#1873 Community report: Strix Halo gfx1151 (GMKtec EVO-X2, 128 GB), UD-Q4_K_XL and IQ2_XS at a 262K context

Descrição

## Title
Community report: Strix Halo gfx1151 (GMKtec EVO-X2, 128 GB), UD-Q4_K_XL and IQ2_XS at a 262K context

## Summary
A results-only report from a GMKtec EVO-X2 (Ryzen AI Max+ 395, Radeon 8060S gfx1151, 128 GB), following docs/COMMUNITY_BENCHMARKS.md. It runs Strata 0.1.41 (`fb58e0d`), built from source for gfx1151 with ROCm 7.14.1 TheRock and the 100401 hipBLASLt table. It covers four configurations, 3 runs each on fresh prompts (no prefix reuse) of 1.3K / 10.6K / 31K / 117K tokens, 512 tokens generated, with the engine's own timing lines.

| UD-Q4_K_XL, 262K context | Prompt tok/s (10.6K-117K) | Decode tok/s |
|---|---|---|
| defaults (A, 131K, up to 31K) | 480-510 | 30.6-32.7 |
| section 5 switches + `STRATA_PF_FUSED_KQ=1`, KV int8 (D) | 1,151-1,197 | 33.9-37.8 |
| same with KV fp16 (Q) | 900-947 | 33.2-35.2 |
| IQ2_XS, as D without `--mtp-q4` | 1,118-1,174 | 38.9-41.4 |

## What changed
Adds only `bench/results/2026-10-11-community-gfx1151-evo-x2/`: the README (hardware, build, configurations, method, results, temperature and power, limitations), the 4 server configs and the environment file (no keys), the Dockerfile used (Chainguard / Wolfi base), the benchmark and summary scripts, the per-request JSONL and a 10-second hwmon log.

## Extra Notes
- On unified memory, `--mmap-experts` without `--resident-budget-gib` keeps a single copy of UD-Q4_K_XL's 71.7 GiB of experts (all 24,576 slots resident, about 85-89 GiB of GTT at 262K). The setup's RAM budget would add a second copy in the same pool.
- Decode is below the ~50 / ~43 tok/s of docs/STRIX_HALO.md. Our runs sample (they are not greedy) and accept about 50% of the drafts. We did not run a greedy A/B.
- `--mtp-q4 all` stops the engine on the IQ2_XS pack: #1867.
- Not tested: greedy decoding, concurrent requests, needle tests, the switches one by one, UD-IQ4_XS, the 112 GiB GTT / `amd_iommu=off` host settings.