A fast, network-connected, differentiable tensor library for TypeScript (and JavaScript). Built with bun + flashlight for software engineers and researchers alike.

⚠️ This is experimental software! ⚠️

• Usage
• Install
• Build from source
• Benchmarks
• Contributing

Quickstart

Install Bun and ArrayFire

then run:

bun install @shumai/shumai

Only macOS and Linux are supported. Linux installs default to GPU computation with CUDA, and macOS to CPU. Detailed install instructions below.

Install is work in progress: please file an issue if you run into problems.

Usage

shumai will always attempt to use an attached GPU or accelerator; although CPU computation will use the ArrayFire CPU backend, which is not well-optimized.

We hope to support the ArrayFire OpenCL backend and other non-ArrayFire tensor backends soon.

If shumai seems unusually slow, please file an issue!

Standard array utilities:

import * as sm from "@shumai/shumai"

// create a 1024 by 1024 tensor, randomly filled with normal distribution
let X = sm.randn([1024, 1024])
let W = sm.identity(1024)
let Y = X.matmul(W)
console.log(Y.shape)

Conversion to and from JavaScript native arrays:

const data : Float32Array = new Float32Array(128)
for (let i = 0; i < 128; ++i) {
  data[i] = Math.random()
}

const X : Tensor = sm.tensor(data)
const pi = sm.scalar(3.14)
const Y = X.mul(pi)

// tensors can be converted back to native JavaScript
const Y_data = Y.toFloat32Array()

// scalar tensors can be converted to JavaScript numbers
const total : number = X.sum().toFloat32()

Gradients:

const W = sm.randn([128, 128])
W.requires_grad = true

const X = sm.randn([128, 128])
const diff = X.sub(W)
const mse = diff.mul(diff).sum()
mse.backward()

W.grad // this gradient is now populated

// copy W without allowing gradient updates
const Y = W.detach()
Y.sum().backward() // nothing changes

Some more examples can be found here.

Supported operators can be found here.

Install

The install procedure is a work in progress! If you have any problems building or installing, we would greatly appreciate filed issues. Please tell us about your platform/OS when you do.

Prerequisites:

• Ensure you have bun installed (https://bun.sh).
• Install ArrayFire. macOS users should install ArrayFire's CPU backend; Linux users should install the CUDA backend^.
  • macOS --- ArrayFire can easily be installed with Homebrew:

    brew install arrayfire

• Linux --- instructions can be found here. On Ubuntu, ArrayFire can be installed via package managers (e.g. apt).

Once bun and ArrayFire are installed, install the package and backing libs with bun:

bun install @shumai/shumai

Windows Support

While not officially supported, Windows users have been successful leveraging Docker + WSL2 + Linux. Including CUDA support.

Building Native Libraries from Source

Note: not required when developing TypeScript/Javascript library components locally.

From source build instructions for:

• macOS
• Linux

This process will build the dependent ffi libraries (libflashlight and libflashlight_binding) and pack them using npm pack to generate a @shumai/shumai_*.tgz package. You can then use npm install $PATH_TO_SOURCE/@shumai/shumai-*.tgz to install the package where you'd like.

Building on macOS from Source

First, install ArrayFire CPU with brew install arrayfire.

Build and install Flashlight:

mkdir -p $HOME/usr/ # installing flashlight here
git clone --recursive --depth 1 https://github.com/flashlight/flashlight.git
cd flashlight
mkdir -p build
cd build
cmake .. \
  -DCMAKE_BUILD_TYPE=Release \
  -DBUILD_SHARED_LIBS=ON  \
  -DCMAKE_INSTALL_PREFIX=$HOME/usr \
  -DFL_USE_ARRAYFIRE=ON \
  -DFL_ARRAYFIRE_USE_CPU=ON \
  -DFL_USE_ONEDNN=OFF \
  -DFL_BUILD_DISTRIBUTED=OFF \
  -DFL_BUILD_TESTS=OFF \
  -DFL_BUILD_EXAMPLES=OFF
make -j$(nproc)
make install

Build Flashlight bindings for Shumai:

cd shumai
mkdir -p build
cd build
cmake .. -Dflashlight_DIR=$HOME/usr/share/flashlight/cmake/
make -j$(nproc)

Profiling

On macOS, you can record perf with xcrun xctrace record --template "Time Profiler" --launch $(which bun) train.js.

Building on Linux from Source

First install ArrayFire. The Linux build for shumai uses the CUDA backend, but from source, you can build the CPU backend as well (OpenCL support coming soon).

