Keeping intermediates within target_size

Start from an unsliced path and use Fixed or Dynamic slicing to control intermediate tensor sizes.

First prepare the example network structure, dimensions, and numerical arrays from Tensor-network input in Quick start.

Fixed-path slicing

Use your own inputs, output, and size_dict, or reuse the network in Quick start. For the summation principle, see Slicing principles.

from arctn import arctn_schedule

fixed_report = arctn_schedule(
    inputs, output, size_dict,
    preset="heavy", seed=0,
    target_size=4,
    slicing_mode="fixed",
    use_ssa=True,
)
print(fixed_report["sliced_legs"])
print(fixed_report["sliced_log2_max_size"])

For the quick-start network, target_size=4 actually triggers slicing. fixed is the default mode. It first obtains an unsliced path with the selected Light or Heavy preset, then keeps that path fixed and chooses only the sliced indices.

Alternating slicing and local path reconfiguration

Run Dynamic on the same network and inspect both returned paths in SSA format:

dynamic_report = arctn_schedule(
    inputs, output, size_dict,
    preset="heavy", seed=0,
    target_size=4,
    slicing_mode="dynamic",
    use_ssa=True,
)
print(dynamic_report["path"] == fixed_report["path"])
print(dynamic_report["sliced_legs"])
print(dynamic_report["sliced_log10_flops_total"])

Both fixed and dynamic first optimize an unsliced path with the selected Light or Heavy preset. dynamic permits limited local path changes after slicing. The Boolean above only indicates whether the two calls returned the same final path; it is not a record of changes inside one search. Compare slice count and total work as well as whether the path changed.

Comparing slicing settings

target_size constrains only the element count of per-slice binary contraction results; for a single-tensor network, it checks the final output. It does not limit all unary-operation temporaries and is not a cap on total process memory, GPU memory, or temporary workspace.