resnet::load(t, &format!("{d}.mid.2"), 1e-6)?,
            ),
            levels,
            norm_out: GroupNorm::load(t, &format!("{d}.norm_out"), 32, 1e-6)?,
            conv_out: Conv::load(t, &format!("{d}.conv_out"))?,
        })
    }

    /// The upsampling factor (16 for VQ-16).
    pub fn scale(&self) -> usize {
        1 << self.levels.iter().filter(|l| l.up.is_some()).count()
    }

    /// Image tokens (row-major `gh` x `gw` grid) to RGB in [-1, 1], channels-last
    /// (`[height * width, 3]`); returns `(pixels, height, width)`.
    pub fn decode(&self, tokens: &[u32], gh: usize, gw: usize) -> Result<(Vec<f32>, usize, usize)> {
        if tokens.len() != gh * gw {
            bail!("{} image tokens for a {gh}x{gw} grid", tokens.len());
        }
        let dim = self.dim;
        let mut z = vec![0f32; tokens.len() * dim];
        for (r, &tok) in z.chunks_mut(dim).zip(tokens) {
            let tok = tok as usize;
            if tok >= self.codes {
                bail!("image token {tok} is outside the codebook ({} codes)", self.codes);
            }
            r.copy_from_slice(&self.codebook[tok * dim..(tok + 1) * dim]);
        }
        let z = self.post_quant.forward(&z, gh * gw);
        let (mut x, mut h, mut w) = self.conv_in.forward(&z, gh, gw, 1);
        x = self.mid.0.forward(&x, h, w, None);
        self.mid.1.forward(&mut x, h * w);
        x = self.mid.2.forward(&x, h, w, None);
        for l in &self.levels {
            for (i, r) in l.res.iter().enumerate() {
                x = r.forward(&x, h, w, None);
                if let Some(a) = l.attn.get(i) {
                    a.forward(&mut x, h * w);
                }
            }
            if let Some(u) = &l.up {
                let c = x.len() / (h * w);
                let up = upsample2x(&x, h, w, c);
                (h, w) = (h * 2, w * 2);
                x = u.forward(&up, h, w, 1).0;
            }
        }
        let (y, _, _) = self.conv_out.forward_pre(&x, h, w, 1, Pre::Group(&self.norm_out, true));
        Ok((y, h, w))
    }
}

/// RGB in [-1, 1] to bytes as the reference does: `clip((x + 1) / 2 * 255, 0, 255)` stored
/// into uint8 (truncation).
pub fn to_rgb8(x: &[f32]) -> Vec<u8> {
    x.iter().map(|&v| ((v + 1.0) / 2.0 * 255.0).clamp(0.0, 255.0) as u8).collect()
}