User Guide¶
Clifft separates four decisions that are easy to conflate: how a circuit enters Clifft, how it is compiled, which scientific result to request, and how that work runs on the available hardware. Start with the result you need; execution tuning comes last.
See Performance for near-Clifford comparisons, release history, dense Quantum Volume results, and the measurement contracts behind them.
1. Provide a circuit¶
Clifft's primary circuit format follows Stim and adds non-Clifford gates and additional noise models. Clifft also accepts supported OpenQASM 2 text and circuits from Qiskit or Cirq companion packages.
Circuit Inputs compares these paths, their examples, and their limitations.
2. Compile it¶
For the usual workflow, clifft.compile() turns the input into a reusable Program. The compiler resolves Clifford coordinates, symbolic dependencies, and active-state actions before sampling begins.
Compiling Circuits explains the default path first, then the lower-level inspection and customization APIs for power users.
3. Choose a workflow¶
The workflow determines the meaning and shape of the result:
| Question | Guide |
|---|---|
| I need ordinary shots or post-selected survivors. | Sampling and Results |
| I need rare-event estimates conditioned on fault count. | Importance Sampling |
| I need to model leakage or loss outside the computational subspace. | Leakage and Loss |
| I need exact probabilities for selected states or measurement records. | Exact Probabilities |
The workflow chooser provides more detail when the appropriate API is not obvious.
4. Choose execution settings¶
Most users should keep the automatic batching policy and the single-thread default. CPU Execution and Tuning documents worker budgets, packed sampling, reproducibility, and memory tradeoffs for workloads that need explicit tuning.
The AMD HIP backend is a separate, source-build-only experiment. It is never selected by the regular CPU API.