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

//! `oxinfer --model deepseek-ai/Janus-Pro-1B --generate-image TEXT -o out.png`: text to image
//! through `oxinfer::janus`; with `--prompt` (and `--image FILE`), questions about images.

use crate::Args;
use anyhow::{Context, Result, bail};
use oxinfer::generate::Flow;
use oxinfer::janus::{ChatOptions, ImageOptions, Janus, JanusOptions};
use std::io::{Read, Write};
use std::path::{Path, PathBuf};
use std::time::Instant;

pub fn run(args: &Args) -> Result<()> {
    let generate = args.generate_image.is_some() || (args.output.is_some() && args.images.is_empty());
    let opts = JanusOptions {
        quant: args.quant,
        mmap: args.mmap,
        device: args.device,
        cache_dir: args.cache_dir.clone(),
        generation: generate,
        understanding: !generate,
    };
    let jm = Janus::load(&args.model.to_string_lossy(), &opts)?;
    let c = jm.config();
    eprintln!(
        "[Janus-Pro: {} layers, hidden {}, {} | loaded in {:.1}s]",
        c.num_hidden_layers,
        c.hidden_size,
        args.quant.as_str(),
        jm.load_secs
    );
    let prompt = match args.generate_image.clone().or_else(|| args.prompt.clone()) {
        Some(p) => p,
        None => {
            let mut s = String::new();
            std::io::stdin().read_to_string(&mut s)?;
            s.trim().to_string()
        }
    };
    if generate {
        image(args, &jm, &prompt)
    } else {
        answer(args, &jm, &prompt)
    }
}

/// `out.png` for one image, `out-0.png`, `out-1.png`, ... for several.
fn numbered(out: &Path, i: usize, n: usize) -> PathBuf {
    if n == 1 {
        return out.to_path_buf();
    }
    let stem = out.file_stem().map(|s| s.to_string_lossy().to_string()).unwrap_or_default();
    let ext = out
        .extension()
        .map(|e| e.to_string_lossy().to_string())
        .unwrap_or_else(|| "png".into());
    out.with_file_name(format!("{stem}-{i}.{ext}"))
}

fn image(args: &Args, jm: &Janus, prompt: &str) -> Result<()> {
    let o = ImageOptions {
        cfg_weight: args.guidance.unwrap_or(5.0),
        temperature: if args.temp_given { args.temperature } else { 1.0 },
        seed: args.seed,
        parallel_size: args.num_images.max(1),
        ..Default::default()
    };
    eprintln!(
        "[{} image(s), cfg {}, temperature {}, seed {}]",
        o.parallel_size, o.cfg_weight, o.temperature, o.seed
    );
    let t0 = Instant::now();
    oxinfer::prof::reset();
    let (imgs, st) = jm.generate_images(prompt, &o, &mut |p| {
        if p.step % 8 == 0 || p.step == p.steps {
            eprint!("\r[image token {}/{}]", p.step, p.steps);
        }
        Flow::Continue
    })?;
    eprintln!();
    let out = args.output.clone().unwrap_or_else(|| PathBuf::from("image.png"));
    let mut written = Vec::new();
    for (i, img) in imgs.iter().enumerate() {
        let p = numbered(&out, i, imgs.len());
        std::fs::write(&p, oxinfer::png::encode_rgb(img.width, img.height, &img.rgb)).with_context(|| format!("write {}", p.display()))?;
        written.push(p.display().to_string());
    }
    eprintln!(
        "[{} prompt tokens | prefill {:.2}s | {} steps x {} sequences: {:.2}s ({:.1} ms/step) | VQ decode {:.2}s | total {:.2}s | wrote {}]",
        st.prompt_tokens,
        st.prefill,
        st.steps,
        2 * imgs.len(),
        st.tokens,
        st.tokens * 1e3 / st.steps.max(1) as f64,
        st.decode,
        t0.elapsed().as_secs_f64(),
        written.join(", ")
    );
    if args.profile {
        // Prefill and every image-token step (the VQ decoder's convolutions are not counted).
        oxinfer::prof::report(st.prefill + st.tokens, st.steps);
    }
    Ok(())
}

fn answer(args: &Args, jm: &Janus, question: &str) -> Result<()> {
    if !args.audios.is_empty() || !args.videos.is_empty() {
        bail!("Janus takes images only");
    }
    #[cfg(feature = "media")]
    let images = {
        let m = oxinfer::media::Media::new();
        args.images.iter().map(|p| m.load_image(p)).collect::<Result<Vec<_>>>()?
    };
    #[cfg(not(feature = "media"))]
    let images: Vec<oxinfer::video::RgbImage> = if args.images.is_empty() {
        Vec::new()
    } else {
        bail!("--image needs the `media` feature (image decoding)")
    };
    let o = ChatOptions {
        max_tokens: args.max_tokens,
        temperature: args.temperature,
        top_k: args.top_k,
        top_p: args.top_p,
        seed: args.seed,
        system_prompt: args.system.clone(),
    };
    let t0 = Instant::now();
    let mut out = std::io::stdout();
    let text = jm.understand(question, &images, &o, &mut |d| {
        let _ = out.write_all(d.as_bytes());
        let _ = out.flush();
        Flow::Continue
    })?;
    println!();
    eprintln!("[{} chars in {:.2}s]", text.len(), t0.elapsed().as_secs_f64());
    Ok(())
}