stabilizer-qec

A surface-code simulator and decoder written in Rust, with Python bindings.

• Circuits and detector error models in Stim's text formats, and error models built by walking a circuit backwards, identical to Stim's on every circuit checked.
• Exact minimum-weight matching by sparse blossom, and correlated matching, agreeing with PyMatching shot for shot except for ties.
• A bit-parallel sampler, 64 shots per machine word, that runs REPEAT blocks without flattening.
• Raw measurements to detection events, reproducing Google's Willow and Sycamore files bit for bit.
• Sliding and parallel window decoders for real-time decoding, and decoding as a service (decode_async, a decoder queue with statistics, a decoder server).
• The APIs of Stim, PyMatching, sinter, ldpc, Tsim, tqec, Deltakit, PECOS and seventeen more QEC tools under their own names, each port checked against its original.
• Clifford+T circuits with a coherent error on every gate: a stabilizer-rank sampler for the gates and a joint sampler that reweights the twirl of the errors around them.
• Fault-tolerant compilation (tqec's block graphs, Loom's lattice surgery), resource estimates (Azure's and Qualtran's), and sampling and min-sum BP on the GPU through wgpu.
• Belief propagation reproducing the ldpc library's arithmetic bit for bit, and belief-matching equal to the beliefmatching package's.
• BP+OSD (OSD-0, OSD-E, OSD-CS), whose corrections equal ldpc's BpOsdDecoder's, and IBM's bivariate bicycle codes: the gross code [[144, 12, 12]] and [[72, 12, 6]], with the depth-8 syndrome cycle of Bravyi et al. (Nature 627, 778, 2024).
• On the gross code, every automorphism's action on the logical qubits, and the gauging measurement of a logical operator as a circuit, its error model equal to Stim's.
• Stim's whole circuit language (every gate, all three bases, MPP, SPP, correlated and heralded noise, measurement feedback, instruction tags, Pauli-target observables), checked against Stim on random circuits that use every gate.
• Lattice surgery as single circuits, their error models equal to Stim's: Z⊗Z and X⊗X merges, a logical CNOT, merges in a row, and lines of patches merged at once.
• sinter, whole (stabilizer_qec.sinter), for threshold sweeps with nothing else installed, and objects that pickle for multiprocessing.

import stabilizer_qec as sq

circuit = sq.memory_circuit(distance=5, rounds=5, p=0.004)               # a d = 5 SD6 memory
dets, obs = circuit.compile_detector_sampler(seed=1).sample(100_000, separate_observables=True, threads=0)
dem = circuit.detector_error_model(decompose_errors=True)
pred = sq.Matching(dem, enable_correlations=True).decode_batch(dets, enable_correlations=True, threads=0)
print((pred != obs).any(axis=1).mean())                                  # logical error rate

The gross code, decoded by BP+OSD:

gross = sq.BivariateBicycleCode("gross")                                 # [[144, 12, 12]], 12 cycles
c = gross.memory_circuit(12, 0.003)
dets, obs = c.compile_detector_sampler(seed=2).sample(2_000, separate_observables=True)
pred = sq.BpOsd(c.detector_error_model()).decode_batch(dets, threads=0)
print((pred != obs).any(axis=1).mean())                                  # any of the 12 logicals wrong

Threshold sweeps with stabilizer_qec.sinter, which is sinter whole (matplotlib to plot):

from stabilizer_qec import sinter

stats = sinter.collect(tasks=tasks, decoders=["pymatching", "belief_matching", "bposd"],
                       num_workers=8, max_shots=100_000)

The names follow Stim (Circuit, compile_detector_sampler, separate_observables), PyMatching (Matching, decode_batch, enable_correlations) and ldpc (BpDecoder, BpOsdDecoder, error_channel). A seed gives the same shots on any machine and thread count. See the changelog.

The explainer, its figures running the engine in the browser, is at https://qcompiler.jaspersands.com, with a technical report. The source and the full write-up are at https://github.com/Jaspersands/quantum-simulator-qec.