spd

State-of-the-art GPU code for sparse-Pauli dynamics (SPD), with a NumPy CPU backend and a legacy JAX backend.

SPD takes a circuit, evolves a sparse Pauli operator (SPO), computes expectation values, and can backpropagate a sparse Pauli gradient operator (SPGO) through the same circuit.

Installation

For CPU execution with NumPy:

pip install -e .

For NVIDIA GPU execution, Triton is the recommended backend. It requires Linux, a supported NVIDIA GPU and driver, and CUDA-enabled PyTorch:

pip install -e '.[triton]'

The triton extra installs PyTorch and Triton; Triton execution is GPU-only. See the Triton backend guide for capabilities, persistent storage, and validation results.

Add pytket to run the Python circuit examples:

pip install -e '.[pytket]'         # CPU
pip install -e '.[triton,pytket]'  # NVIDIA GPU

JAX remains available as a legacy backend for existing workflows. To use it on an NVIDIA GPU, install its CUDA dependencies and select backend_name="jax" explicitly:

pip install 'jax[cuda12]'

Without an explicit backend, spd.create_spo selects Triton when its dependencies and NVIDIA CUDA are available, otherwise NumPy. Existing states keep their backend through forward and backward execution.

What Is In This Repo

• spd/: package code
• examples/: example scripts
• tests/: test suite
• docs/: notes and design docs

Useful entry points:

• spd/run_circuit.py: public workflow helpers such as create_spo, evolve, init_gradient_spo, and backpropagate
• spd/backend_adapter.py: backend selection and configuration
• spd/pytket_frontend.py: pytket frontend

Recommended examples:

• examples/run_simple_circuit_1.py: smallest forward workflow, including truncation info
• examples/gradient/run_tfi_gs.py: 1D/2D/3D variational TFI optimization

Current Scope

• Frontends:
  • built-in OpenQASM 2
  • pytket
• Backends:
  • Triton: recommended NVIDIA GPU backend, including forward, diagnostics, gradients, and analysis
  • NumPy: default CPU backend
  • JAX: legacy CPU/GPU backend
• Circuit model:
  • static circuits
  • no mid-circuit measurement or feedforward

GPU benchmark

On the published 2D Ising scaling workload, SPD Triton measured 8.3–15.9× faster than the published cuPauliProp GPU results across 36–324 qubits. Triton used one A100-SXM4-80GB; the published GPU reference identifies an A100-SXM-64GB. The figure compares three-trial Triton medians with digitized published curves; Triton compilation is excluded.

In the separate 12×12 fixed-observable test, the final step took 0.834 s for Triton versus 11.287 s for cuPauliProp on the same local GPU. All 28 term counts match the published GPU run, with expectation differences below 6e-16. These are forward-plus-expectation timings, without gradient or truncation-history calculation. MonoProp uses a different truncation rule and is faster at the largest scaling sizes; the comparison does not establish a universal speedup. See the benchmark setup, tables, memory, and reproducible data.

Main Workflow

The usual workflow is:

1. create an SPO with spd.create_spo(...)
2. evolve it with spd.evolve(...)
3. read an expectation value with final_spo.get_expectation_value(...)
4. build the terminal gradient object with spd.init_gradient_spo(...)
5. run reverse propagation with spd.backpropagate(...)

Minimal Forward Example

This matches the structure of examples/run_simple_circuit_1.py.

from pytket.circuit import Circuit

import spd

circ = Circuit(3)
circ.Rz(0.5, 0)
circ.Rx(0.5, 1)
circ.ZZPhase(0.25, 0, 2)
circ.measure_all()

trunc_val = 3e-5
max_num_str = int(1e6)

initial_spo = spd.create_spo({"IZI": 1.0})
final_spo, info = spd.evolve(
    initial_spo,
    circ,
    trunc_val=trunc_val,
    max_num_str=max_num_str,
)

exp_val = final_spo.get_expectation_value()

print("expectation value:", exp_val)
print("final SPO size:", final_spo.get_size())
print("tracked truncation steps:", info["num_steps_tracked"])

Light-cone pruning

For local observables, use pruning="light-cone" to plan once or pruning="light-cone-barrier" to refresh support between circuit barriers:

final_spo, info = spd.evolve(
    initial_spo, circ, trunc_val, max_num_str,
    pruning="light-cone-barrier", progress=False,
)

The default is no pruning. Without barriers, both modes use one plan. The result carries one combined record through init_gradient_spo, so backpropagate uses the full original circuit and replays exactly the retained forward gates. Skipped gates perform no truncation or capping. Both modes work across all three backends and in Triton's evolve_step.

See pruning behavior and usage and the benchmark note.

