#1978 bench/results: gfx906 IQ3_XXS (2x Instinct MI50 / Radeon VII), and the same run with --mmap-experts

说明

Two folders from one machine, deliberately one flag apart — same box, same model, same config:

- resident expert arena (no `--mmap-experts`): `bench/results/2026-10-11-community-2x-instinct-mi50-32gb-qwen3-8-flash-next-uncensored-iq3-xxs/`
- arena read through the page cache (`--mmap-experts`): `bench/results/2026-10-11-community-2x-instinct-mi50-32gb-qwen3-8-flash-next-uncensored-iq3-xxs-low-ram/`

Both are added to the list in `docs/COMMUNITY_BENCHMARKS.md` as well.

**Hardware and model.** AMD Instinct MI50 32 GB + Radeon VII 16 GB (gfx906), Core i5-12400F, 61.6 GiB RAM.
Qwen3.8-Flash-Next-Uncensored IQ3_XXS (pack `orca-iq3_xxs`, profile `expert-profile.bin`), context 131,072,
KV q4_0, prefill 2048, MTP spec 4, `--layer-split auto`, no `--resident-*` flags.

**How it was run.** `tools/strata_bench.py --url http://127.0.0.1:8085 --runs 3 --clients 1`, against a server
that was already running — hence no engine-log excerpt in the folders.

**What the one flag is worth on a model that fits in RAM** (medians over 3 runs):

| | resident | `--mmap-experts` | |
| --- | ---: | ---: | --- |
| 4K prompt tok/s | 546.9 (546.5-549.5) | 473.6 (375.1-477.0) | −13.4% |
| 16K prompt tok/s | 621.8 (619.7-622.0) | 543.9 (543.2-546.1) | −12.5% |
| 4K / 16K first token | 7.38 / 25.87 s | 8.52 / 29.56 s | +14-15% |
| chat decode | 44.0 (41.9-62.1) | 41.2 (34.8-53.0) | ranges overlap |
| agent decode | 41.7 (34.9-62.0) | 44.7 (33.7-53.2) | ranges overlap |
| 1 concurrent client | 31.7 tok/s | 32.5 tok/s | same |

On this model the page-cache mode costs the prompt path about 12-13% and the first token 14-15%; decode is
not attributable at n=3. The wide range in the mmap arm's 4K row is one run paying for a cold page cache
(375.1). Not measured: the case low-RAM exists for — an arena larger than RAM, where the resident mode cannot
run at all.

**Provenance — these are not stock 0.1.43 numbers.** They come from the gfx906 image we maintain
(`xxdoman/strata-mi50`, engine built from tag v0.1.43). That build carries three HIP build gates (without
them the 0.1.40+ tree does not compile on gfx906) and the rocBLAS solution-table reader — the idea is in
#981, which is gated behind a path gfx906 never compiles, so we widened the gate. The image enables that
table together with the FP16 prompt path it needs (`STRATA_HIP_PROMPT_F16=1`, which the engine calls "not
bit-identical to the default"). The table is a prefill lever, not a decode one: measured by environment
switch on the same engine, +9.5% / +10.8% prompt at 16K / 32K, decode unchanged. Model files, packs and the
harness are the stock ones.

If it is more useful as a single folder, say which one and I will drop the other.