logs/output.json

{"model": "llama", "latency_ms": 12.5, "batch_size": 8}