Distributed MPS capacity example

Evolve and differentiate one open-boundary MPS whose sites are owned by different ranks. The example tests state capacity and a complete training step, rather than throughput from independent replicas.

Start with the bounded profile

From an installed checkout on a host with eight suitable GPUs:

timeout --signal=TERM --kill-after=30s 10m \
  torchrun --standalone --nproc-per-node=8 \
  examples/distributed_mps/variable_bond_capacity_8gpu.py \
  --profile smoke --output /tmp/fq-mps-smoke.json

Inspect all-rank completion, numerical checks, gradient and optimizer updates, truncation, memory, and boundary communication before attempting capacity runs.

Capacity runbook contains the matched baseline commands, workload sizes, historical measurements, and the scientific VQE probe. Keep workload and precision fixed when comparing devices. A completed run is not by itself a release-grade scalability claim.

Capability catalog · Evidence policy