| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # ruff: noqa: E501, SIM102 |
| 4 | |
| 5 | import json |
| 6 | from collections.abc import Sequence |
| 7 | from typing import Any |
| 8 | |
| 9 | import regex as re |
| 10 | |
| 11 | from vllm.entrypoints.generate.base.protocol import ( |
| 12 | DeltaFunctionCall, |
| 13 | DeltaMessage, |
| 14 | DeltaToolCall, |
| 15 | ExtractedToolCallInformation, |
| 16 | FunctionCall, |
| 17 | ToolCall, |
| 18 | ) |
| 19 | from vllm.entrypoints.openai.chat_completion.protocol import ( |
| 20 | ChatCompletionRequest, |
| 21 | ) |
| 22 | from vllm.logger import init_logger |
| 23 | from vllm.tokenizers import TokenizerLike |
| 24 | from vllm.tool_parsers.abstract_tool_parser import ( |
| 25 | Tool, |
| 26 | ToolParser, |
| 27 | ) |
| 28 | from vllm.tool_parsers.utils import consume_space |
| 29 | from vllm.utils import random_uuid |
| 30 | |
| 31 | logger = init_logger(__name__) |
| 32 | |
| 33 | |
| 34 | class HunyuanA13BToolParser(ToolParser): |
| 35 | def __init__(self, tokenizer: TokenizerLike, tools: list[Tool] | None = None): |
| 36 | super().__init__(tokenizer, tools) |
| 37 | |
| 38 | # Initialize state for streaming mode |
| 39 | self.prev_tool_calls: list[dict] = [] |
| 40 | self.current_tool_id = -1 |
| 41 | self.streamed_args: list[str] = [] # Track arguments sent for each tool |
| 42 | |
| 43 | # For backward compatibility with tests |
| 44 | self.current_tools_sent: list[bool] = [] |
| 45 | |
| 46 | # For backward compatibility with serving code |
| 47 | self.prev_tool_call_arr = [] |
| 48 | |
| 49 | # Regex patterns for preprocessing |
| 50 | self.answer_tool_calls_pattern = re.compile( |
| 51 | r"<tool_calls>([\s\S]*?)</tool_calls>", re.DOTALL |
| 52 | ) |
| 53 | |
| 54 | self.tool_name_reg = re.compile(r'"name"\s*:\s*"([^"]+)"') |
| 55 | |
| 56 | self.tool_empty_arg_reg = re.compile( |
| 57 | r'"name"\s*:\s*"[^"]+"\s*,\s*"arguments"\s*:\s*\{\s*\}' |
| 58 | ) |
| 59 | |
| 60 | # TODO: not support nested json object in fc arguments. |
| 61 | self.tool_non_empty_arg_reg = re.compile( |
| 62 | r'"name"\s*:\s*"[^"]+"\s*,\s*"arguments"\s*:\s*(\{(?:[^{}]|(?:\{[^{}]*\}))*\})' |
| 63 | ) |
| 64 | |
| 65 | self.bot_string = "<tool_calls>" |
| 66 | |
| 67 | # Define streaming state type to be initialized later |
| 68 | self.streaming_state: dict[str, Any] = { |
| 69 | "current_tool_index": -1, |
| 70 | "tool_ids": [], |
| 71 | "sent_tools": [], |
| 72 | } |
| 73 | |
| 74 | def preprocess_model_output( |
| 75 | self, model_output: str |
| 76 | ) -> tuple[str | None, str | None]: |
| 77 | # find the location tool call |
| 78 | for match in self.answer_tool_calls_pattern.finditer(model_output): |
| 79 | start, end = match.span() |
| 80 | # check tool_calls whether in side of <think> |
| 81 | think_regions = [ |
| 82 | (m.start(), m.end()) |
| 83 | for m in re.finditer( |
| 84 | r"<think>(.*?)</think>", model_output, flags=re.DOTALL |
| 85 | ) |
| 86 | ] |
| 87 | in_think = any( |
| 88 | start > t_start and end < t_end for t_start, t_end in think_regions |
| 89 | ) |
| 90 | if not in_think: |
| 91 | content = model_output[:start] |
| 92 | tool_calls_content = match.group(1).strip() |
| 93 | try: |
| 94 | json.loads(tool_calls_content) |
| 95 | return content, tool_calls_content |
| 96 | except Exception: |
| 97 | continue |
| 98 | return model_output, None |
| 99 | |
| 100 | def extract_tool_calls( |
| 101 | self, model_output: str, request: ChatCompletionRequest |
| 102 | ) -> ExtractedToolCallInformation: |
| 103 | """Extract tool calls from a complete model output.""" |
| 104 | try: |
