feat(parallel): add topology-aware DLO planning

[0z5a]

Summary

• add a fail-closed parser for nvidia-smi topo -m and a topology-aware DP/SP device-order planner for distributed layerwise offload
• select DLO groups with vLLM-Omni-compatible semantics while excluding TP, and prefer measured collective results over topology-label heuristics
• map logical torchrun ranks to physical local devices without changing LOCAL_RANK; DDP consumes the resolved LOCAL_DEVICE
• add planning and real NCCL benchmark CLIs, regression tests, launch documentation, machine-readable benchmark evidence, and an AstrAI-targeted InfraSWE Draft

Validation

• 631 passed, 103 skipped in the complete AstrAI test suite
• Ruff formatting and import-order checks pass
• InfraSWE Draft resolution: explicit local AstrAI target, no audit flags
• InfraSWE Draft/system-path engine tests: 53 passed

8x RTX 5060 Ti NCCL benchmark

The workload reconstructs a 256 MiB finalized one-byte INT8/FP8 DLO weight and runs a 64 MiB-per-rank SP all-to-all. Each result uses five warmups, 20 iterations, seven trials, and the slowest rank.

Shape	Mapping	Combined median	Result
DP2 x SP4	natural	75.8654 ms	baseline mapping
DP2 x SP4	PHB-paired topology	71.4831 ms	5.78% faster, DLO p99 67.73 -> 52.83 ms
DP4 x SP2	natural	55.2465 ms	9.67% faster than label-only choice
DP4 x SP2	label optimum	60.5890 ms	rejected by measured planner

The DP4 result is an important negative control: minimizing DLO link labels alone can hurt SP traffic, so the planner keeps natural order when the measured combined cost is lower.

MiniMax-H3 end-to-end

Validated current vLLM-Omni with online FP8, DP1 x SP8, text TP8, VAE tile8, and DLO AllGather:

• first valid 2-step request: 17.317 s client E2E
• warm 2-step request: 13.714 s client E2E
• 10-step stability request: 71.719 s client E2E
• all eight GPUs reached 100% utilization; peak memory was 14,206 MiB of 16,311 MiB
• output was verified as H.264 832x480@24 FPS plus stereo AAC 32 kHz

InfraSWE diagnostic

• diagnostic ProjectFit: 86.73/100
• BenchmarkTrust: 93.06/100
• official ProjectFit: unresolved

The diagnostic score is intentionally non-official. The checked-in score card contains every subcomponent input and rationale. Official scoring remains blocked on Draft seal, hidden probes, and a verified evidence manifest.