GPU optimization / live API

Bounded-bond tensor-network optimization.

A real transfer-tensor contraction for sparse QUBO and Ising models. The frontier is truncated to a declared bond dimension, every returned objective is recomputed, and heuristic runs never claim a certified bound or global optimum.

Maximum variables

256

Bond dimension

2–4,096

A10G p95 / 256

80.840 ms

Production boundary

Exact enumeration is available through 18 variables. Above that threshold, bounded-bond-transfer-contraction-torch-v1 is explicitly heuristic. Inputs are bounded inline data and are not echoed by the job API.

const result = await maha.optimization.solveTensorNetwork({
  clientRequestId: crypto.randomUUID().replaceAll('-', ''),
  problem: { formulation: 'qubo', size: 3, terms },
  solver: { bondDimension: 256, exactThreshold: 18 }
})