#2225 Configuration report: RTX 5090, IQ3_S + INT8 KV + BF16 vision, four 192K slots

Descripción

## Configuration/compatibility field report

Sharing a bounded long-context serving result, not a bug report or an engine-comparison benchmark. Preparation pinned Strata `7695859aa13afb248bdb4d36f9b2e95707db079d` and llama.cpp `3cf03257`; final runtime logs identify engine 0.1.43. The final runtime directory was an archive without Git metadata, so the preparation revision is provenance, not an independently verified final-tree identity. Final engine SHA256: `28d54efd142f75b47c9b1bc53cb213b6471fdd4be552ff819757a8c72d643621`.

### System and model

- Linux VM, RTX 5090 32 GB (compute capability 12.0), 16 vCPUs, 64 GiB guest RAM, 62 GiB container limit; CUDA 13.4 build environment.
- Qwen3.8-Flash-Next-GSQ-RCO IQ3_S, two GGUF shards, native packed experts; BF16 mmproj on GPU.
- INT8 KV, four serving slots, context 196608 per slot, MTP four drafts; suffix drafting and lookup chain disabled.
- Expert RAM budget 12 GiB, host conversation cache 12 GiB / four slots / 8 GiB minimum-free floor, KV resident 32768, VRAM reserve 1400 MiB.

Relevant engine options (model/pack paths omitted):

```text
--serve --expert-cache auto --prefill auto --spec 4 --spec-min-p 0.5
--max-context 196608 --kv int8 --vision --vram-reserve-mib 1400
--conversation-cache-mib 12288 --conversation-cache-slots 4
--conversation-cache-min-free-mib 8192 --kv-resident 32768
--resident-budget-gib 12 --suffix-draft 0 --lookup-chain 0 --batch 4
```

### Synthetic workload results

A historical exact-profile rotation used five long synthetic conversation primes followed by 15 four-request full-history continuation batches, with 512-token output budgets and retrieval/word-limit checks scored separately:

- Retrieval: 65/65; word-limit adherence: 65/65.
- Continuation prefix-cache hits: 37/60 (not all full-history requests hit).
- Median TTFT: 52.40 s; median per-request decode: 23.00 tok/s, not aggregate server throughput.

A separate fresh-server four-cold-prompt gate (~188416 filler tokens per prompt before chat overhead, with 2 s arrival staggering) passed 4/4 retrieval checks and 3/4 word-limit checks. TTFT: 48.90, 97.32, 147.80, 197.69 s. These arrival/result timings do **not** demonstrate simultaneous four-way cold prefill; cold admission/prefill can serialize substantially.

### Compatibility and limits

Authenticated OpenAI-compatible Chat Completions and Responses, image input, JSON object/schema, and automatic/required/named tool-choice smoke checks passed. Stream cancellation after output and subsequent Chat/Responses/vision recovery passed. Bounded executable coding fixtures passed; these are not broad coding or reasoning-quality evaluations.

This supports controlled use of this particular configuration, not zero-latency cold prompts, guaranteed cache hits, unrestricted context growth, or perfect instruction following. Comparisons with other experiments changed model variant, quantization, KV format and RAM budgets, so no engine-only superiority claim is made.

No private prompts, hostnames, addresses, credentials, user paths or raw deployment logs are included.