Nabla is a Machine Learning library for the emerging Mojo/Python ecosystem, featuring:
- Gradient computation the PyTorch way (imperatively via .backward())
- Purely-functional, JAX-like composable function transformations:
grad,vmap,jit, etc. - Custom differentiable CPU/GPU kernels
For tutorials and API reference, visit: nablaml.com
pip install nabla-mlThe most simple, but fully functional Neural Network training setup:
import nabla as nb
# Defines MLP forward pass and loss.
def loss_fn(params, x, y):
for i in range(0, len(params) - 2, 2):
x = nb.relu(x @ params[i] + params[i + 1])
predictions = x @ params[-2] + params[-1]
return nb.mean((predictions - y) ** 2)
# JIT-compiled training step via SGD
@nb.jit(auto_device=True)
def train_step(params, x, y, lr):
loss, grads = nb.value_and_grad(loss_fn)(params, x, y)
return loss, [p - g * lr for p, g in zip(params, grads)]
# Setup network (hyper)parameters.
LAYERS = [1, 32, 64, 32, 1]
params = [p for i in range(len(LAYERS) - 1) for p in (nb.glorot_uniform((LAYERS[i], LAYERS[i + 1])), nb.zeros((1, LAYERS[i + 1])),)]
# Run training loop.
x, y = nb.rand((256, 1)), nb.rand((256, 1))
for i in range(1001):
loss, params = train_step(params, x, y, 0.01)
if i % 100 == 0: print(i, loss.to_numpy())- Clone the repository
- Create a virtual environment (recommended)
- Install dependencies
git clone https://github.com/nabla-ml/nabla.git
cd nabla
python3 -m venv venv
source venv/bin/activate
pip install -r requirements-dev.txt
pip install -e ".[dev]"nabla/
├── nabla/ # Core Python library
│ ├── core/ # Tensor class and MAX compiler integration
│ ├── nn/ # Neural network modules and models
│ ├── ops/ # Mathematical operations (binary, unary, linalg, etc.)
│ ├── transforms/ # Function transformations (vmap, grad, jit, etc.)
│ └── utils/ # Utilities (formatting, types, MAX-interop, etc.)
├── tests/ # Comprehensive test suite
├── tutorials/ # Notebooks on Nabla usage for ML tasks
└── examples/ # Example scripts for common use cases
Contributions welcome! Discuss significant changes in Issues first. Submit PRs for bugs, docs, and smaller features.
Nabla is licensed under the Apache-2.0 license.
