noniml 0.1.2

Noni (WIP)

A minimal tensor library with autograd flexible for building good enough deep learning models.

Familiar API

a = Tensor([[1., 2.], [3., 4.]], requires_grad=True)
b = Tensor([[0.5, -1.], [2., 0.]], requires_grad=True)

# Each op records its backward function
c = a * b        # op="*",  backward: dc/da = b, dc/db = a
d = c.sum()      # op="sum", backward: ones

d.backward()     # topological sort → apply each _backward in reverse

print(a.grad)    # dL/da = b.data = [[0.5, -1.], [2., 0.]]
print(b.grad)    # dL/db = a.data = [[1., 2.], [3., 4.]]

Common Modules for everything

from noni.nn import Linear, LayerNorm, MultiHeadAttention, CrossEntropyLoss

# A simple 2-layer MLP
W1 = Linear(784, 256)
W2 = Linear(256, 10)

x = Tensor(some_batch)
h = W1(x).relu()
logits = W2(h)

loss = CrossEntropyLoss()(logits, targets)
loss.backward()   # gradients in W1.weight.grad, W2.weight.grad etc.

Build your own

Noni has three built-in backends:

Move tensors and modules to any backend with .to():

from noni import Tensor
from noni.nn import Linear

# Apple Silicon — runs matmul through Metal Performance Shaders
lin = Linear(512, 256)
lin.to("mlx")
x = Tensor(data, device="mlx")
y = lin(x)

There is also work going on to support CUDA natively as well as Vulkan compute and Triton. You can always implement and register your own backend if you prefer.

from noni.backends import Backend, register_backend


class MyDevice(Backend):
	...

register_backend("mygpu", MyDevice())

Building wheels

python -m build
twine upload dist/*