| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Unit tests for ThinkingBudgetStateHolder batch index moves.""" |
| 4 | |
| 5 | import torch |
| 6 | |
| 7 | from tests.v1.sample.utils import create_mock_reasoning_config |
| 8 | from vllm.sampling_params import SamplingParams |
| 9 | from vllm.v1.sample.logits_processor.interface import ( |
| 10 | BatchUpdate, |
| 11 | MoveDirectionality, |
| 12 | ) |
| 13 | from vllm.v1.sample.thinking_budget_state import ThinkingBudgetStateHolder |
| 14 | |
| 15 | |
| 16 | def _make_holder() -> ThinkingBudgetStateHolder: |
| 17 | return ThinkingBudgetStateHolder( |
| 18 | create_mock_reasoning_config([151667], [151668]), |
| 19 | 8, |
| 20 | 0, |
| 21 | torch.device("cpu"), |
| 22 | False, |
| 23 | ) |
| 24 | |
| 25 | |
| 26 | def test_swap_budgeted_with_unbudgeted_clears_empty_side(): |
| 27 | """Asymmetric SWAP must not leave the empty index sharing state.""" |
| 28 | h = _make_holder() |
| 29 | h.sync_batch( |
| 30 | BatchUpdate( |
| 31 | batch_size=2, |
| 32 | removed=(), |
| 33 | added=[ |
| 34 | (0, SamplingParams(thinking_token_budget=5), None, []), |
| 35 | (1, SamplingParams(), None, []), |
| 36 | ], |
| 37 | moved=(), |
| 38 | ) |
| 39 | ) |
| 40 | assert list(h._state.keys()) == [0] |
| 41 | budget_state = h._state[0] |
| 42 | |
| 43 | h.sync_batch( |
| 44 | BatchUpdate( |
| 45 | batch_size=2, |
| 46 | removed=(), |
| 47 | added=(), |
| 48 | moved=[(0, 1, MoveDirectionality.SWAP)], |
| 49 | ) |
| 50 | ) |
| 51 | assert list(h._state.keys()) == [1] |
| 52 | assert h._state[1] is budget_state |
| 53 | assert h._state[1]["thinking_token_budget"] == 5 |
| 54 | |
| 55 | h.sync_batch( |
| 56 | BatchUpdate( |
| 57 | batch_size=2, |
| 58 | removed=(), |
| 59 | added=(), |
| 60 | moved=[(0, 1, MoveDirectionality.SWAP)], |
| 61 | ) |
| 62 | ) |
| 63 | assert list(h._state.keys()) == [0] |
| 64 | assert h._state[0] is budget_state |
| 65 | |
| 66 | |
| 67 | def test_swap_exchanges_two_budgeted_states(): |
| 68 | h = _make_holder() |
| 69 | h.sync_batch( |
| 70 | BatchUpdate( |
| 71 | batch_size=2, |
| 72 | removed=(), |
| 73 | added=[ |
| 74 | (0, SamplingParams(thinking_token_budget=3), None, []), |
| 75 | (1, SamplingParams(thinking_token_budget=7), None, []), |
| 76 | ], |
| 77 | moved=(), |
| 78 | ) |
| 79 | ) |
| 80 | b0 = h._state[0]["thinking_token_budget"] |
| 81 | b1 = h._state[1]["thinking_token_budget"] |
| 82 | h.sync_batch( |
| 83 | BatchUpdate( |
| 84 | batch_size=2, |
| 85 | removed=(), |
| 86 | added=(), |
| 87 | moved=[(0, 1, MoveDirectionality.SWAP)], |
| 88 | ) |
| 89 | ) |
| 90 | assert h._state[0]["thinking_token_budget"] == b1 |
| 91 | assert h._state[1]["thinking_token_budget"] == b0 |