Compile Clifford computations

Plan before calling torch.compile. Planning resolves the algebraic operation and storage contracts; compilation then sees its tensor execution. Keep Python configuration and plan construction outside the compiled function.

import torch
from clifra import make_algebra

algebra = make_algebra(3, dtype=torch.float64)
vector = algebra.layout((1,))
even = algebra.layout((0, 2))
product = algebra.plan_product(left=vector, right=vector, output=even)
compiled = torch.compile(product, backend="aot_eager", fullgraph=True)

x = torch.tensor([[1., 2., 0.]], dtype=torch.float64, requires_grad=True)
y = torch.tensor([[0., 1., 3.]], dtype=torch.float64, requires_grad=True)

expected, actual = product(x, y), compiled(x, y)

torch.testing.assert_close(actual, expected)

expected_grad = torch.autograd.grad(expected.square().sum(), (x, y))
actual_grad = torch.autograd.grad(actual.square().sum(), (x, y))
for got, want in zip(actual_grad, expected_grad):
    torch.testing.assert_close(got, want)

aot_eager checks graph capture and differentiation without generating an optimized device kernel. It is also the backend used by the repository's representative compilation tests; those tests verify graph capture and differentiation, not speedup or support for every optimizing backend. Select and test an optimizing backend for your device, dtype, operation, and workload separately.

Batch shapes remain ordinary tensor shapes. PyTorch may specialize or recompile for new shapes; a reusable clifra plan accommodates this flexibility but does not promise a single compiled graph across all batch sizes. Move the module to its intended device and dtype before compiling. Data-dependent strict geometry checks may raise from compiled execution when their mathematical preconditions fail.