Execution environments

Execution environments: local CPU, Slurm, GPU and QPU

Four independent choices

A run is identified by workflow + software environment + execution provider + resource profile. Changing GPU node or QPU backend is not equivalent to changing the scientific method. The portal owns authorization/catalog; this repository owns runner software and scientific results.

Execution target	Expected path	Caveat
Local laptop / Jupyter CPU	Python 3.14 venv, native scientific SDKs	Good for notebooks and small references
Slurm CPU	authorized submission, versioned OCI→Apptainer runner	Separate workbench from durable job
A100/H200 simulator	GPU-enabled image and scheduler allocation	CUDA/runtime/driver and dependency compatibility essential
Remote QPU	explicit provider credential and allocation	Provider availability, queue, shots/credits and calibration vary

Current templates

The source tree contains Slurm templates under slurm/cpu-smoke.sbatch, hello-ibm.sbatch, trotter.sbatch, and sqd.sbatch; current OCI build contexts are runners/qiskit-cpu/ and runners/sqd/. Treat GPU runners/accelerators and vendor SDK matrices in planning docs as target environments until an actual image, resource policy and acceptance result exist.

CPU → GPU comparison

Run the same circuit, observable, seed/shot plan, tolerance and input with CPU reference first. Pin the correct CUDA-compatible qiskit-aer GPU or PennyLane Lightning GPU dependency in a dedicated runner; default pip install may install a CPU-only wheel. Run inside an allocated GPU node, not a cheap web workbench. Report accelerator model, driver/CUDA versions, method, precision, runtime, memory, power if measured and numerical agreement.

Hardware access

Use a real provider entitlement and official SDK credentials through protected providers/secret brokers. Never paste tokens into notebooks, issues or an agent chat. Reserve hardware use as a deliberate opt-in; plan finite shot budgets, queue delays, noisy baselines and reproducible calibration references.

See environments, workbench-to-execution tutorial and run architecture.