---

name: core-solvers description: "Use POT core solvers for exact and regularized optimal transport on cost matrices or sample clouds, including unified APIs, EMD, Sinkhorn, 1D, circle, sparse, lazy helpers, and OTResult validation." disable-model-invocation: true metadata: disco-role: operating license: MIT

POT Core Solvers

Use this sub-skill when the task is balanced discrete optimal transport between histograms or sample clouds: exact network-simplex OT, entropic or L2 regularized OT, ot.solve, ot.solve_sample, ot.emd/emd2, ot.sinkhorn/sinkhorn2, cost construction with ot.dist, uniform weights with ot.unif, and OTResult inspection.

For verified signatures, return objects, solver routing, weight conventions, sparse/lazy support, gradient modes, and parameter defaults, read references/api-reference.md.

For copy-pasteable recipes that convert between classical and unified APIs, validate OTResult attributes, handle sample-cloud workflows, and run 1D/circle/sparse/lazy checks, read references/workflows.md.

For failure diagnosis covering mass mismatch, invalid weights, too-small regularization, invalid methods, compiled EMD import/build issues, n_threads/OpenMP caveats, sparse matrices, and gradient memory, read references/troubleshooting.md.

To verify a tiny deterministic installation-level core-solver path without plotting or external data, run scripts/core_solver_smoke.py with python scripts/core_solver_smoke.py --mode all.

Route barycenters to barycenters, GW/FGW to gromov, unbalanced or partial OT workflows to unbalanced-partial, optional backend installation and batched arrays to backend-and-batch, and large-scale approximation families such as sliced, Gaussian, low-rank, Nystroem, factored, or BSP-heavy workflows to sliced-gaussian-large-scale unless this task only needs to recognize the solve_sample entry point.