Contraction Path Optimization in cuquantum.tensornet.experimental: Is Fine-Grained Control Possible?

[seba2390]

As of cuQuantum 25.03.0, there are several ways to calculate expectation values—i.e., to contract the tensor network corresponding to ⟨ψ|𝐻|ψ⟩ for some operator 𝐻 and quantum circuit state |ψ⟩.

In practice, the quality of the contraction path has a significant impact on both the speed and memory consumption of the contraction. When using the cuquantum.tensornet.contract(...) interface, the user can provide OptimizerOptions as follows:

from cuquantum import contract
from cuquantum.tensornet import (
    OptimizerOptions,
    PathFinderOptions,
    ReconfigOptions,
    SlicerOptions,
)

pathfinder_options = PathFinderOptions(
    num_partitions=...,
    cutoff_size=...,
    algorithm=...,
    imbalance_factor=...,
    num_iterations=...,
    num_cuts=...
)

slicer_options = SlicerOptions(
    disable_slicing=...,
    memory_model=...,
    memory_factor=...,
    min_slices=...,  
    slice_factor=...
)

reconfig_options = ReconfigOptions(
    num_iterations=...,
    num_leaves=...
)

optimizer_options = OptimizerOptions(
    samples=...,
    threads=...,
    path=pathfinder_options,
    slicing=slicer_options,
    reconfiguration=reconfig_options,
    seed=...,
    cost_function=...,
    smart=...
)

out = contract(..., optimize=optimizer_options, ...)

This allows for fairly fine-grained control over the contraction path optimization process. However, in my case, both the operator 𝐻 and the state |ψ⟩ can be described using only real numbers. As discussed in post #182, the only proper way to enforce a real dtype (which is meaningful as it presumably reduces memory usage and speeds up contraction) is by using the circuit = cuquantum.tensornet.experimental.NetworkState(...) and hamiltonian = cuquantum.tensornet.experimental.NetworkOperator(...) interfaces, followed by computing the expectation value via circuit.compute_expectation(hamiltonian, ...).

Another reason for choosing this interface is that it supports a degree of caching in a VQE-like setting, as outlined in the cuQuantum documentation.

My question is therefore: is there any way to achieve the same level of control over contraction path optimization when using the cuquantum.tensornet.experimental layer?

Best regards,
Sebastian Yde Madsen

[yangcal]

Thanks for reporting this need. Currently we don't support all these fine-grained control over the path optimization and I will report to the internal dev team