{
  "purpose": "Contexte des mesures de référence Qwen3.8 sur RTX 3090, 24 octobre 2026",
  "os_pretty_name": "Debian GNU/Linux forky/sid",
  "cpu_model": "Intel(R) Xeon(R) W-1270 CPU @ 3.40GHz",
  "cpu_logical_count": 16,
  "memory_total_kib": 65670800,
  "ollama_environment_explicit": {
    "OLLAMA_CONTEXT_LENGTH": "65536"
  },
  "client_isolation": {
    "service": "hermes-gateway.service",
    "state_before_benchmark": "frozen",
    "note": "Service gelé temporairement; la requête déjà envoyée à Ollama doit finir avant les essais.",
    "state_after_benchmark": "running"
  },
  "gpu_unlock_baseline": "module absent; aucun paramètre gpu_unlock dans la ligne de démarrage",
  "model_update_policy": "aucun pull ni changement de poids entre les essais",
  "runner_options_verified_in_ollama_journal": {
    "timestamp_local": "2026-10-24T14:21:18+02:00",
    "spec_type": "draft-mtp",
    "spec_draft_n_max": 4,
    "parallel_slots": 1,
    "context": 16384,
    "batch": 512,
    "ubatch": 512,
    "flash_attention": "auto"
  },
  "measurement_window_utc": {
    "start": "2026-10-24T12:21:14.385714+00:00",
    "end": "2026-10-24T12:24:20.896836+00:00"
  }
}