Adapter de benchmark isole (hors runtime Lea) ciblant un serveur
/v1/chat/completions a support vision (vLLM/SGLang/TGI), pour comparer
plus tard a Ollama via LeaBench. Ne controle jamais le desktop.
- core/evaluation/openai_compat_lea_bench_adapter.py : payload data-URL
image_url, parsing choices[0].message.content. Reutilise par import la
logique prompt/parse/normalisation de ollama_lea_bench_adapter (zero refactor).
- tools/lea_bench_openai_compat.py : wrapper CLI (--base-url defaut :8001).
- tests/unit/test_openai_compat_lea_bench_adapter.py : 6 tests mockes HTTP
(data URL, pas de fuite expectation/click_region, prediction valide,
abstain safe sur HTTP!=200 et reponse malformee, JSONL rechargeable).
Aucun runtime Lea modifie. Aucun service lance.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
0f122a512f