Denograd is a rewrite of tinygrad in TypeScript. JS ecosystem is very large, but it didn't have a good ML library for model inference and training. Since tinygrad doesn't use any external python libraries, has potential to be the fastest way to run models, is quite simple compared to others and supports many runtimes, I decided to rewrite it in TS to get the same experience in browser and in deno/node/bun.
Why you should use Denograd?
- 0 dependencies
- will be fast (not yet)
- Multiple runtime backends (WebGPU, CLANG, + others coming soon)
- Clean, modern API inspired by tinygrad's elegant design
- Works in browser and in Deno (Node and Bun support coming soon)
See MNIST inference and training example with WebGPU on denograd.com
There are multiple ways to use denograd:
<html>
<head>
<!-- Makes `dg` available globally -->
<script src="https://denograd.com/denograd.js"></script>
<script>
const run = async () => {
const mnist = await new dg.MNIST().load()
console.log(await mnist.call(dg.Tensor.ones([1, 1, 28, 28])).tolist())
}
run()
</script>
</head>
<body></body>
</html>import { Tensor, MNIST } from "https://denograd.com/denograd.mjs"
const mnist = await new MNIST().load()
console.log(await mnist.call(Tensor.ones([1, 1, 28, 28])).tolist()) Install package from jsr.io
# with deno
deno add jsr:@denograd/denograd
# with npm
npx jsr add @denograd/denograd
# with yarn
yarn dlx jsr add @denograd/denograd
# with pnpm
pnpm dlx jsr add @denograd/denograd
# with bun
bunx jsr add @denograd/denogradand then import with
import { Tensor, MNIST } from "@denograd/denograd"
const mnist = await new MNIST().load()
console.log(await mnist.call(Tensor.ones([1, 1, 28, 28])).tolist()) Soon everything should work like this in browser and server with no install step, while still being fast:
const llama = await new Llama({ model: '3.1-3B' }).load()
const res = await llama.run({ prompt: 'Hello how are you?' })const llama = await new Llama({ model: '3.1-3B', device: 'CLOUD', host: process.env.CLOUD_HOST }).load()
const res = await llama.run({ prompt: 'Hello how are you?' })const whisper = await new Whisper({ model: "large-v2" }).load()
const listening = whisper.startListening()
// after some time
const text = await listening.stop()const tts = await new TTS()
const audio = await tts.run({ text: "Hello how are you?" })
audio.play()class MNIST extends Model {
layers: Layer[] = [
new nn.Conv2d(1, 32, 5),
Tensor.relu,
new nn.Conv2d(32, 32, 5),
Tensor.relu,
new nn.BatchNorm(32),
Tensor.max_pool2d,
new nn.Conv2d(32, 64, 3),
Tensor.relu,
new nn.Conv2d(64, 64, 3),
Tensor.relu,
new nn.BatchNorm(64),
Tensor.max_pool2d,
(x) => x.flatten(1),
new nn.Linear(576, 10),
]
}
const [X_train, Y_train, X_test, Y_test] = await mnist()
const model = new MNIST()
const opt = Adam(get_parameters(model))
const train_step = async (): Promise<Tensor> => {
Tensor.training = true
opt.zero_grad()
const samples = Tensor.randint([BS], undefined, X_train.shape[0])
const loss = model.call(X_train.get(samples)).sparse_categorical_crossentropy(Y_train.get(samples)).backward()
await opt.step()
Tensor.training = false
return loss
}
const get_test_acc = (): Tensor => model.call(X_test).argmax(1).eq(Y_test).mean().mul(100)
let test_acc = NaN
const t = new Tqdm(range(get_number_env('STEPS', 70)))
for await (const i of t) {
const loss = await (await train_step()).item()
if (i % 10 === 9) test_acc = await get_test_acc().item()
t.set_description(`loss: ${loss.toFixed(2)}, test_accuracy: ${test_acc.toFixed(2)}`)
}
await model.save('./mnist.safetensors')- rewrite all the necesary parts of tinygrad for MNIST, with 'JS' runtime, with tests comparing the python and TS implemenations
- Github CI
- CLANG runtime
- get MNIST training
- get working inside browser with JS runtime
- WebGPU runtime (MNIST inference running in WebGPU: https://karelnagel.github.io/denograd/, training coming soon)
- delete lazy + other tinygrad updates
- add all the missing parts of Tensor and other code that were left out in the beginning.
- get hand_coded_optimisations working correctly
- WEBGPU training MNIST in browser
- docs
- some LLM
- whisper
- bun support
- node support
- WASM runtime
- have popular models as a package, maybe even as prebuilt binaries with
deno compile - CLOUD runtime
- METAL runtime
- AMD runtime
- Nvidia runtime
- JIT
- Multi