| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import ast |
| 4 | import json |
| 5 | from dataclasses import dataclass, field |
| 6 | from functools import lru_cache |
| 7 | from typing import TYPE_CHECKING |
| 8 | |
| 9 | import torch |
| 10 | from transformers import PreTrainedTokenizerBase |
| 11 | |
| 12 | from vllm.exceptions import VLLMValidationError |
| 13 | from vllm.sampling_params import SamplingParams |
| 14 | from vllm.utils.import_utils import LazyLoader |
| 15 | from vllm.utils.torch_utils import PIN_MEMORY |
| 16 | from vllm.v1.structured_output.backend_types import ( |
| 17 | StructuredOutputBackend, |
| 18 | StructuredOutputGrammar, |
| 19 | StructuredOutputOptions, |
| 20 | ) |
| 21 | from vllm.v1.structured_output.utils import compile_regex_with_timeout |
| 22 | |
| 23 | if TYPE_CHECKING: |
| 24 | import lmformatenforcer |
| 25 | import lmformatenforcer.integrations.vllm as lmfe_vllm |
| 26 | else: |
| 27 | lmformatenforcer = LazyLoader("lmformatenforcer", globals(), "lmformatenforcer") |
| 28 | lmfe_vllm = LazyLoader( |
| 29 | "lmformatenforcer.integrations.vllm", |
| 30 | globals(), |
| 31 | "lmformatenforcer.integrations.vllm", |
| 32 | ) |
| 33 | |
| 34 | |
| 35 | @lru_cache |
| 36 | def _cached_build_vllm_token_enforcer_tokenizer_data( |
| 37 | tokenizer: PreTrainedTokenizerBase, vocab_size: int |
| 38 | ) -> "lmfe_vllm.TokenEnforcerTokenizerData": |
| 39 | return lmfe_vllm.build_vllm_token_enforcer_tokenizer_data( |
| 40 | tokenizer, use_bitmask=True, vocab_size=vocab_size |
| 41 | ) |
| 42 | |
| 43 | |
| 44 | @dataclass |
| 45 | class LMFormatEnforcerGrammar(StructuredOutputGrammar): |
| 46 | token_enforcer: lmformatenforcer.TokenEnforcer |
| 47 | current_tokens_prefix: list[int] = field(default_factory=list) |
| 48 | |
| 49 | def accept_tokens(self, request_id: str, tokens: list[int]) -> bool: |
| 50 | original_len = len(self.current_tokens_prefix) |
| 51 | for token in tokens: |
| 52 | if not self.token_enforcer.get_allowed_tokens( |
| 53 | self.current_tokens_prefix |
| 54 | ).is_token_allowed(token): |
| 55 | # Rollback partial updates to ensure atomicity. |
| 56 | del self.current_tokens_prefix[original_len:] |
| 57 | return False |
| 58 | self.current_tokens_prefix.append(token) |
| 59 | return True |
| 60 | |
| 61 | def validate_tokens(self, tokens: list[int]) -> list[int]: |
| 62 | for prefix_length in range(len(tokens)): |
| 63 | prefix = tokens[:prefix_length] |
| 64 | next_token = tokens[prefix_length] |
| 65 | if not self.token_enforcer.get_allowed_tokens( |
| 66 | self.current_tokens_prefix + prefix |
| 67 | ).is_token_allowed(next_token): |
| 68 | break |
| 69 | else: |
| 70 | return tokens |
| 71 | |
| 72 | return tokens[:prefix_length] |
| 73 | |
| 74 | def rollback(self, num_tokens: int) -> None: |
| 75 | self.current_tokens_prefix = self.current_tokens_prefix[:-num_tokens] |
| 76 | |
| 77 | def fill_bitmask(self, bitmask: torch.Tensor, batch_index: int) -> None: |
| 78 | allowed_tokens = self.token_enforcer.get_allowed_tokens( |
| 79 | self.current_tokens_prefix |
| 80 | ) |
| 81 | bitmask[batch_index] = allowed_tokens.allowed_tokens |
| 82 | |
| 83 | def is_terminated(self) -> bool: |
| 84 | # We are considered terminated if the prefix ends with eos_token_id |
| 85 | return_value = ( |
| 86 | len(self.current_tokens_prefix) > 0 |
| 87 | and self.current_tokens_prefix[-1] == self.token_enforcer.eos_token_id |
| 88 | ) |
| 89 | return return_value |
| 90 | |
| 91 | def reset(self): |
| 92 | self.current_tokens_prefix = [] |
| 93 | |
| 94 | |
| 95 | @dataclass |
| 96 | class LMFormatEnforcerBackend(StructuredOutputBackend): |
