| """Conserved work and honest cost-model accounting, not a speedup acceptance test.""" |
| from __future__ import annotations |
|
|
| import pytest |
|
|
| from offload.bg_prefill.simulate import Cost, simulate |
|
|
|
|
| @pytest.mark.parametrize('backgrounds', [1, 2]) |
| @pytest.mark.parametrize('priority', [0, 10]) |
| def test_equal_full_work_and_reported_fairness_cost(backgrounds: int, priority: int) -> None: |
| control = simulate(backgrounds, priority, False) |
| candidate = simulate(backgrounds, priority, True) |
| for row in (control, candidate): |
| assert row['background_prompt_tokens_total'] == backgrounds * 60000 |
| assert row['estimated_foreground_tokens'] == 600 |
| |
| assert row['all_work_s'] == pytest.approx(300 * .039 + backgrounds * 60000 / 2200) |
| assert candidate['background_tokens_at_foreground_end'] == backgrounds * 12 * 32 |
| assert candidate['max_background_service_gap_calls_during_overlap'] == 24 |
|
|
|
|
| def test_simulation_refuses_invalid_costs_and_workloads() -> None: |
| with pytest.raises(ValueError): |
| Cost(decode_ms=0) |
| with pytest.raises(ValueError): |
| simulate(3, 0, True) |
|
|
