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
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

//! Gemma 4 "assistant" drafter for speculative decoding: a 4-layer, 256-wide decoder whose
//! layers all cross-attend to the main model's shared KV caches. Input is the concatenation of
//! the main model's (scaled) embedding of the last token and the main model's hidden state of
//! the token before it; output is an argmax draft token (over a centroid-selected vocabulary
//! subset) and a projected hidden state to feed the next draft step.

use crate::config::TextConfig;
use crate::kernels::*;
use crate::model::{LayerCache, attention};
use crate::safetensors::ModelFiles;
use anyhow::{Context, Result, bail};
use std::path::Path;

struct DLayer<'a> {
    sliding: bool,
    hd: usize,
    q: &'a [u16],
    o: &'a [u16],
    q_norm: &'a [u16],
    ln_in: &'a [u16],
    ln_post_attn: &'a [u16],
    ln_pre_ffn: &'a [u16],
    ln_post_ffn: &'a [u16],
    gate: &'a [u16],
    up: &'a [u16],
    down: &'a [u16],
    layer_scalar: f32,
}

pub struct Drafter<'a> {
    pub cfg: TextConfig,
    pub backbone_hidden: usize,
    pre_proj: &'a [u16],  // [H, 2*B]
    post_proj: &'a [u16], // [B, H]
    embed: &'a [u16],     // [V, H] (tied LM head)
    centroids: &'a [u16], // [C, H]
    token_ordering: Vec<u32>,
    num_centroids: usize,
    top_k_centroids: usize,
    layers: Vec<DLayer<'a>>,
    final_norm: &'a [u16],
    inv_freq_sliding: Vec<f32>,
    inv_freq_full: Vec<f32>,
}

impl<'a> Drafter<'a> {
    pub fn load(dir: &Path, files: &'a ModelFiles) -> Result<Self> {
        let raw: serde_json::Value = serde_json::from_slice(&crate::fsx::read(&dir.join("config.json"))?)?;
        let backbone_hidden = raw["backbone_hidden_size"].as_u64().context("backbone_hidden_size")? as usize;
        let num_centroids = raw["num_centroids"].as_u64().unwrap_or(2048) as usize;
        let top_k_centroids = raw["centroid_intermediate_top_k"].as_u64().unwrap_or(32) as usize;
        if raw["use_ordered_embeddings"].as_bool() != Some(true) {
            bail!("drafter without ordered embeddings is not supported");
        }
        let cfg = TextConfig::load(dir)?;
        let h = cfg.hidden_size;
        let nh = cfg.num_attention_heads;
        let v = cfg.vocab_size;

        let get = |name: &str, shape: &[usize]| -> Result<&'a [u16]> {
            let info = files.info(name).with_context(|| format!("missing tensor {name}"))?;
            if info.shape != shape {
                bail!("tensor {name} has shape {:?}, expected {:?}", info.shape, shape);
            }
            files.bf16(name)
        };
        let token_ordering: Vec<u32> = {
            let info = files.info("masked_embedding.token_ordering").context("token_ordering")?;
            if info.dtype != "I64" || info.shape != [v] {
                bail!("unexpected token_ordering dtype/shape");
            }
            let bytes = files.bytes("masked_embedding.token_ordering")?;
            bytes.as_chunks::<8>().0.iter().map(|c| i64::from_le_bytes(*c) as u32).collect()
        };
        let mut layers = Vec::new();
        for l in 0..cfg.num_hidden_layers {
            let p = format!("model.layers.{l}.");
            let g = |n: &str, s: &[usize]| get(&format!("{p}{n}"), s);
            let hd = cfg.layer_head_dim(l);
            let layer_scalar = g("layer_scalar", &[1]).map(|s| bf16_to_f32(s[0])).unwrap_or(1.0);
            layers.push(DLayer {
                sliding: cfg.is_sliding(l),
                hd,
                q: g("self_attn.q_proj.weight", &[nh * hd, h])?,
                o: g("self_attn.o_proj.weight", &[h, nh * hd])?,
                q_norm: g("self_attn.q_norm.weight", &[hd])?,
                ln_in: g("input_layernorm.weight", &[h])?,
                ln_post_attn: g("post_attention_layernorm.weight", &[h])?,
                ln_pre_ffn: g("pre_feedforward_layernorm.weight", &[h])?,
                ln_post_ffn: g("post_feedforward_layernorm.weight", &[h])?,
                gate: g("mlp.gate_proj.weight", &[cfg.intermediate_size, h])?,
                up: g("mlp.up_proj.weight", &[cfg.intermediate_size, h])?,
                down: g("mlp.down_proj.weight", &[h, cfg.intermediate_size])?,
                layer_scalar,
            });
        }
        let rs = cfg.rope_spec(true);
        let rf = cfg.rope_spec(false);
        Ok(Self {
            backbone_hidden,
            pre_proj: get("pre_projection.weight", &[h, 2 * backbone_hidden])?,
            post_proj: get("post_projection.weight", &[backbone_hidden, h])?,
            embed: get("model.embed_tokens.weight", &[v, h])?,
            centroids: get("masked_embedding.centroids.weight", &[num_centroids, h])?,
            token_ordering,
            num_centroids,
            top_k_centroids,
            layers,
            final_norm: get("model.norm.weight", &[h])?,
            inv_freq_sliding: rope_inv_freq(cfg.head_dim, rs.theta, rs.proportion),
            inv_freq_full: rope_inv_freq(cfg.global_head_dim, rf.theta, rf.proportion),
            cfg,
        })
    }

