#moonxi-net-gpu

#Prerequisites

Note: The standard MoonBit toolchain does not support CUDA compilation. You must use the CUDA-patched fork.

#Installation

moon add chnlkw/moonxi-net-gpu

#Quick Example

///|
fn main {
  // Initialize CUDA context
  @cuda.init()

  // Use GpuTensor instead of NpArray
  let params : Array[@gt.GpuTensor] = train_linear()
  // ...same generic train_linear() function, no code changes needed

  @cuda.shutdown()
}

moon run moonxi-net-gpu/examples/linear --target native --release
# Learned: w=3.00, b=1.00

#Usage

#Import in moon.pkg

{
  "import": [
    "chnlkw/moonxi-net" @tensor,
    "chnlkw/moonxi-net-gpu/cuda" @cuda,
    "chnlkw/moonxi-net-gpu/tensor" @gt,
    "chnlkw/moonxi-net-gpu/train" @gpu_train,
    "chnlkw/moonxi-net-gpu/event_tensor" @et
  ]
}

#Core Packages

#GPU Training Workflow

///|
fn main {
  // Initialize CUDA
  @cuda.init()

  // Create tensors using GpuTensor (GPU backend)
  let w = @grad.no_grad(@gt.GpuTensor::zeros([10, 10]))
  let b = @grad.no_grad(@gt.GpuTensor::zeros([10]))

  // Same autograd API as CPU
  @grad.clear_tape()
  let loss = /* ... your forward pass ... */
  loss.backward()

  // Access gradients
  match w.grad {
    Some(Some(g)) => w.value = w.value.sub(g.scale(0.01))
    _ => ()
  }

  @cuda.shutdown()
}

#MNIST Training

bash scripts/download_mnist.sh
moon run moonxi-net-gpu/examples/mnist --target native --release                    # CPU backend
moon run moonxi-net-gpu/examples/mnist --target native --release -- --gpu           # GPU backend
moon run moonxi-net-gpu/examples/mnist --target native --release -- -e 20 -b 128    # Custom config

#CIFAR-10 ResNet-18 Training

bash scripts/download_cifar10.sh
moon run moonxi-net-gpu/examples/cifar10 --target native --release                   # 20 epochs, momentum SGD
moon run moonxi-net-gpu/examples/cifar10 --target native --release -- -e 10          # Custom epochs
moon run moonxi-net-gpu/examples/cifar10 --target native --release -- -o adam        # Use Adam optimizer

#Build & Test

# Requires CUDA-patched MoonBit + CUDA toolkit
moon test --target native
moon build --target native

#License