• himmetozcan's avatar
    Sadece GPU sunucusundaki GLiNER2 servisi · 0bea2e43
    himmetozcan authored
    Repo icerigi 10.10.23.10:/home/srv/Projects/Data_Privacy/GLINER2 ile birebir ayni.
    
    - FastAPI servisi, NVIDIA L4 (GPU 0) uzerinde fp16 cikarim, 4 uvicorn worker
    - Dinamik micro-batching (max_batch=32, 8ms toplama penceresi)
    - Sinirli kuyruk: dolu ise HTTP 429 (sessiz fail-open yerine acik red)
    - Siniflar: national_id_number (TC), iban, person (isim/soyisim), api_key
    - TC/IBAN icin checksum tabanli deterministik katman
      (resmi TC algoritmasi + IBAN MOD-97) + model yanlis pozitif vetosu
    - LiteLLM generic_guardrail_api sozlesmesi (/beta/litellm_basic_guardrail_api)
    - Prometheus metrikleri, /health, /detect, /redact, /batch
    - Docker Compose: GPU 0'a sabitli, restart unless-stopped (rebootta otomatik kalkar)
    - Sinif politikasi: GLINER2_BLOCK_LABELS (su an bos = sadece maskeleme)
    - Test: scripts/variant_test.py (19/19), scripts/bench.py (yuk testi)
    - Olculen: ~380 req/s sunucu ici, ~290 req/s ag uzerinden
    0bea2e43