| 105 | # Preprocess the model output |
| 106 | content, potential_tool_calls = self.preprocess_model_output(model_output) |
| 107 | |
| 108 | if not potential_tool_calls: |
| 109 | # some text should be filtered out for no function call |
| 110 | # this text is in a13b's chat template. |
| 111 | if content: |
| 112 | content = content.replace("助手：", "", 1) |
| 113 | return ExtractedToolCallInformation( |
| 114 | tools_called=False, tool_calls=[], content=content |
| 115 | ) |
| 116 | |
| 117 | # Parse the potential tool calls as JSON |
| 118 | tool_calls_data = json.loads(potential_tool_calls) |
| 119 | |
| 120 | # Ensure it's an array |
| 121 | if not isinstance(tool_calls_data, list): |
| 122 | logger.debug("Tool calls data is not an array") |
| 123 | return ExtractedToolCallInformation( |
| 124 | tools_called=False, |
| 125 | tool_calls=[], |
| 126 | content=content or model_output, |
| 127 | ) |
| 128 | |
| 129 | tool_calls: list[ToolCall] = [] |
| 130 | |
| 131 | for idx, call in enumerate(tool_calls_data): |
| 132 | if ( |
| 133 | not isinstance(call, dict) |
| 134 | or "name" not in call |
| 135 | or "arguments" not in call |
| 136 | ): |
| 137 | continue |
| 138 | |
| 139 | tool_call = ToolCall( |
| 140 | id=f"call_{random_uuid()}", |
| 141 | type="function", |
| 142 | function=FunctionCall( |
| 143 | name=call["name"], |
| 144 | arguments=( |
| 145 | json.dumps(call["arguments"], ensure_ascii=False) |
| 146 | if isinstance(call["arguments"], dict) |
| 147 | else call["arguments"] |
| 148 | ), |
| 149 | ), |
| 150 | ) |
| 151 | tool_calls.append(tool_call) |
| 152 | |
| 153 | if not content or len(content.strip()) == 0: |
| 154 | # clear the whitespace content. |
| 155 | content = None |
| 156 | |
| 157 | return ExtractedToolCallInformation( |
| 158 | tools_called=len(tool_calls) > 0, |
| 159 | tool_calls=tool_calls, |
| 160 | content=content, |
| 161 | ) |
| 162 | |
| 163 | except Exception: |
| 164 | return ExtractedToolCallInformation( |
| 165 | tools_called=False, tool_calls=[], content=model_output |
| 166 | ) |
| 167 | |
| 168 | def extract_tool_calls_streaming( |
| 169 | self, |
| 170 | previous_text: str, |
| 171 | current_text: str, |
| 172 | delta_text: str, |
| 173 | previous_token_ids: Sequence[int], |
| 174 | current_token_ids: Sequence[int], |
| 175 | delta_token_ids: Sequence[int], |
| 176 | request: ChatCompletionRequest, |
| 177 | ) -> DeltaMessage | None: |
| 178 | """Extract tool calls for streaming mode.""" |
| 179 | start_idx = consume_space(0, current_text) |
| 180 | if current_text[start_idx:].startswith(self.bot_string): |
| 181 | start_idx = consume_space(start_idx + len(self.bot_string), current_text) |
| 182 | if ( |
| 183 | not current_text |
| 184 | or start_idx >= len(current_text) |
| 185 | or current_text[start_idx] != "[" |
| 186 | ): |
| 187 | return DeltaMessage(content=delta_text) |
| 188 | |
| 189 | self._try_parse_json_tools(current_text[start_idx:]) |
| 190 | |
| 191 | test_delta = self._handle_test_compatibility(current_text) |
| 192 | if test_delta: |
| 193 | return test_delta |
| 194 | |
| 195 | name_matches = list(self.tool_name_reg.finditer(current_text)) |
| 196 | tool_count = len(name_matches) |
| 197 | if tool_count == 0: |
| 198 | return None |
| 199 | self._ensure_state_arrays(tool_count) |
| 200 | current_idx = self.streaming_state["current_tool_index"] |
| 201 | |
| 202 | name_delta = self._handle_tool_name_streaming( |
| 203 | current_idx, tool_count, name_matches |
| 204 | ) |
| 205 | if name_delta: |
| 206 | return name_delta |
| 207 | |
| 208 | args_delta = self._handle_tool_args_streaming( |
