Skip to content

Compute Backends (MLX / CUDA / CPU)

Q-RLSTC supports hardware-accelerated computation for distance calculations, array operations, and optimizer steps via three backends.

Backend Priority

When --backend auto is used (default), the system auto-detects:

  1. Apple MLX — M1/M2/M3/M4 Apple Silicon via mlx.core
  2. NVIDIA CUDA — via PyTorch (torch.cuda)
  3. CPU — NumPy fallback (always available)

Module: q_rlstc.accelerator

Component Purpose
ComputeBackend (Enum) CPU, MLX, CUDA, AUTO
detect_best_backend() Auto-detect best available backend
resolve_backend(name) Resolve "auto"/"cpu"/"mlx"/"cuda" to ComputeBackend
AcceleratedArray Unified array wrapper (NumPy/MLX/Torch)
pairwise_distances(X, Y) Accelerated Euclidean distance matrix
get_device_info() Dict of available hardware and versions

CLI Usage

# Auto-detect (recommended)
python experiments/run_full_benchmark.py

# Force specific backend
python experiments/run_full_benchmark.py --backend mlx
python experiments/run_full_benchmark.py --backend cuda
python experiments/run_full_benchmark.py --backend cpu

# Compare all available backends
python experiments/run_full_benchmark.py --compare-backends

Python API

from q_rlstc.accelerator import (
    resolve_backend,
    pairwise_distances,
    get_device_info,
)

# Check what's available
info = get_device_info()
print(info)
# {'numpy_version': '1.26.4', 'backends': {'cpu': True, 'mlx': True, 'cuda': False}, 'best': 'mlx'}

# Use accelerated distances
backend = resolve_backend("mlx")
distances = pairwise_distances(X, Y, backend=backend)

Configuration

The QRLSTCConfig master config includes a compute_backend field:

from q_rlstc.config import QRLSTCConfig

config = QRLSTCConfig(
    version="A",
    compute_backend="mlx",  # "auto", "cpu", "mlx", "cuda"
)

Installing Backend Dependencies

# Apple MLX (macOS Apple Silicon only)
pip install mlx

# CUDA (requires NVIDIA GPU + CUDA toolkit)
pip install torch  # with CUDA support

# CPU — no extra dependencies needed

Backend Comparison Plots

When running --compare-backends, the CLI generates a timing comparison chart in the output directory:

outputs/benchmark_20260220/plots/backend_comparison.png

This shows grouped bars of runtime per run per backend, making it easy to see where MLX or CUDA provide speedup.

Important Notes

  • Quantum circuit simulation always uses Qiskit Aer (CPU-based) regardless of compute backend. The compute backend only accelerates array operations (distance matrices, clustering, feature extraction).
  • MLX uses float32 precision for best GPU utilisation on Apple Silicon.
  • CUDA uses float64 to match NumPy CPU precision exactly.
  • Lazy imports: mlx.core and torch are only imported when their backend is selected.

← Visualization & Plotting | Benchmarking →