Crate oxillama_wasm 

WebAssembly bindings for OxiLLaMa.

Exposes GGUF header parsing, Q4_0 dequantization, and full text generation to JavaScript/TypeScript via wasm-bindgen.

Feature flags

Feature	Default	Description
inference	yes	Enables generate() via oxillama-runtime with
		the pure-Rust unstable_wasm tokenizer backend.

Usage (generate)

import init, { generate } from './oxillama_wasm.js';
await init();

const modelResp = await fetch('model.gguf');
const modelBytes = new Uint8Array(await modelResp.arrayBuffer());
const tokenizerResp = await fetch('tokenizer.json');
const tokenizerJson = await tokenizerResp.text();

// Streaming: pass a callback to receive each token as it is generated.
const text = generate(modelBytes, tokenizerJson, "Hello, world!", 128,
    (token) => process.stdout.write(token));
console.log(text);

Re-exports

pub use service_worker::get_service_worker_script;

pub use service_worker::register_service_worker;

pub use service_worker::ServiceWorkerOptions;

pub use simd_check::get_simd128_status;

pub use streaming_loader::StreamingGgufLoader;

pub use streaming_loader::StreamingLoadOptions;

Modules

gpu_bridge

Async WebGPU bridge for OxiLLaMa WASM.

idb_cache

IndexedDB model cache for persisting GGUF model bytes across page reloads.

service_worker

Service-worker registration helper for the OxiLLaMa WASM browser build.

simd_check

SIMD128 capability detection for the OxiLLaMa WASM build.

streaming_load

Streaming / chunked GGUF loader.

streaming_loader

Streaming GGUF loader with LRU tensor cache and on-demand byte-range fetching.

webgpu

WebGPU compute bridge for accelerating dequantization in the browser.

worker

Web-worker message-passing helpers for offloaded inference.

Structs

GgufMetadataJs

Typed GGUF metadata returned by parse_gguf_metadata.

WasmEngine

Opaque handle wrapping a loaded InferenceEngine for use from JS.

Functions

dequant_q4_0

Dequantize a buffer of Q4_0 blocks to an array of f32 values.

dequant_q4_k

Dequantize a buffer of Q4_K blocks to an array of f32 values.

dequant_q5_k

Dequantize a buffer of Q5_K blocks to an array of f32 values.

dequant_q6_k

Dequantize a buffer of Q6_K blocks to an array of f32 values.

generate

Run full text generation from an in-memory GGUF model.

init

Initialize the WASM module (sets up panic hook for better error messages).

list_tensor_names

Return all tensor names stored in a GGUF file as a JS array of strings.

load_model_from_bytes_with_progress

Load a GGUF model from raw bytes, reporting progress via an optional JS callback.

parse_gguf_header

Parse a GGUF file header from raw bytes and return key metadata as a JS object.

parse_gguf_metadata

Parse a GGUF file and return typed metadata as a JS object.