contraction-tuner

Contraction Tuner

Use this skill when a task asks to optimize, compare, or benchmark tensor-network contraction paths, slicing targets, OMECO, cotengra, contraction width, FLOPs, write, or slice count for TensorCircuit-NG.

Core Principle

Treat contraction tuning as an empirical search problem. Do not infer quality from width alone. Always report at least:

• path/search time and slicer/reconfiguration time separately
• log10(total FLOPs)
• log2(max intermediate size)
• log2(total write)
• number of sliced indices and total slice tasks
• package versions and topology size (ntensors, nindices)

For dim-2 circuit indices, k sliced indices means 2**k slice tasks. A path with more slices can still have lower total FLOPs if per-slice contractions are much cheaper.

Metric Hygiene

• cotengra target_size is a linear tensor element count. For target width W, pass target_size=2**W, not W.
• cotengra ContractionTree.total_flops() and total_write() include slicing multiplicity.
• OMECO SlicedEinsum.complexity() is per-slice. For fair comparison, convert the OMECO sliced tree to a cotengra ContractionTree and apply remove_ind_() for each sliced.slicing() label, or multiply per-slice FLOPs/write by prod(size_dict[ix] for ix in sliced.slicing()).
• OMECO SlicedEinsum.num_slices() is the number of sliced labels, not total slice tasks.
• A result where sliced.slicing() is empty but width improves means OMECO slicer reoptimized the tree enough to meet the target without actual slicing.

OMECO Workflow

Use OMECO as the default first pass for large/deep tensor networks because it often finds strong seed trees much faster than cotengra.

Recommended seed search grid:

seed_configs = ["8x48", "12x64", "24x64"]
score = omeco.ScoreFunction(
    tc_weight=1.0,
    sc_weight=0.0,
    rw_weight=64.0,
    sc_target=64.0,
)

Parse AxB as ntrials=A, niters=B, and use len(betas)=B, for example:

betas = [float(1.0 / x) for x in np.geomspace(2.0, 0.05, niters)]
tree = omeco.optimize_code(
    inputs,
    output,
    sizes,
    omeco.TreeSA(ntrials=ntrials, niters=niters, betas=betas, score=score),
)