| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import numpy as np |
| 4 | import torch |
| 5 | |
| 6 | from vllm.sampling_params import SamplingParams |
| 7 | from vllm.v1.sample.ops.topk_topp_sampler import apply_top_k_top_p |
| 8 | from vllm.v1.worker.gpu.buffer_utils import UvaBackedTensor |
| 9 | from vllm.v1.worker.gpu.sample.gumbel import apply_temperature |
| 10 | from vllm.v1.worker.gpu.sample.min_p import apply_min_p |
| 11 | |
| 12 | NO_LOGPROBS = -1 |
| 13 | _NP_INT64_MIN = np.iinfo(np.int64).min |
| 14 | _NP_INT64_MAX = np.iinfo(np.int64).max |
| 15 | |
| 16 | |
| 17 | class SamplingStates: |
| 18 | def __init__(self, max_num_reqs: int, vocab_size: int): |
| 19 | self.max_num_reqs = max_num_reqs |
| 20 | self.vocab_size = vocab_size |
| 21 | |
| 22 | self.temperature = UvaBackedTensor(max_num_reqs, dtype=torch.float32) |
| 23 | self.top_k = UvaBackedTensor(max_num_reqs, dtype=torch.int32) |
| 24 | self.top_p = UvaBackedTensor(max_num_reqs, dtype=torch.float32) |
| 25 | self.min_p = UvaBackedTensor(max_num_reqs, dtype=torch.float32) |
| 26 | self.seeds = UvaBackedTensor(max_num_reqs, dtype=torch.int64) |
| 27 | # Tracks whether `seed` was set explicitly by the user, so callers |
| 28 | # can fall back from RNG paths that don't honor per-request seeds. |
| 29 | self.seeds_set = np.zeros(max_num_reqs, dtype=bool) |
| 30 | |
| 31 | # Initialize top_k and top_p manually because 0 is an invalid value for them. |
| 32 | self.top_k.np.fill(self.vocab_size) |
| 33 | self.top_k.copy_to_uva() |
| 34 | self.top_p.np.fill(1.0) |
| 35 | self.top_p.copy_to_uva() |
| 36 | |
| 37 | self.num_logprobs = np.empty(self.max_num_reqs, dtype=np.int32) |
| 38 | # -1 means no logprobs are requested. |
| 39 | self.num_logprobs.fill(NO_LOGPROBS) |
| 40 | |
| 41 | def add_request(self, req_idx: int, sampling_params: SamplingParams) -> bool: |
| 42 | temperature = sampling_params.temperature |
| 43 | top_k = sampling_params.top_k |
| 44 | top_p = sampling_params.top_p |
| 45 | min_p = sampling_params.min_p |
| 46 | |
| 47 | self.temperature.np[req_idx] = temperature |
| 48 | self.top_p.np[req_idx] = top_p |
| 49 | if top_k <= 0 or top_k > self.vocab_size: |
| 50 | top_k = self.vocab_size |
| 51 | self.top_k.np[req_idx] = top_k |
| 52 | self.min_p.np[req_idx] = min_p |
| 53 | |
| 54 | seed = sampling_params.seed |
| 55 | self.seeds_set[req_idx] = seed is not None |
| 56 | if seed is None: |
| 57 | seed = np.random.randint(_NP_INT64_MIN, _NP_INT64_MAX) |
| 58 | self.seeds.np[req_idx] = seed |
| 59 | |
| 60 | num_logprobs = sampling_params.logprobs |
| 61 | if num_logprobs is None: |
| 62 | num_logprobs = NO_LOGPROBS |
| 63 | elif num_logprobs == -1: |
| 64 | num_logprobs = self.vocab_size |
| 65 | self.num_logprobs[req_idx] = num_logprobs |
| 66 | |
| 67 | return temperature != 0.0 and ( |
| 68 | temperature != 1.0 |
| 69 | or min_p != 0.0 |
| 70 | or top_k != self.vocab_size |
| 71 | or top_p != 1.0 |
| 72 | ) |
| 73 | |
| 74 | def apply_staged_writes(self) -> None: |
| 75 | self.temperature.copy_to_uva() |
| 76 | self.top_p.copy_to_uva() |
| 77 | self.top_k.copy_to_uva() |
| 78 | self.min_p.copy_to_uva() |
| 79 | self.seeds.copy_to_uva() |
| 80 | |
| 81 | def apply_temperature( |
| 82 | self, |
| 83 | logits: torch.Tensor, |
| 84 | expanded_idx_mapping: torch.Tensor, |
| 85 | idx_mapping_np: np.ndarray, |
| 86 | ) -> None: |
| 87 | temp_np = self.temperature.np[idx_mapping_np] |
| 88 | if np.all((temp_np == 0.0) | (temp_np == 1.0)): |
| 89 | # No request requires temperature. Skip the kernel launch. |
| 90 | return |
| 91 | |
| 92 | apply_temperature(logits, expanded_idx_mapping, self.temperature.gpu) |
| 93 | |
| 94 | def apply_min_p( |
| 95 | self, |
| 96 | logits: torch.Tensor, |
| 97 | expanded_idx_mapping: torch.Tensor, |
| 98 | idx_mapping_np: np.ndarray, |
| 99 | ) -> None: |
| 100 | if np.all(self.min_p.np[idx_mapping_np] == 0.0): |
| 101 | # No request uses min_p. Skip the kernel launch. |
| 102 | return |
| 103 | apply_min_p(logits, expanded_idx_mapping, self.min_p.gpu) |
| 104 | |
| 105 | def get_top_k_top_p( |
| 106 | self, expanded_idx_mapping: torch.Tensor, idx_mapping_np: np.ndarray |
| 107 | ) -> tuple[torch.Tensor | None, torch.Tensor | None]: |
| 108 | do_top_k = np.any(self.top_k.np[idx_mapping_np] != self.vocab_size) |
| 109 | do_top_p = np.any(self.top_p.np[idx_mapping_np] != 1.0) |
| 110 | top_k = self.top_k.gpu[expanded_idx_mapping] if do_top_k else None |
| 111 | top_p = self.top_p.gpu[expanded_idx_mapping] if do_top_p else None |
| 112 | return top_k, top_p |
| 113 | |
| 114 | def apply_top_k_top_p( |
| 115 | self, |
| 116 | logits: torch.Tensor, |
| 117 | expanded_idx_mapping: torch.Tensor, |
| 118 | idx_mapping_np: np.ndarray, |
| 119 | ) -> torch.Tensor: |
| 120 | top_k, top_p = self.get_top_k_top_p(expanded_idx_mapping, idx_mapping_np) |
| 121 | if top_k is None and top_p is None: |
| 122 | return logits |
| 123 | return apply_top_k_top_p(logits, top_k, top_p) |
| 124 | |
| 125 | def any_greedy(self, idx_mapping_np: np.ndarray) -> bool: |
| 126 | return bool(np.any(self.temperature.np[idx_mapping_np] == 0.0)) |
| 127 | |
| 128 | def any_explicit_seed(self, idx_mapping_np: np.ndarray) -> bool: |
| 129 | return bool(np.any(self.seeds_set[idx_mapping_np])) |
| 130 | |
| 131 | def max_num_logprobs(self, idx_mapping_np: np.ndarray) -> int: |
| 132 | return int(np.max(self.num_logprobs[idx_mapping_np])) |