Crate wasmicro 

wasmicro

Tiny transformer inference for the web. One file. No build step.

Design rules

1. Tiny WASM bundle (target: < 250 KB after wasm-opt -Oz).
2. Fast cold start (target: < 500 ms model load + first inference).
3. Forward inference only — no autograd, no optimizers, no training.
4. Owned tensors, no Rc<RefCell> indirection.
5. Minimal dependencies. The full default build pulls in only bytemuck.
6. Same code runs natively and in WASM. The library never opens files — callers pass bytes in via ModelFile::parse.

Quick start

use std::fs;
use wasmicro::{
    models::bert::{BertConfig, BertModel},
    ModelFile, WordPieceTokenizer,
};

let model_bytes = fs::read("model.safetensors").unwrap();
let vocab_bytes = fs::read("vocab.txt").unwrap();
let file = ModelFile::parse(&model_bytes).unwrap();
let tokenizer = WordPieceTokenizer::from_vocab_bytes(&vocab_bytes).unwrap();

let config = BertConfig::mini_lm_l6_v2();
let model = BertModel::from_safetensors(&file, config, "").unwrap();

let embedding = model.embed_text(&tokenizer, "hello world", 128).unwrap();
println!("embedding dim: {:?}", embedding.shape().as_slice());

Re-exports

pub use error::Error;

pub use error::Result;

pub use loader::Dtype;

pub use loader::ModelFile;

pub use loader::TensorView;

pub use pipeline::Pipeline;

pub use quant::QuantizedTensorI8;

pub use quant::QuantizedTensorQ4;

pub use quant::QuantizedTensorU8;

pub use tensor::Shape;

pub use tensor::Tensor;

pub use tokenizer::EncodedInput;

pub use tokenizer::WordPieceOptions;

pub use tokenizer::WordPieceTokenizer;

pub use tokenizer::bpe::BpeTokenizer;

Modules

error

Error types used across the crate.

loader

Safetensors model file loader.

models

Pre-built transformer architectures.

ops

Forward-only tensor operations.

pipeline

High-level pipeline API — one call to load and run any supported model.

quant

Small weight-only quantized tensor types.

tensor

Plain, forward-only tensor.

tokenizer

Minimal WordPiece tokenizer with Unicode support.