Parallel Backends

Network construction (make_networks) builds n_nets PC networks on independent cell subsamples. Each network is independent, so the heavy step accepts a backend / n_jobs pair.

Backends

Backend	Parallelism	Extra dependency	Best for
"serial" (default)	none	none	Small data, reproducibility checks, debugging.
"joblib-loky"	processes	none	CPU-bound runs on a single machine; bypasses the GIL.
"joblib-threading"	threads	none	NumPy/BLAS-heavy runs that release the GIL; lower memory.
"ray"	processes / cluster	ray>=2 (extra)	Larger fan-out or running across multiple machines.

Usage

from scTenifold import make_networks

nets = make_networks(data, n_nets=10, backend="joblib-loky", n_jobs=4)

The same applies via nc_kws in the class API or network_kws / backend / n_jobs in compare_networks and virtual_knockout:

result = compare_networks(
    x, y,
    backend="joblib-loky",
    n_jobs=8,
    network_kws={"n_nets": 10},   # backend/n_jobs propagated automatically
)

n_jobs Semantics

• 1 (default): a single worker.
• N > 1: N workers.
• -1: all available cores. Internally mapped to None for joblib and to num_cpus=None for Ray.

Reproducibility

All backends share the same random_state propagated through randomized_svd. The cell-subsample RNG is also seeded from random_state, so identical seeds produce identical networks regardless of backend, modulo BLAS-level non-determinism in the SVD.

Deprecated Alias

n_cpus is kept as an alias for n_jobs during the 0.2.x series and emits a DeprecationWarning. It will be removed in 0.3.