rullama 0.4.0

Browser-resident Gemma 4 inference: pure Rust → WebAssembly + WebGPU. Loads Ollama's on-disk GGUF blobs and runs the forward pass on the local GPU via hand-written WGSL.
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

//! Quick token-id → string lookup using the vocab embedded in the GGUF. No tokenizer
//! mechanics — just an array indexing.

use std::env;
use std::fs;
use std::process::ExitCode;

use rullama::gguf::GgufReader;

fn main() -> ExitCode {
    let mut args = env::args().skip(1);
    let path = match args.next() {
        Some(p) => p,
        None => {
            eprintln!("usage: decode_ids <gguf> <id1> <id2> ...");
            return ExitCode::from(2);
        }
    };
    let bytes = fs::read(&path).expect("read");
    let r = GgufReader::new(bytes).expect("parse");
    let tokens = r
        .get("tokenizer.ggml.tokens")
        .expect("vocab")
        .as_string_array()
        .expect("strs");
    for arg in args {
        let id: usize = arg.parse().expect("id");
        if id < tokens.len() {
            println!("{:>7}  {:?}", id, tokens[id]);
        } else {
            println!("{id} OUT OF RANGE");
        }
    }
    ExitCode::SUCCESS
}