[AlanJager]

Problem and result

This results-only report adds Linux/RTX 3090 evidence for the existing mixed-size resident exchange rotation switch. One unmodified engine 0.1.55 build at c62d5e3a15b40ea51c96aae9e8d36928965f53ab ran four fresh-process A→B→B→A starts; only STRATA_EXCHANGE_ROTATE changed.

512-token case	Off median tok/s	On median tok/s	Change
Python source generation	118.5	137.1	+15.7%
Chinese explanation	86.1	107.5	+24.8%

Code ranges slightly overlap; Chinese ranges are separate. All four natural-stop repairs passed 14/14 checks, with lengths A 499/311 and B 352/1070 tokens. Repair E2E is reported per output, including its higher on-arm median, rather than presented as equal-work latency improvement. Outputs also vary within each arm.

Related discussion

Existing implementation and maintainer measurements: #2042. This PR adds measurements and explanation; it does not resubmit the implementation or change defaults.

Approach

Report, raw records and reproduction include all 28 requests, sanitized logs/configuration, the frozen expert profile, engine/client timings, output text, grader and artifact hashes. An optional standalone HTML explains ownership transfer and displays the same data.

The official 16 staging slots were retained, and CPU fallback occurred. GPU cache geometry matched across all starts. The GPU retains the previous hardware experiment state, disclosed in the report; both arms use that state.

The memory formula covers distinct block sizes and async bounce buffers:

extra exchange payload = (N - b) * (sum(distinct sizes) - largest size)

Here the exchange pool grows from 138.281 to 365.625 MiB: +227.344 MiB pinned host RAM, with GPU expert-cache capacity unchanged. Ownership metadata and driver mapping costs are outside this payload equation and were not separately measured.

Validation

Passed:

• Same-day, same-binary, same-profile ABBA: 28/28 completed streams, no payload-cache reuse, all four repairs natural-stop and 14/14. Individual timing ranges and all outputs are attached.
• Full Release CUDA 12.8 / GCC 11.4 / sm_86 build on Ubuntu 22.04. Exact commands and logs: build.json, build-0.log, build-1.log in the report directory.
• Official ctest --test-dir <build> -R '^(exchange_storage_test|file_expert_source_test)$' --output-on-failure and <build>/file_expert_source_test --rotation-gpu: exit 0, logs attached.
• python3 verify_public.py . from the report directory: hashes, 28 raw records, timing formulas, summary statistics, score counts and toggle-only fingerprint environment verified.
• Original service/source/selectors restored and checked; no new NVRM Xid.

Failed: A first observer admission attempt found no engine PID because it only searched the server main thread's child list. It ended before any inference request. Error/logs and recovery are retained separately; the corrected observer searched all threads and a fresh directory holds the complete ABBA. No inference request was retried or replaced.

Not run: standard benchmark protocol, Compute Sanitizer, native token-ID/logit parity, other hardware/backends and browser screenshot verification for the optional HTML.

Limits

This is a custom workload report under COMMUNITY_BENCHMARKS.md, not a standard-protocol or approved result. It covers one VM, GPU, quantization and one fixed repair task. Publisher revisions are unknown; actual local model/pack/tokenizer SHA-256 values are attached. Seven workers and approximately 295 MiB initial free VRAM trigger tuning warnings, preserved with both arms fixed. Actual GPU paging and the cause of output differences were not isolated. No engine code, preset, default or top-level benchmark index is changed.

---

AI-assisted experiment orchestration and report preparation with Codex.