FlagOS statevector capacity expansion (F5)

F5 is a matched, fail-closed development measurement of one statevector that does not fit on a single A800 but completes when sharded across eight A800s through FlagQuantum's public backend="flagos" and device="flagos" boundary.

The authoritative artifact is artifacts/flagos_statevector_capacity_f5_a800_20260827.json, with SHA-256 128a6e7db8eee153d1bf1577d16ca55a38a7f7718d6da4661576ff9a759903cb.

Measured result

The exact matched workload is a 32-qubit, complex128, forward-only circuit. Its logical statevector contains $2^{32}$ amplitudes and requires 64 GiB of state storage before runtime scratch space.

Mode	World size	Result	Exact boundary
Single device	1	Expected OOM measured	64 GiB allocation failed on one A800; peak allocated bytes 68,719,477,248
Replicated	8	Expected OOM on every rank	Each rank attempted the same full 64 GiB state and failed
Sharded	8	Passed	Each rank owned 536,870,912 amplitudes (8 GiB state); no full-state materialization

Every sharded rank completed in approximately 1,910 seconds, reported zero global norm error at the complex128 tolerance of 2e-11, recorded 128 communication operations and 17,179,869,184 communication bytes, and stayed below physical memory with peak allocated bytes 21,483,356,160 out of 85,093,777,408 bytes.

The capacity run performs one full-width forward execution and validates its global norm with FP64 scalar aggregation. It deliberately does not replay the 64 GiB logical workload: deterministic numerical behavior is covered by the bounded F2 correctness cases, while F5 answers only the matched capacity question. An initial two-replay protocol reached its 3,600-second subprocess limit and was rejected; it is not counted as successful evidence.

Evidence boundary

This result demonstrates development capacity expansion for this exact single-node workload. It does not establish general scalability, performance, convergence, backward or optimizer capacity, multi-node behavior, production support, or release readiness.

The measured outer backend is FlagOS. The provider's inner communication route was not independently identified, so the artifact records:

• inner_communication_route = "unattributed";
• flagcx_route_verified = false;
• host_staging_observed = null;
• communication, scalability, production, and release claim flags as false.

The environment used FlagQuantum source revision 348482a8f1278bcd406f66126f99bdf52755e6f5, Torch-FL source revision 1b19383c24d9aa54637dd23557a5a20ef6d52ec5, PyTorch 2.11.0+cu130, and eight NVIDIA A800-SXM4-80GB devices. Torch-FL was built in an isolated CUDA-boxing reference environment with FlagGems disabled.

Reproduction and audit

The frozen manifest is benchmarks/manifests/flagos_statevector_capacity_f5.json. On an isolated eight-device FlagOS environment, run:

python tools/validate_flagos_statevector_capacity.py \
  --manifest benchmarks/manifests/flagos_statevector_capacity_f5.json \
  --output artifacts/flagos_statevector_capacity_f5_a800_20260827.json \
  --timeout-seconds 900

The controller runs single, eight-rank replicated, and eight-rank sharded modes against the same canonical workload hash. The checked-in benchmark-contract test reconstructs the profile from raw rank records and rejects mismatched workloads, incomplete ranks, unmeasured OOMs, full-state materialization, missing communication, invalid norm evidence, memory overflow, and provider route overclaims.