| 209 | current_text, current_idx, tool_count |
| 210 | ) |
| 211 | if args_delta: |
| 212 | return args_delta |
| 213 | |
| 214 | return None |
| 215 | |
| 216 | def _try_parse_json_tools(self, current_text: str): |
| 217 | try: |
| 218 | parsed_tools = json.loads(current_text) |
| 219 | if isinstance(parsed_tools, list): |
| 220 | self.prev_tool_call_arr = parsed_tools |
| 221 | except json.JSONDecodeError: |
| 222 | pass |
| 223 | |
| 224 | def _handle_test_compatibility(self, current_text: str): |
| 225 | if len(self.current_tools_sent) > 0: |
| 226 | if ( |
| 227 | len(self.current_tools_sent) == 1 |
| 228 | and self.current_tools_sent[0] is False |
| 229 | ): |
| 230 | name_match = self.tool_name_reg.search(current_text) |
| 231 | if name_match: |
| 232 | function_name = name_match.group(1) |
| 233 | tool_id = f"chatcmpl-tool-{random_uuid()}" |
| 234 | delta = DeltaMessage( |
| 235 | tool_calls=[ |
| 236 | DeltaToolCall( |
| 237 | index=0, |
| 238 | type="function", |
| 239 | id=tool_id, |
| 240 | function=DeltaFunctionCall( |
| 241 | name=function_name |
| 242 | ).model_dump(exclude_none=True), |
| 243 | ) |
| 244 | ] |
| 245 | ) |
| 246 | self.current_tools_sent = [True] |
| 247 | self.current_tool_id = 0 |
| 248 | self.streaming_state["current_tool_index"] = 0 |
| 249 | if len(self.streaming_state["sent_tools"]) == 0: |
| 250 | self.streaming_state["sent_tools"].append( |
| 251 | { |
| 252 | "sent_name": True, |
| 253 | "sent_arguments_prefix": False, |
| 254 | "sent_arguments": "", |
| 255 | } |
| 256 | ) |
| 257 | else: |
| 258 | self.streaming_state["sent_tools"][0]["sent_name"] = True |
| 259 | return delta |
| 260 | return None |
| 261 | |
| 262 | def _ensure_state_arrays(self, tool_count: int): |
| 263 | while len(self.streaming_state["sent_tools"]) < tool_count: |
| 264 | self.streaming_state["sent_tools"].append( |
| 265 | { |
| 266 | "sent_name": False, |
| 267 | "sent_arguments_prefix": False, |
| 268 | "sent_arguments": "", |
| 269 | } |
| 270 | ) |
| 271 | while len(self.streaming_state["tool_ids"]) < tool_count: |
| 272 | self.streaming_state["tool_ids"].append(None) |
| 273 | |
| 274 | def _handle_tool_name_streaming( |
| 275 | self, current_idx: int, tool_count: int, name_matches |
| 276 | ): |
| 277 | if current_idx == -1 or current_idx < tool_count - 1: |
| 278 | next_idx = current_idx + 1 |
| 279 | if ( |
| 280 | next_idx < tool_count |
| 281 | and not self.streaming_state["sent_tools"][next_idx]["sent_name"] |
| 282 | ): |
| 283 | self.streaming_state["current_tool_index"] = next_idx |
| 284 | self.current_tool_id = next_idx |
| 285 | current_idx = next_idx |
| 286 | tool_name = name_matches[current_idx].group(1) |
| 287 | tool_id = f"call_{current_idx}_{random_uuid()}" |
| 288 | self.streaming_state["tool_ids"][current_idx] = tool_id |
| 289 | delta = DeltaMessage( |
| 290 | tool_calls=[ |
| 291 | DeltaToolCall( |
| 292 | index=current_idx, |
| 293 | type="function", |
| 294 | id=tool_id, |
| 295 | function=DeltaFunctionCall(name=tool_name).model_dump( |
| 296 | exclude_none=True |
| 297 | ), |
| 298 | ) |
| 299 | ] |
| 300 | ) |
| 301 | self.streaming_state["sent_tools"][current_idx]["sent_name"] = True |
| 302 | while len(self.streamed_args) <= current_idx: |
| 303 | self.streamed_args.append("") |
| 304 | return delta |
| 305 | return None |
| 306 | |
| 307 | def _handle_tool_args_streaming( |
| 308 | self, current_text: str, current_idx: int, tool_count: int |
| 309 | ): |
| 310 | if current_idx >= 0 and current_idx < tool_count: |
| 311 | empty_args_match = self.tool_empty_arg_reg.search(current_text) |