| 97 | def __post_init__(self): |
| 98 | self.tokenizer_data = _cached_build_vllm_token_enforcer_tokenizer_data( |
| 99 | self.tokenizer, self.vocab_size |
| 100 | ) |
| 101 | |
| 102 | def compile_grammar( |
| 103 | self, |
| 104 | request_type: StructuredOutputOptions, |
| 105 | grammar_spec: str, |
| 106 | stop_token_ids: set[int] | None = None, |
| 107 | ) -> StructuredOutputGrammar: |
| 108 | character_level_parser: lmformatenforcer.CharacterLevelParser |
| 109 | if request_type == StructuredOutputOptions.JSON: |
| 110 | spec_dict = json.loads(grammar_spec) |
| 111 | character_level_parser = lmformatenforcer.JsonSchemaParser(spec_dict) |
| 112 | elif request_type == StructuredOutputOptions.JSON_OBJECT: |
| 113 | character_level_parser = lmformatenforcer.JsonSchemaParser(None) |
| 114 | elif request_type == StructuredOutputOptions.REGEX: |
| 115 | character_level_parser = compile_regex_with_timeout( |
| 116 | lmformatenforcer.RegexParser, |
| 117 | grammar_spec, |
| 118 | ) |
| 119 | elif request_type == StructuredOutputOptions.CHOICE: |
| 120 | choices = ast.literal_eval(grammar_spec) |
| 121 | character_level_parser = lmformatenforcer.UnionParser( |
| 122 | [lmformatenforcer.StringParser(choice) for choice in choices] |
| 123 | ) |
| 124 | else: |
| 125 | raise ValueError( |
| 126 | f"Invalid request type for LM Format Enforcer backend({request_type!s})" |
| 127 | ) |
| 128 | max_rollback_tokens = ( |
| 129 | self.vllm_config.speculative_config.num_speculative_tokens |
| 130 | if self.vllm_config.speculative_config is not None |
| 131 | else 0 |
| 132 | ) |
| 133 | |
| 134 | if max_rollback_tokens > 0: |
| 135 | raise ValueError( |
| 136 | "LM Format Enforcer backend does not support speculative tokens" |
| 137 | ) |
| 138 | |
| 139 | token_enforcer = lmformatenforcer.TokenEnforcer( |
| 140 | tokenizer_data=self.tokenizer_data, |
| 141 | parser=character_level_parser, |
| 142 | ) |
| 143 | return LMFormatEnforcerGrammar(token_enforcer) |
| 144 | |
| 145 | def allocate_token_bitmask(self, max_num_seqs: int) -> torch.Tensor: |
| 146 | return torch.full( |
| 147 | (max_num_seqs, (self.vocab_size + 31) // 32), |
| 148 | -1, |
| 149 | dtype=torch.int32, |
| 150 | pin_memory=PIN_MEMORY, |
| 151 | ) |
| 152 | |
| 153 | def destroy(self): |
| 154 | pass |
| 155 | |
| 156 | |
| 157 | def validate_structured_output_request_lm_format_enforcer(params: SamplingParams): |
| 158 | if params.structured_outputs is None: |
| 159 | return |
| 160 | |
| 161 | so_params = params.structured_outputs |
| 162 | |
| 163 | if so_params.regex: |
| 164 | try: |
| 165 | compile_regex_with_timeout( |
| 166 | lmformatenforcer.RegexParser, |
| 167 | so_params.regex, |
| 168 | ) |
| 169 | except Exception as err: |
| 170 | raise VLLMValidationError( |
| 171 | f"Failed to compile regex for lm-format-enforcer: {err}" |
| 172 | ) from err |
| 173 | return |
| 174 | elif so_params.json: |
| 175 | if isinstance(so_params.json, str): |
| 176 | try: |
| 177 | # make sure schema is valid json |
| 178 | json.loads(so_params.json) |
| 179 | except json.JSONDecodeError as e: |
| 180 | raise VLLMValidationError("Invalid JSON grammar specification.") from e |
| 181 | else: |
| 182 | try: |
| 183 | json.dumps(so_params.json) |
| 184 | except Exception as e: |
| 185 | raise VLLMValidationError( |
| 186 | f"Error serializing structured outputs jsonschema: {e}" |
| 187 | ) from e |
| 188 | return |
| 189 | elif so_params.choice: |
| 190 | return |
| 191 | elif so_params.grammar: |
| 192 | raise VLLMValidationError( |
| 193 | "LM Format Enforcer structured outputs backend " |
| 194 | "does not support grammar specifications" |
| 195 | ) |