Build and install Flashlight:

mkdir -p $HOME/usr/ # installing flashlight here
git clone --recursive --depth 1 https://github.com/flashlight/flashlight.git
cd flashlight
mkdir -p build
cd build
cmake .. \
  -DCMAKE_BUILD_TYPE=RelWithDebInfo \ # or as specified
  -DFL_ARRAYFIRE_USE_CPU=OFF \
  \ # swap with the above to build for CPU
  -DFL_ARRAYFIRE_USE_CUDA=ON \ 
  -DFL_BUILD_DISTRIBUTED=OFF \
  -DFL_USE_ONEDNN=OFF \
  -DFL_BUILD_TESTS=OFF \
  -DFL_BUILD_EXAMPLES=OFF \
  -DFL_BUILD_SCRIPTS=OFF \
  -DCMAKE_INSTALL_PREFIX=$HOME/usr/
make -j$(nproc)
make install

Build bindings for shumai:

mkdir -p build && cd build
cmake .. \
    -DBUILD_SHARED_LIBS=ON \
    -DCMAKE_BUILD_TYPE=RelWithDebInfo \ # or as specified
    -Dflashlight_DIR=${FLASHLIGHT_INSTALL_PREFIX}/share/flashlight/cmake \
    -DArrayFire_DIR=${ARRAYFIRE_INSTALL_PREFIX}/share/ArrayFire/cmake # if built from source, else not needed
make -j$(nproc)

Why build this?

With Shumai, we hope to make

• Creating datasets easier
  • JavaScript, with native typed arrays and a JIT compiler, is perfect for twiddling with data before it can be made into big, flat GPU-compatible arrays.
• Training small models faster
  • FFI bindings in Bun are crazy fast (~3ns), so JS gets out of the way when training small models
• Advanced/fine-grained training/inference logic more expressive
  • Bun uses the JSC JIT compiler, meaning you can confidently write complex training logic without needing a native C++ implementation
• Building applications enoyable
  • JavaScript has a large HUGE ecosystem, which facilitates better application development

Benchmarks

Benchmark data is collected from https://github.com/shumai-org/benchmarks

On an Apple M1 Pro:

On an Nvidia GP100:

Memory Usage

While the out of the box memory management may suffice in many cases, tuning memory usage can greatly improve performance by reducing unnecessary overhead from the Garbage Collector.

import { util } from '@shumai/shumai'

util.memoryOptions({
  lowerBoundThreshold: 100e6, // 100MB
  upperBoundThreshold: 5e9, // 5GB
  delayBetweenGCs: 1000 // 1s
})

Pay special attention to upperBoundThreshold which if exceeded will force GC for every allocated tensor, ignoring delayBetweenGCs. Supplying a value that will fully utilize your hardware can greatly improve performance.

Statistics

graph TD
  OpA(Op A) --> statsA{{"stats A"}};
  OpB(Op B) --> statsA;
  statsA --> LoggerA{{"LoggerConsole A"}};
  LoggerA --> Stdout(("Stdout"));
  OpC(Op C) --> statsA;
  OpD(Op D) --> statsA;
  statsA --> LoggerB("LoggerCustom B");
  LoggerB --> Disk(("Disk"));

Basic usage of gathering statistics is as simple adding a collector using the default StatsLoggerConsole.

import { stats, StatsLoggerConsole, rand, matmul } from '@shumai/shumai'

stats.enabled = true // all ops following will capture stats

// perform ops...

stats.enabled = false // all ops following will no longer capture stats

While the above examples may suffice for simple use cases, if you're looking to capture stats across multiple threads, processes, and/or hosts, StatsLoggerHttp has you covered.

graph TD
  subgraph Host C
    Processor("LoggerHttp Processor")
    style Processor stroke:#222,stroke-width:4px,stroke-dasharray:5 5
  end
  subgraph Host A
    OpA(Op A) --> statsA{{"stats A"}};
    OpB(Op B) --> statsA;
    statsA --> LoggerA{{"LoggerHttp A"}};
    LoggerA --> Processor;
  end
  subgraph Host B
    OpC(Op C) --> statsB{{"stats B"}};
    OpD(Op D) --> statsB;
    statsB --> LoggerB{{"LoggerHttp B"}};
    LoggerB --> Processor;
  end

import { StatsLoggerHttp } from '@shumai/shumai'

stats.logger = new StatsLoggerHttp({ url: 'http://localhost:4242' })

For more custom needs you can supply your own logger:

import { StatsLogger, StatsLoggerData } from '@shumai/shumai'

class CustomLogger implements StatsLogger {
  async process(data: StatsLoggerData): Promise<void> {
    const summary = data.collector.getSummary()
    console.log('Collector stats:', summary)
  }
}

stats.logger = new CustomLogger()

By default stack tracing is disabled as it adds 50%+ overhead, but can be enabled via stats.collectStacks = true.

Scoped Statistics

If you wish to isolate stats profiling you can do this as well:

import { collectStats } from '@shumai/shumai'

const scopedStats = collectStats(() => {
  // perform ops...
}/*, StatsCollectorOptions | StatsLogger */)
console.log(scopedStats.getSummary())

Contributing

If you'd like to make changes to the core bindings or ffi, first build from source.

All files ending in *.inl or *_gen.ts are generated. These can be modified by editing scripts/gen_binding.py and running ./scripts/gen_all_binding.sh.

See the CONTRIBUTING file for style guidance and more info on how to help out. 😁

License

shumai is MIT licensed, as found in the LICENSE file.