Forward + Backward Example

This is the core workflow used in examples/gradient/run_tfi_gs.py.

import spd

backend = spd.BackendAdapter.from_name("triton", precision="double")
# Use "numpy" above to run this example on CPU.

initial_spo = spd.create_spo(ham_dict, backend=backend)
final_spo, forward_info = spd.evolve(
    initial_spo,
    circ,
    trunc_val=trunc_val,
    max_num_str=max_num_str,
    backend=backend,
)

exp_val = final_spo.get_expectation_value(basis=basis)

initial_spgo = spd.init_gradient_spo(
    final_spo,
    basis=basis,
    backend=backend,
)
final_spgo, raw_grads, backward_info = spd.backpropagate(
    initial_spgo,
    circ,
    trunc_val=trunc_val,
    max_num_str=max_num_str,
    backend=backend,
)

backward_spo = final_spgo.to_spo()
overlap = initial_spo.dot(backward_spo)

In the TFI example, VariationalCircuit.parameter_gradients(...) combines raw_grads into parameter gradients for the optimizer. For overlap diagnostics, to_spo() extracts the primal SPO from the backward object and dot(...) compares matching Pauli-string coefficients. If you want a quantity that should be close to 1, normalize that overlap in user code.

Variational Circuits

VariationalCircuit keeps a pytket circuit together with the mapping from its rotation gates to optimizer parameters. Repeated parameter indices represent shared parameters. parameter_gradients(...) sums their gate gradients and applies pytket's phase-to-angle conversion.

from spd.ansatz import tfi_1d_hva

ansatz = tfi_1d_hva(params, system_size=12, basis="+")

final_spo, _ = spd.evolve(initial_spo, ansatz.circuit, trunc_val, max_num_str)
initial_spgo = spd.init_gradient_spo(final_spo, basis="+")
_, gate_grads, _ = spd.backpropagate(
    initial_spgo,
    ansatz.circuit,
    trunc_val,
    max_num_str,
)
parameter_grads = ansatz.parameter_gradients(gate_grads)

The metadata follows parameterized rotation commands in pytket command order. Do not rebase or structurally modify the circuit after constructing the VariationalCircuit. Fixed rotations use parameter index -1. The optional gate factors describe affine relations such as Rx(params[k] / 2, qubit).

Stable periodic TFI HVA generators are available as tfi_1d_hva, tfi_2d_hva, and tfi_3d_hva in spd.ansatz. Their interaction terms are scheduled as disjoint brickwork layers separated by barriers. Odd periodic dimensions are rejected because they cannot be split into two disjoint even/odd bond coverings. See examples/variational_tfi.py for a complete forward and backward calculation.

L2 Gradient Support

For loss_type="l2_difference", spd.init_gradient_spo(...) builds the terminal gradient object on the support of the current spo.

If you need the union support of the current and target operators, use the backend helper init_gradient_from_l2_difference_union(...) directly.

Input To create_spo

create_spo(...) accepts two common forms:

• a Pauli-string dictionary such as {"IZI": 1.0, "ZZI": -0.5}
• a list of qubit indices such as [0, 2], together with system_size=...

Examples:

spo_1 = spd.create_spo({"Z": 1.0})
spo_2 = spd.create_spo([0, 2], system_size=3)

Backends

spd.create_spo(...) automatically uses Triton on an available NVIDIA CUDA GPU when the optional dependencies are installed, and NumPy otherwise. Explicit selection overrides this choice:

cpu_spo = spd.create_spo({"Z": 1.0}, backend_name="numpy")
gpu_spo = spd.create_spo({"Z": 1.0}, backend_name="triton")  # requires CUDA

For a reusable configuration, use spd.BackendAdapter.from_name("triton", precision="double") and pass backend=backend to the workflow helpers. JAX is available with "jax". Explicit Triton requests raise an error when CUDA is unavailable.

evolve, init_gradient_spo, and backpropagate infer the backend from their input state. Forward and backward evolution preserve their inputs by default. Triton also supports in_place=True in evolve, backpropagate, and backpropagate_noise_analysis to retain mutable storage across circuit calls; NumPy and JAX raise NotImplementedError for this option.

OpenQASM Note

SPD also includes a built-in OpenQASM 2 frontend. If you need it, see spd/openqasm_frontend.py and examples/open_qasm/run_openqasm_file.py.

Rotation Convention

SPD uses

exp(-i * theta * P / 2)

for a Pauli rotation generated by P.

For pytket, this means the frontend converts from

exp(-i * param * pi * P / 2)

to theta = param * pi.

Run Tests

pytest tests

Assumptions

• The examples above use pytket when they build circuits in Python.
• The backward example assumes ham_dict, circ, basis, trunc_val, and max_num_str already exist, just like in examples/gradient/run_tfi_gs.py.