Skip to content

vllm.v1.worker.gpu.sample.thinking_budget

Classes:

ThinkingBudgetState

Model Runner V2 state for per-request thinking token budgets.

Source code in vllm/v1/worker/gpu/sample/thinking_budget.py 
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
class ThinkingBudgetState:
    """Model Runner V2 state for per-request thinking token budgets."""

    def __init__(
        self,
        req_states: RequestState,
        reasoning_config: "ReasoningConfig | None",
    ):
        self.req_states = req_states
        self.max_num_reqs = req_states.max_num_reqs
        self.device = req_states.device

        start_ids = (
            []
            if reasoning_config is None
            else reasoning_config.reasoning_start_token_ids or []
        )
        end_ids = (
            []
            if reasoning_config is None
            else reasoning_config.reasoning_end_token_ids or []
        )
        natural_end_ids = (
            []
            if reasoning_config is None
            else reasoning_config.natural_reasoning_end_token_ids or []
        )
        self.enabled = bool(start_ids and end_ids and natural_end_ids)
        if not self.enabled:
            return

        self.thinking_token_budget = UvaBackedTensor(
            self.max_num_reqs, dtype=torch.int32
        )
        self.thinking_token_budget.np.fill(-1)
        self.thinking_token_budget.copy_to_uva()
        self.use_thinking_budget = np.zeros(self.max_num_reqs, dtype=bool)

        self.cached_last_start = torch.full(
            (self.max_num_reqs,), -1, dtype=torch.int32, device=self.device
        )
        self.cached_last_end = torch.full(
            (self.max_num_reqs,), -1, dtype=torch.int32, device=self.device
        )
        self.cached_scan_pos = torch.zeros(
            self.max_num_reqs, dtype=torch.int32, device=self.device
        )
        self._reset_reqs: list[int] = []
        self._budget_dirty = False

        self.reasoning_start_token_ids = torch.tensor(
            start_ids, dtype=torch.int32, device=self.device
        )
        self.natural_reasoning_end_token_ids = torch.tensor(
            natural_end_ids, dtype=torch.int32, device=self.device
        )
        self.reasoning_end_token_ids = torch.tensor(
            end_ids, dtype=torch.int32, device=self.device
        )

    def add_request(self, req_idx: int, sampling_params: SamplingParams) -> None:
        if not self.enabled:
            return
        budget = sampling_params.thinking_token_budget
        self.use_thinking_budget[req_idx] = budget is not None
        if budget is None:
            budget = -1
        else:
            budget = min(budget, _INT32_MAX)
            self._reset_reqs.append(req_idx)
        if self.thinking_token_budget.np[req_idx] != budget:
            self.thinking_token_budget.np[req_idx] = budget
            self._budget_dirty = True

    def apply_staged_writes(self) -> None:
        if not self.enabled:
            return
        if self._reset_reqs:
            idx = async_tensor_h2d(
                self._reset_reqs, dtype=torch.int64, device=self.device
            )
            self.cached_last_start.index_fill_(0, idx, -1)
            self.cached_last_end.index_fill_(0, idx, -1)
            self.cached_scan_pos.index_fill_(0, idx, 0)
            self._reset_reqs.clear()
        if self._budget_dirty:
            self.thinking_token_budget.copy_to_uva()
            self._budget_dirty = False

    def apply(
        self,
        logits: torch.Tensor,
        expanded_idx_mapping: torch.Tensor,
        idx_mapping: torch.Tensor,
        idx_mapping_np: np.ndarray,
        input_ids: torch.Tensor,
        expanded_local_pos: torch.Tensor,
    ) -> None:
        if not self.enabled or not np.any(self.use_thinking_budget[idx_mapping_np]):
            return

        apply_thinking_budget(
            logits,
            idx_mapping,
            expanded_idx_mapping,
            self.thinking_token_budget.gpu,
            self.req_states.all_token_ids.gpu,
            self.req_states.total_len.gpu,
            input_ids,
            expanded_local_pos,
            self.cached_last_start,
            self.cached_last_end,
            self.cached_scan_pos,
            self.reasoning_start_token_ids,
            self.natural_reasoning_end_token_ids,
            self.reasoning_end_token_ids,
        )