[Benchmark] record the layer split (stages, split point, trim, residency) in the standard report #4495

Area

Benchmark

What do you want to do?

Record a layer split in the standard report: which card runs which layers, the split point, --trim-stage-weights, and the expert residency per stage.

Today hardware.gpus lists the cards in nvidia-smi order with no role or stage, configuration.offloading holds --layer-split as a bare option, and --trim-stage-weights, --peer-device and --ple-io fall into extensions.other_engine_options (GROUPS in strata_cli/bench_facts.py ~507-516). docs/BENCHMARK_PROTOCOL.md (~236) asks for "all inference GPUs, selected device indices", but selected is computed and then dropped. The server already has what is needed: /v1/telemetry carries topology.stages and selected_devices (serve/server.py ~7020-7040), and the serving-suite runner records them; the standard runner never asks. Expert residency per stage and the per-request hit_rate / pcie_share the server keeps (~5661-5683) have no field either, though they decide what a split's prompt and decode numbers mean.

Suggested: an extensions.topology block (stages with device index, name, layer range, trim, resident experts and VRAM) taken from /v1/telemetry, plus a per-case median of hit_rate and pcie_share in the summaries. Two-card reports (#4468 among them) would then be comparable without reading the engine log.

What do you do today instead?

Describe the split by hand in the README and attach the engine log where the runner kept it.

Hardware: 2x RTX 3090 24 GB (PCIe 3.0 x16, one per socket), 2x Xeon E5-2620 v3, 256 GB, Ubuntu 26.10. Version: 0.2.7 (engine 0.1.52, Control 0.1.52); code checked on main 46c152a. Model: Flash-Next IQ3_S, layer split 25 with --trim-stage-weights, conversation cache 40 GiB / 16 slots; the full profile is in #4468.