bitorch is a compact boolean tensor library with a PyTorch-shaped API.

Implemented surface:

• BitTensor: torch.Tensor subclass backed by torch.bool
• Custom STE-style backward traversal with float-valued .grad
• Boolean tensor operators: AND, OR, XOR, NOT
• BitParameter, BitModule, and BitLinear
• BitSGD optimizer over latent float parameter states
• BitStepLR scheduler

BitLinear uses boolean algebra rather than arithmetic:

• pairwise multiply -> AND
• reduction sum -> OR
• bias combine -> XOR

import torch

from bitorch import BitTensor
from bitorch.nn import BitLinear
from bitorch.optim import BitSGD

layer = BitLinear(3, 2)
inputs = BitTensor([[True, False, True]], requires_grad=True)

outputs = layer(inputs)
outputs.backward(torch.ones_like(outputs.to_tensor(), dtype=torch.float32))

optimizer = BitSGD(layer.parameters(), lr=0.1)
optimizer.step()
optimizer.zero_grad()