#1894 bench: community report, RX 6900 XT (gfx1030): STRATA_ROUTE_TAIL_SKIP=7 on HIP, decode x1.24 (Unsloth UD-Q4_K_XL, 0.1.42)

Description

Results only, no engine change. Adds `bench/results/2026-10-10-community-rx-6900xt-tail-skip-hip/` plus its row in `bench/results/COMMUNITY.md` and its entry in `docs/COMMUNITY_BENCHMARKS.md`.

**What it measures.** 0.1.42 turns `STRATA_ROUTE_TAIL_SKIP=7` on by default on CUDA only; on HIP the default stays off. The skip itself is host code (`expert_pool_dispatch_multi` in `src/core/expert_source.cpp`) and reads the variable on any build, so this sets it by hand on a HIP build and compares decode against leaving it unset. Every `tail7` engine log printed `STRATA_ROUTE_TAIL_SKIP=7 is ON` and skipped 77-81K missed experts per start; no `off` log printed either line.

**Machine.** RX 6900 XT 16 GB (gfx1030, PCIe 4.0 x16), Ryzen 9 5950X, 128 GB RAM, Arch Linux, ROCm 7.2.4, Strata `61b3fb5` (0.1.42) HIP source build. Unsloth UD-Q4_K_XL at 131K, int8 KV, `--kv-resident 32768`, 2,831 expert slots in VRAM, 63.5 GiB of experts pinned in RAM, `STRATA_HIP_PROMPT_F16=1 STRATA_SH_STREAM=1` in both arms.

**Method.** 5 rounds, a fresh server per arm and round, order alternating. Per start: a story, a code answer and a 7.9K-token document summary (300 greedy tokens each, all 300 generated), then needle prompts at ~8K and ~32K. Decode tok/s from the server's own `timings`.

| Request | `off` median (range) | `tail7` median (range) | ratio median | `tail7` faster |
| --- | --- | --- | --- | ---: |
| story | 28.3 (26.9-28.7) | 31.8 (26.7-34.3) | 1.116 | 4 of 5 |
| code | 31.9 (31.4-32.4) | 39.6 (39.1-39.8) | 1.238 | 5 of 5 |
| 7.9K document | 28.2 (27.5-29.0) | 35.2 (34.1-36.0) | 1.248 | 5 of 5 |

Pooled: x1.237, faster in 14 of 15 pairs. Needles 10/10 in both arms. Decode hit rate on this machine is about 60-74% without the skip (#1884's V100, where the skip was slower, sits at 99.9%). The story's draft acceptance drops 0.60 → 0.55 with the skip, as #1884 saw, but the time saved on missed experts outweighs it here.

**Not tested:** KL or task scores against the unskipped path, long or sampled replies, other models or AMD cards, Windows HIP, `--batch`, layer splits. One machine.

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

https://claude.ai/code/session_01FwWW9Ex5exNhhuwd3mts7D