    /// One draft step. `tok_emb`: the main model's scaled embedding of the last token; `hidden`:
    /// the main-side hidden state of the token before it; `pos`: the last token's position;
    /// `cache_s`/`cache_g`: the main model's shared sliding/global KV caches (positions `< pos`
    /// are attended). Returns the argmax draft token and the hidden state for the next step.
    pub fn step(&self, tok_emb: &[f32], hidden: &[f32], pos: usize, cache_s: &LayerCache, cache_g: &LayerCache) -> (u32, Vec<f32>) {
        let cfg = &self.cfg;
        let h = cfg.hidden_size;
        let nh = cfg.num_attention_heads;
        let eps = cfg.rms_norm_eps as f32;
        let b = self.backbone_hidden;
        debug_assert_eq!(tok_emb.len(), b);
        debug_assert_eq!(hidden.len(), b);

        let mut input = Vec::with_capacity(2 * b);
        input.extend_from_slice(tok_emb);
        input.extend_from_slice(hidden);
        let mut x = vec![0f32; h];
        matmul(self.pre_proj, h, 2 * b, &input, 1, &mut x);

        let (cos_s, sin_s) = rope_tables(&[pos], &self.inv_freq_sliding);
        let (cos_f, sin_f) = rope_tables(&[pos], &self.inv_freq_full);

        let mut hn = vec![0f32; h];
        for layer in &self.layers {
            let hd = layer.hd;
            let (cos, sin, cache, window) = if layer.sliding {
                (&cos_s, &sin_s, cache_s, Some(cfg.sliding_window))
            } else {
                (&cos_f, &sin_f, cache_g, None)
            };
            rmsnorm(&x, Some(layer.ln_in), eps, &mut hn);
            let mut q = vec![0f32; nh * hd];
            matmul(layer.q, nh * hd, h, &hn, 1, &mut q);
            for qh in q.chunks_mut(hd) {
                rmsnorm_inplace(qh, Some(layer.q_norm), eps);
                apply_rope(qh, cos, sin);
            }
            let mut att = vec![0f32; nh * hd];
            // The query sits at `pos` but only positions `< pos` are cached; `attention` treats the
            // query's own slot as cached, so anchor it at `pos - 1` to cover exactly [pos-window, pos).
            debug_assert_eq!(cache.len, pos, "drafter expects the cache to hold exactly the tokens before `pos`");
            attention(&q, cache, nh, 1, pos - 1, window, &mut att);
            let mut o = vec![0f32; h];
            matmul(layer.o, h, nh * hd, &att, 1, &mut o);
            rmsnorm_inplace(&mut o, Some(layer.ln_post_attn), eps);
            add_inplace(&mut x, &o);

            rmsnorm(&x, Some(layer.ln_pre_ffn), eps, &mut hn);
            let inter = cfg.intermediate_size;
            let mut g = vec![0f32; inter];
            let mut u = vec![0f32; inter];
            matmul_multi(&mut [(layer.gate, inter, &mut g[..]), (layer.up, inter, &mut u[..])], h, &hn, 1);
            gelu_mul(&mut g, &u);
            matmul(layer.down, h, inter, &g, 1, &mut o);
            rmsnorm_inplace(&mut o, Some(layer.ln_post_ffn), eps);
            add_inplace(&mut x, &o);
            if layer.layer_scalar != 1.0 {
                scale_inplace(&mut x, layer.layer_scalar);
            }
        }
        rmsnorm_inplace(&mut x, Some(self.final_norm), eps);

        let mut next_hidden = vec![0f32; b];
        matmul(self.post_proj, b, h, &x, 1, &mut next_hidden);

        // Centroid-masked LM head: top-k centroids, then exact logits over their token clusters.
        let mut cl = vec![0f32; self.num_centroids];
        matmul(self.centroids, self.num_centroids, h, &x, 1, &mut cl);
        let mut idx: Vec<usize> = (0..self.num_centroids).collect();
        let k = self.top_k_centroids.min(self.num_centroids);
        idx.select_nth_unstable_by(k - 1, |&a, &b| cl[b].partial_cmp(&cl[a]).unwrap_or(std::cmp::Ordering::Equal));
        let per = cfg.vocab_size / self.num_centroids;
        let sel = &idx[..k];
        let mut scores = vec![0f32; k * per];
        let embed = self.embed;
        let ordering = &self.token_ordering;
        let xr = &x;
        for_chunks(&mut scores, per, true, |i, chunk| {
            let c = sel[i];
            for (j, &tok) in ordering[c * per..(c + 1) * per].iter().enumerate() {
                let row = &embed[tok as usize * h..(tok as usize + 1) * h];
                chunk[j] = dot_bf16(row, xr);
            }
        });
        let mut best = (f32::NEG_INFINITY, 0u32);
        for (i, &c) in sel.iter().enumerate() {
            for j in 0..per {
                let sc = scores[i * per + j];
                if sc > best.0 {
                    best = (sc, ordering[c * per + j]);
                }
            }
        }
        (best.1, next_hidden)
    }
}