#3426 [Benchmark] 2x RX 6900 XT (gfx1030), 0.2.0-beta stock, Flash-Next IQ3_S: strata_bench suite + real-text reads and decode

说明

## Problem and result

A results-only report (no engine change): upstream 0.2.0-beta stock on 2x RX 6900 XT (gfx1030, layer split), Flash-Next IQ3_S, measured with `tools/strata_bench.py` and, on the same server, real-text prompt reads and single-request decode. Folder: `bench/results/2026-10-20-community-2x-rx-6900xt-0.2.0/` (README in the template's format + `results.json`).

| | 4K | 16K | 32K | 119K real text |
|---|---|---|---|---|
| prompt tok/s (median, 3 runs) | 979.0 | 1,415.6 | 1,661.5 | 1,893 |

Short chat decode 61.6 tok/s; one request summarizing 4K / 30K tokens of real text: 66.5 / 66.0 tok/s; agent session cache hit 69.5%; two clients 42.8 tok/s together.

## Related discussion

None. Earlier reports from this box: #927, #1270, #1657, #1927, #2237 (older versions).

## Approach

Main e88ee39d (v0.2.0-beta + 12 docs/packaging commits), source build with setup.py's HIP options (`STRATA_PREFILL_MMQ`, `STRATA_Q6K_EXPERTS`, `STRATA_MMQ_KQUANTS` on, gfx1030, the pinned ggml), no local patches; the configuration and the non-default environment are listed in the README. A cold server, then the suite, then the real-text runs.

## Validation

**Passed:** `python tools/strata_bench.py --url http://127.0.0.1:8089 --clients 2 --long --runs 3` (224 s, 0 errors); real-text reads 3x each with 0 tokens reused; the same reads on 20 more cold starts of builds whose default path is byte-identical to this one (two PR branches with their variables unset) agree within 0.7%. The README and `results.json` were checked for home paths, host names and addresses.

**Failed:** none observed.

**Not run:** correctness checks (speed only); a single-card or helper configuration; today's main 9ae93f7b (the 0.2.1 candidate).

## Limits

One machine, one model. The real-text corpus is private (mixed English / Chinese / code notes) and not attached; the README describes how the prompts were built. With batch slots on a layer split, 0.2.0 turns `--pipeline-windows` and `--batch-mtp` off, so the single-request decode here is lower than this box reaches with them (PR #3095).

Developed with an AI coding assistant; every number above was measured on the machine described here.

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