#1855 bench: community results on one MI50 32 GB (gfx906), 262,144 context

Beschreibung

Results-only. One AMD Instinct MI50 32 GB (gfx906) runs a Qwen3.8-Flash-Next unsloth-UD-IQ4_XS
pack at the model's native 262,144-token context **on a single card**. The AMD notes list "a
single-card run" as not done for gfx906, and the existing gfx906 report is two 16 GB cards with a layer
split.

**Speed** (the engine's own timing lines; 3 runs each, all `reused 0`, so every run is new work):

| prompt tokens | prompt tok/s | decode tok/s |
| --- | --- | --- |
| 3,141 | 393.8 [361.7..394.3] | 44.7 [38.3..45.4] |
| 26,038-26,040 | 460.7 [457.3..461.5] | 40.7 [38.4..47.5] |
| 93,337-93,339 | 458.8 [458.8..460.7] | 40.7 [35.4..46.8] |

**Recall** (`tools/needle_bench.py`, 32k and 128k at depths 10/50/90): **6 of 6 found**
(126,638-126,640 actual prompt tokens at 128k). 262k is skipped by the script's own context check and is
recorded as a skipped case, not a failure.

The folder also carries a `--kv` comparison that differs in one setting: 4-bit KV frees about 3.8 GiB,
the engine spends it on **1,714 more resident experts** (9,761 vs 8,047 slots), the decode cache hit
rate rises 3.4-7.7 points and decode runs **7.7%-16.6% faster** at equal prefill (within 1%); recall is
6/6 on both arms, 0 disagreements.

**Hardware:** 1x MI50 32 GB (gfx906) on a Gen3 x16 link, Xeon E5-2696 v4 without AVX-512, 128.7 GiB
RAM, ROCm 7.14 from the `mixa3607/rocm-gfx906` image, Strata 0.1.41 release image.

**Limits** (fuller list in `README.md`): single card, single user, no concurrency measurement, TTFT not
measured (non-streaming requests), and the weights are a community abliterated variant rather than an
official quantization.

No engine changes -- results only.