This paper introduces ATHENA, a compiler designed to reduce costly qubit teleportations in distributed quantum computers. Distributed machines link small quantum chips with photonic channels. To run a two-qubit operation between chips, the system must teleport qubits or use special remote gates. Those non-local operations are much slower (4.3–7.7×) and have higher error rates (about 4×) than local operations inside a chip, so reducing them helps program fidelity and speed.

Existing compilers group two-qubit gates with overlapping qubits into blocks and optimize teleportations within each block. That block-level approach has two main problems. First, it picks the best schedule for one block before looking far enough ahead, so a teleportation that seems good now can force many extra teleportations later. The authors show a concrete example where using the next 10 blocks for lookahead still leaves more teleportations than the true optimum. Second, compilers often delay planning teleportations for future blocks until earlier blocks are fully scheduled. That deferral creates waiting and extra latency; in one benchmark the authors found 45.4% of teleportations waited about 7.5 ms on average.

ATHENA attacks both problems with two ideas. The first is Utility-driven Lookahead with Multi-candidate block Scheduling (UMS). Rather than trying to consider all future blocks, ATHENA only looks ahead to future blocks that are “useful” — those that share qubits with the block currently being scheduled. UMS also keeps several candidate schedules at once instead of committing early to the single best-looking choice. This lets the compiler avoid getting stuck on a short-term choice that hurts long-term performance. To keep runtime practical, ATHENA prunes the search tree and keeps only the top-w candidate schedules.

The second idea is EPR-Capacity-Aware Early Scheduling (EES). Teleportation uses entangled pairs called EPR pairs, and each chip can hold only a limited number of external qubits at a time (its EPR capacity). EES identifies gates that could be scheduled early once their data dependencies are resolved, and it schedules their teleportations early only if EPR capacity is available and doing so does not increase total teleportations. The compiler tracks how early moves change qubit positions and EPR availability, and avoids early scheduling that would raise costs for intermediate operations.