Phase 7 of the vision-LLM benchmark plan. Adds:
- docs/benchmarks/2026-05-10-vision-llm.md — curated report (TL;DR,
per-model analysis, top-N agreement, cost vs cloud APIs, sample
captions). Verdict: qwen3vl-4b for the request path (3.55 s p50,
100% parse, decisive top-N distro); qwen3vl-8b for caption polish.
- docs/benchmarks/benchmark-2026-05-10-1424.json — raw 300-row dump
for diff-checking against future runs.
- main.tf: -fa -> -fa on (b9085 llama.cpp removed the no-value form
of the flash-attention flag; without the value llama-server exits
before serving any request).
- llama-cpp.md architecture doc links the report so future operators
land on the deployed-and-evaluated model from one entry point.
300/300 calls, 0 parse errors, 33m32s wall on a single T4 with the
GPU exclusively allocated. immich-ml was scaled to 0 for the run
(node1 RAM constraint, not GPU - bumping node1 RAM is tracked as a
follow-up).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Single Deployment of mostlygeek/llama-swap:cuda hot-swaps three
GGUF vision models (qwen3vl-8b, minicpm-v-4-5, qwen3vl-4b) at one
OpenAI-compat /v1 endpoint on Service llama-swap.llama-cpp.svc.
Idle TTL 10min so models unload between benchmark batches.
Storage: NFS-RWX from /srv/nfs-ssd/llamacpp (30Gi). One-shot
download Job pulls Q4_K_M GGUF + mmproj per model, creates stable
model.gguf / mmproj.gguf symlinks so the llama-swap config is
filename-agnostic, then warms the kernel page cache.
GPU: nvidia.com/gpu=1 = whole T4 — operator must scale immich-ml
to 0 during benchmark windows. wait_for_rollout=false so apply
doesn't block on GPU availability.
Initial use case: vision-LLM benchmark for instagram-poster
candidate scoring; future consumers (HA, agentic tooling) hit
the same endpoint via LiteLLM at the gateway.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
