#3646 [Benchmark] RX 6800 XT (gfx1030) / Flash-Next IQ3_S / ROCm 10.0 / v0.2.1 + v0.2.2 (vibe-coded)

Description

**Vibe-coded disclosure:** the owner of this RX 6800 XT directed the work; an AI agent (Anthropic's Claude Code) installed ROCm, built the engines, ran every benchmark and wrote the report and this PR. All numbers are measured outputs (per-run JSON included); unmeasured items are marked.

## Problem and result
No report covered a **single RX 6800 XT (gfx1030)** on Linux, and the RDNA2 section asks for numbers from 6800-class cards and for what a card-calibrated rocBLAS table changes. Results-only PR, no engine changes:
- `bench/results/2026-10-21-community-rx-6800-xt-linux/`: community report (README + per-arm `summary.json` + two rocBLAS tables calibrated on this card).
- `bench/results/2026-10-21-standard-unknown-35dea42b/`: `strata bench run` (protocol `strata-qwen38-thinking-v1-draft.1`, track **custom**), unedited runner output. The runner could not read the GPU name ("AMD Radeon"), hence `unknown` in the folder name; it is the RX 6800 XT.
- one entry in `docs/COMMUNITY_BENCHMARKS.md`.

Highlights (original Flash-Next IQ3_S, 131K context, ROCm 10.0.0, kernel 7.0.0-38, Ryzen 9 5950X, 128 GB):
- v0.2.1, stock vs `STRATA_HIP_PROMPT_F16=1 STRATA_SH_STREAM=1 STRATA_HIP_ADAPT_KERNEL_COPY=1`: prompt 521 / 546 / 525 -> **986 / 1,112 / 1,033 tok/s** at 4K / 32K / 128K, decode 48-52.
- v0.2.2: same speed as v0.2.1 here; `--prefill auto:32768` +3% at 32K, **+4.4% at 125K** (1,010 -> 1,054).
- `--resident-experts`: -21% at 4K although fully covered (`1874 of 1874` lendable slots, file tier 0.0 GiB); the log says adaptive overlap is off in resident RAM mode. With 128 GB the default arena is already in pinned RAM.
- rocBLAS table: none vs shipped 6900 XT table = +14% at 4K; a table calibrated on this card (`tune_rocblas`) = the shipped one. ROCm 10.1.0 = 10.0.0.
- `STRATA_DENSE_MMQ`, `STRATA_SELECT_SIMT`, `STRATA_HC_PERSIST`: within noise at 4K/32K on one card.
- Conversation cache 24 GiB: a 7.7K-token chat resumed after another request in 2.5 s vs 8.6 s (7,688 tokens restored in 36 ms).
- Standard benchmark: 21/21 requests, **12/12 sanity**, prompt 952 (4K) / 1,100 (16K) tok/s, decode 43-48 tok/s.

## Related discussion
none (results-only report, per CONTRIBUTING: benchmark reports can go straight to a PR).

## Approach
`bench/results/2026-10-09-community-2x-rx-6900xt-0.1.41/benchmark.py` unchanged (fresh nonce prompts, greedy, 256-token cap, server timings), a fresh server per arm, 2 runs per size; same model files (same sha256) as the 2x RX 6900 XT reports. Source builds via `./setup.sh --backend hip` against a system ROCm 10.0.0 copied out of AMD's `rocm/dev-ubuntu-24.04:10.0.0-full` image (the README notes the `/etc/alternatives` symlink trap). Everything personal was left out; hardware, versions, settings and data are included.

## Validation
**Passed:**
- `strata bench run --url http://127.0.0.1:8080 --track custom ...` on v0.2.2 `011c9e60` (engine 0.1.47), freshly loaded: complete, 21/21, 12/12 sanity, 4 min 2 s.
- community `benchmark.py` arms listed in the report, v0.2.1-3-gc093ac26 (10-20) and v0.2.2 (10-21), every arm 2/2 requests per size.
- No secrets or personal data in the added files (checked before commit).

**Failed:**
- none observed. (One arm first ran with mis-named table files, so the engine ran without a table; it is reported as "no rocBLAS table", then rerun with correct names.)

**Not run:**
- needle/recall in the community sweep; two-card layer split; Windows; the `rx6000` engine package; `--resident-experts` with `STRATA_ADAPT_OVERLAP_RESIDENT=1`.

## Limits
One machine, one card, two repetitions per arm, single sessions; 1-3% differences are noise. On 10-21 the card ran at PCIe 4.0 x8 (an idle RTX 4070 in the second CPU slot). The answer-quality figure in the README comes from a private suite and is context only.

🤖 Generated with [Claude Code](https://claude.com/claude-code)