Getting Started

InstallationCopy Anchor Link

# CPU only (default)
uv pip install portc

# With NVIDIA GPU support (CUDA 13.x, driver >= 580)
uv pip install "portc[gpu-nvidia]"

# With AMD GPU support (ROCm)
uv pip install "portc[gpu-rocm]"

The runtime dependencies are NumPy, opt_einsum and array-api-compat. The GPU extras add CuPy; see GPU execution.

The three entry pointsCopy Anchor Link

from portc import apply, compress, src

• apply contracts an MPO with an MPS or another MPO and compresses the product.
• compress compresses a single MPS or MPO.
• src contracts a whole stack of trains in one sweep; apply and compress are its two- and one-train special cases.

All three are pure: they never modify their inputs, and return a new list of site arrays in right-canonical form. Whether a train is an MPS or an MPO is inferred from the rank of its first site tensor, so there is no wrapper type.

A first exampleCopy Anchor Link

Apply a random MPO to a random MPS and keep the result at bond dimension 16. Trains are plain lists of arrays in the quimb layout, so quimb can build the inputs and check the result:

import quimb.tensor as qtn

from portc import apply

H = qtn.MPO_rand(10, bond_dim=4, phys_dim=2, seed=0)
psi = qtn.MPS_rand_state(10, bond_dim=8, phys_dim=2, seed=1)

result = apply(H.arrays, psi.arrays, chi_out=32, seed=2)
phi = qtn.MatrixProductState(result)

exact = H.apply(psi, compress=False)
print(f"max bond: {phi.max_bond()}")
print(f"relative error: {phi.distance(exact) / exact.norm():.2e}")

The exact product has bond dimension 32, so chi_out=32 reproduces it essentially exactly; lower it to trade accuracy for size. Passing seed makes the random sketches, and therefore the result, reproducible.

Networks with fewer than three sites fall back to an exact SVD-based contraction and log a warning.

Next stepsCopy Anchor Link

• Concepts explains what SRC does and the index conventions.
• Features covers stacks, adaptive truncation and GPUs.
• Python API lists every argument.