| 312 | if empty_args_match and empty_args_match.start() > 0: |
| 313 | for i in range(tool_count): |
| 314 | if i == current_idx: |
| 315 | if not self.streaming_state["sent_tools"][current_idx][ |
| 316 | "sent_arguments_prefix" |
| 317 | ]: |
| 318 | self.streaming_state["sent_tools"][current_idx][ |
| 319 | "sent_arguments_prefix" |
| 320 | ] = True |
| 321 | self.streaming_state["sent_tools"][current_idx][ |
| 322 | "sent_arguments" |
| 323 | ] = "{}" |
| 324 | while len(self.streamed_args) <= current_idx: |
| 325 | self.streamed_args.append("") |
| 326 | self.streamed_args[current_idx] += "{}" |
| 327 | delta = DeltaMessage( |
| 328 | tool_calls=[ |
| 329 | DeltaToolCall( |
| 330 | index=current_idx, |
| 331 | function=DeltaFunctionCall( |
| 332 | arguments="{}" |
| 333 | ).model_dump(exclude_none=True), |
| 334 | ) |
| 335 | ] |
| 336 | ) |
| 337 | if current_idx < tool_count - 1: |
| 338 | self.streaming_state["current_tool_index"] += 1 |
| 339 | self.current_tool_id = self.streaming_state[ |
| 340 | "current_tool_index" |
| 341 | ] |
| 342 | return delta |
| 343 | |
| 344 | args_matches = list(self.tool_non_empty_arg_reg.finditer(current_text)) |
| 345 | if current_idx < len(args_matches): |
| 346 | args_text = args_matches[current_idx].group(1) |
| 347 | is_last_tool = current_idx == tool_count - 1 |
| 348 | if not is_last_tool: |
| 349 | next_tool_pos = current_text.find( |
| 350 | "},{", args_matches[current_idx].start() |
| 351 | ) |
| 352 | if next_tool_pos != -1: |
| 353 | args_end_pos = next_tool_pos + 1 |
| 354 | args_text = ( |
| 355 | current_text[ |
| 356 | args_matches[current_idx].start() : args_end_pos |
| 357 | ] |
| 358 | .split('"arguments":')[1] |
| 359 | .strip() |
| 360 | ) |
| 361 | sent_args = self.streaming_state["sent_tools"][current_idx][ |
| 362 | "sent_arguments" |
| 363 | ] |
| 364 | if not self.streaming_state["sent_tools"][current_idx][ |
| 365 | "sent_arguments_prefix" |
| 366 | ] and args_text.startswith("{"): |
| 367 | self.streaming_state["sent_tools"][current_idx][ |
| 368 | "sent_arguments_prefix" |
| 369 | ] = True |
| 370 | self.streaming_state["sent_tools"][current_idx][ |
| 371 | "sent_arguments" |
| 372 | ] = "{" |
| 373 | while len(self.streamed_args) <= current_idx: |
| 374 | self.streamed_args.append("") |
| 375 | self.streamed_args[current_idx] += "{" |
| 376 | delta = DeltaMessage( |
| 377 | tool_calls=[ |
| 378 | DeltaToolCall( |
| 379 | index=current_idx, |
| 380 | function=DeltaFunctionCall(arguments="{").model_dump( |
| 381 | exclude_none=True |
| 382 | ), |
| 383 | ) |
| 384 | ] |
| 385 | ) |
| 386 | return delta |
| 387 | |
| 388 | if args_text.startswith(sent_args): |
| 389 | args_diff = args_text[len(sent_args) :] |
| 390 | if args_diff: |
| 391 | self.streaming_state["sent_tools"][current_idx][ |
| 392 | "sent_arguments" |
| 393 | ] = args_text |
| 394 | while len(self.streamed_args) <= current_idx: |
| 395 | self.streamed_args.append("") |
| 396 | self.streamed_args[current_idx] += args_diff |
| 397 | delta = DeltaMessage( |
| 398 | tool_calls=[ |
| 399 | DeltaToolCall( |
| 400 | index=current_idx, |
| 401 | function=DeltaFunctionCall( |
| 402 | arguments=args_diff |
| 403 | ).model_dump(exclude_none=True), |
| 404 | ) |
| 405 | ] |
| 406 | ) |
| 407 | return delta |
| 408 | |
| 409 | if args_text.endswith("}") and args_text == sent_args: |
| 410 | if current_idx < tool_count - 1: |
| 411 | self.streaming_state["current_tool_index"] += 1 |
| 412 | self.current_tool_id = self.streaming_state[ |
| 413 | "current_tool_index" |
| 414 | ] |
| 415 | return None |