from inferscale import run_simulation


def cfg(cache: bool):
    return {
        "model": "Qwen2.5-3B",
        "accelerator": "L4",
        "quantization": "int8",
        "scheduler": "continuous_fcfs",
        "duration_s": 6,
        "request_rate_rps": 2,
        "prompt_tokens_mean": 512,
        "prompt_tokens_cv": 0,
        "output_tokens_mean": 8,
        "output_tokens_cv": 0,
        "seed": 2,
        "prefix_cache_enabled": cache,
        "shared_prefix_tokens": 256,
        "prefix_reuse_fraction": 1.0,
    }


def test_prefix_cache_reduces_prefill_work_and_is_reported():
    uncached = run_simulation(cfg(False))
    cached = run_simulation(cfg(True))
    assert cached["resource"]["prefix_cache_hit_rate"] == 1.0
    assert cached["resource"]["prefill_tokens_saved"] > 0
    assert cached["latency"]["ttft_ms"]["p50"] < uncached["latency"]["ttft_ms"]["p50"]