Multi-GPU Workflows
mgpu) and multi-QPU (mqpu) backends are planned; the CPU-only reference build does not include them. This page documents the intended interface.Scale beyond a single device by pooling GPU memory (mgpu) or running
many QPUs in parallel (mqpu).
From CPU to GPU
All code on this page is planned (规划中); the CPU reference build ships none of these backends, so every snippet is commented out as an intended-interface illustration.
# Planned — not in the CPU reference build:
# import qalgora
# qalgora.set_target("gpu") # single-GPU statevectorPooling memory across GPUs (mgpu)
One large statevector would be distributed across the memory of multiple GPUs, letting you reach more qubits than fit on a single card.
# Planned (mgpu) — not in the CPU reference build:
# qalgora.set_target("gpu", option="mgpu") # one big state, many GPUs
# counts = qalgora.sample(large_kernel)Parallel virtual QPUs (mqpu)
The mqpu platform would expose each GPU as an independent virtual QPU — here a
"virtual QPU" is a simulator worker, not a physical QPU. Work would be sharded across them.
# Planned (mqpu) — not in the CPU reference build:
# qalgora.set_target("gpu", option="mqpu") # independent virtual QPU per GPU
# energy = qalgora.observe(ansatz, hamiltonian, theta).expectation()Batching Hamiltonian terms
A large Hamiltonian has many Pauli terms. Under the planned mqpu backend,
observe would split the terms across virtual QPUs and sum the partial expectation
values — a long chemistry Hamiltonian could be evaluated with less wall-clock time when enough
independent devices are available; the speedup is workload- and hardware-dependent.
# Planned — 0..N qubit Hamiltonian with hundreds of terms run across all virtual QPUs:
# energy = qalgora.observe(ansatz, big_hamiltonian, params).expectation()Circuit batching
Independent circuits (a parameter sweep, a batch of shots, gradient evaluations) would be dispatched asynchronously, one per virtual QPU:
# Planned — one async circuit per virtual QPU:
# futures = [qalgora.observe_async(ansatz, hamiltonian, t, qpu_id=i % n_qpus)
# for i, t in enumerate(angles)]
# energies = [f.get().expectation() for f in futures]mgpu when one circuit is too big for a single GPU; use
mqpu when you have many independent circuits or Hamiltonian terms to evaluate at once.
They are designed to compose — a multi-node cluster would run mqpu with each virtual
QPU itself an mgpu pool.
多 GPU 工作流
mgpu）与多 QPU（mqpu）后端均属规划中，CPU-only 的参考实现并未包含。本页仅用于说明预期接口。通过显存池化（mgpu）或多 QPU 并行（mqpu）突破单卡限制。
从 CPU 切换到 GPU
本页所有代码均属规划中（规划中）；CPU 参考实现并未包含这些后端，因此每段代码都以注释形式给出，仅用于说明预期接口。
# 规划中——CPU 参考实现暂未包含：
# import qalgora
# qalgora.set_target("gpu") # 单 GPU 态矢量跨 GPU 显存池化 (mgpu)
将一个大型态矢量分布在多块 GPU 的显存中，使可模拟的量子比特数超出单卡上限。
# 规划中 (mgpu)——CPU 参考实现暂未包含：
# qalgora.set_target("gpu", option="mgpu") # 一个大态矢量，跨多块 GPU
# counts = qalgora.sample(large_kernel)并行虚拟 QPU (mqpu)
mqpu 平台会将每块 GPU 作为独立的虚拟 QPU 对外暴露，工作负载分片到各虚拟 QPU 上。这里的“虚拟 QPU”指模拟器工作单元，而非物理 QPU。
# 规划中 (mqpu)——CPU 参考实现暂未包含：
# qalgora.set_target("gpu", option="mqpu") # 每块 GPU 一个独立虚拟 QPU
# energy = qalgora.observe(ansatz, hamiltonian, theta).expectation()哈密顿量项批处理
大型哈密顿量包含大量 Pauli 项。在规划中的 mqpu 后端下，observe 会将这些项分片到各虚拟 QPU 并汇总部分期望值——当有足够多独立设备时，可降低含大量项的化学哈密顿量的墙钟时间；加速比取决于工作负载与硬件。
# 规划中——含数百项的 0..N 比特哈密顿量，各项跨所有虚拟 QPU 运行：
# energy = qalgora.observe(ansatz, big_hamiltonian, params).expectation()电路批处理
独立电路（参数扫描、一批采样、梯度求值）会异步分发，每个虚拟 QPU 承接一个：
# 规划中——每个虚拟 QPU 承接一个异步电路：
# futures = [qalgora.observe_async(ansatz, hamiltonian, t, qpu_id=i % n_qpus)
# for i, t in enumerate(angles)]
# energies = [f.get().expectation() for f in futures]mgpu；需要同时求值多个独立电路或哈密顿量项时选 mqpu。两者设计为可组合使用：多节点集群会在 mqpu 模式下运行，每个虚拟 QPU 本身即为一个 mgpu 显存池。