| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from __future__ import annotations |
| 5 | |
| 6 | import functools |
| 7 | import json |
| 8 | from dataclasses import replace |
| 9 | from itertools import groupby |
| 10 | from typing import TYPE_CHECKING |
| 11 | |
| 12 | from vllm.entrypoints.generate.base.protocol import ( |
| 13 | DeltaFunctionCall, |
| 14 | DeltaMessage, |
| 15 | DeltaToolCall, |
| 16 | ) |
| 17 | from vllm.parser.engine.events import EventType, SemanticEvent |
| 18 | from vllm.parser.engine.parser_engine import ParserEngine |
| 19 | from vllm.parser.engine.parser_engine_config import ( |
| 20 | ParserEngineConfig, |
| 21 | ParserState, |
| 22 | Transition, |
| 23 | ) |
| 24 | |
| 25 | if TYPE_CHECKING: |
| 26 | from vllm.entrypoints.openai.chat_completion.protocol import ( |
| 27 | ChatCompletionRequest, |
| 28 | ) |
| 29 | from vllm.entrypoints.openai.responses.protocol import ResponsesRequest |
| 30 | from vllm.tokenizers import TokenizerLike |
| 31 | from vllm.tool_parsers.abstract_tool_parser import Tool |
| 32 | |
| 33 | BEGIN_TOOL_REQUESTS = "<|plamo:begin_tool_requests:plamo|>" |
| 34 | END_TOOL_REQUESTS = "<|plamo:end_tool_requests:plamo|>" |
| 35 | BEGIN_TOOL_REQUEST = "<|plamo:begin_tool_request:plamo|>" |
| 36 | END_TOOL_REQUEST = "<|plamo:end_tool_request:plamo|>" |
| 37 | BEGIN_TOOL_NAME = "<|plamo:begin_tool_name:plamo|>" |
| 38 | END_TOOL_NAME = "<|plamo:end_tool_name:plamo|>" |
| 39 | BEGIN_TOOL_ARGUMENTS = ( |
| 40 | "<|plamo:begin_tool_arguments:plamo|><|plamo:constrain|>json<|plamo:msg|>" |
| 41 | ) |
| 42 | END_TOOL_ARGUMENTS = "<|plamo:end_tool_arguments:plamo|>" |
| 43 | BEGIN_THINK = "<|plamo:begin_think:plamo|>" |
| 44 | END_THINK = "<|plamo:end_think:plamo|>" |
| 45 | EOT = "<|plamo:tag|>" |
| 46 | |
| 47 | # PLaMo3 control markers are split into multiple tokenizer tokens. These |
| 48 | # atomic pieces must not be treated as unrelated special tokens by the |
| 49 | # parser-engine drop-token handling. |
| 50 | PLAMO_MARKER_TOKENS = frozenset( |
| 51 | { |
| 52 | "<|plamo:begin_", |
| 53 | "<|plamo:end_", |
| 54 | ":plamo|>", |
| 55 | } |
| 56 | ) |
| 57 | |
| 58 | |
| 59 | @functools.cache |
| 60 | def plamo3_config(thinking: bool = True) -> ParserEngineConfig: |
| 61 | if thinking: |
| 62 | reasoning_terminals = { |
| 63 | "THINK_START": BEGIN_THINK, |
| 64 | "THINK_END": END_THINK, |
| 65 | } |
| 66 | reasoning_transitions = { |
| 67 | (ParserState.REASONING, "THINK_START"): Transition( |
| 68 | ParserState.REASONING, |
| 69 | ), |
| 70 | (ParserState.CONTENT, "THINK_START"): Transition( |
| 71 | ParserState.REASONING, |
| 72 | (EventType.REASONING_START,), |
| 73 | ), |
| 74 | (ParserState.REASONING, "THINK_END"): Transition( |
| 75 | ParserState.CONTENT, |
| 76 | (EventType.REASONING_END,), |
| 77 | ), |
| 78 | (ParserState.CONTENT, "THINK_END"): Transition( |
| 79 | ParserState.CONTENT, |
| 80 | (), |
| 81 | ), |
| 82 | (ParserState.REASONING, "TOOL_REQUESTS_START"): Transition( |
| 83 | ParserState.TOOL_PREAMBLE, |
| 84 | (EventType.REASONING_END,), |
| 85 | ), |
| 86 | } |
| 87 | initial_state = ParserState.REASONING |
| 88 | else: |
| 89 | reasoning_terminals = {} |
| 90 | reasoning_transitions = {} |
| 91 | initial_state = ParserState.CONTENT |
| 92 | |
| 93 | return ParserEngineConfig( |
| 94 | name="plamo3", |
| 95 | initial_state=initial_state, |
| 96 | terminals={ |
| 97 | **reasoning_terminals, |
| 98 | "TOOL_REQUESTS_START": BEGIN_TOOL_REQUESTS, |
| 99 | "TOOL_REQUESTS_END": END_TOOL_REQUESTS, |
| 100 | "TOOL_REQUEST_START": BEGIN_TOOL_REQUEST, |
| 101 | "TOOL_REQUEST_END": END_TOOL_REQUEST, |
| 102 | "TOOL_NAME_START": BEGIN_TOOL_NAME, |
| 103 | "TOOL_NAME_END": END_TOOL_NAME, |
| 104 | "TOOL_ARGS_START": BEGIN_TOOL_ARGUMENTS, |
| 105 | "TOOL_ARGS_END": END_TOOL_ARGUMENTS, |
| 106 | "EOT": EOT, |
| 107 | }, |
| 108 | transitions={ |
| 109 | **reasoning_transitions, |
| 110 | (ParserState.CONTENT, "TOOL_REQUESTS_START"): Transition( |
| 111 | ParserState.TOOL_PREAMBLE, |
| 112 | ), |
| 113 | (ParserState.TOOL_PREAMBLE, "TOOL_REQUEST_START"): Transition( |
| 114 | ParserState.TOOL_PREAMBLE, |
| 115 | ), |
| 116 | (ParserState.TOOL_PREAMBLE, "TOOL_NAME_START"): Transition( |
| 117 | ParserState.TOOL_NAME, |
| 118 | (EventType.TOOL_CALL_START,), |
| 119 | ), |
| 120 | (ParserState.TOOL_NAME, "TOOL_NAME_END"): Transition( |
| 121 | ParserState.TOOL_PREAMBLE, |
| 122 | ), |
| 123 | (ParserState.TOOL_PREAMBLE, "TOOL_ARGS_START"): Transition( |
| 124 | ParserState.TOOL_ARGS, |
| 125 | ), |
| 126 | (ParserState.TOOL_ARGS, "TOOL_ARGS_END"): Transition( |
| 127 | ParserState.TOOL_BETWEEN, |
| 128 | ), |
| 129 | (ParserState.TOOL_BETWEEN, "TOOL_REQUEST_END"): Transition( |
| 130 | ParserState.TOOL_PREAMBLE, |
| 131 | (EventType.TOOL_CALL_END,), |
| 132 | ), |
| 133 | (ParserState.TOOL_PREAMBLE, "TOOL_REQUESTS_END"): Transition( |
| 134 | ParserState.CONTENT, |
| 135 | ), |
| 136 | # EOT is normally consumed by generation stop handling. Keep it |
| 137 | # in the grammar as well so text-only parsing and explicitly |
| 138 | # supplied stop tokens cannot leak it into user-visible output. |
| 139 | (ParserState.REASONING, "EOT"): Transition( |
| 140 | ParserState.CONTENT, |
| 141 | (EventType.REASONING_END,), |
| 142 | ), |
| 143 | (ParserState.CONTENT, "EOT"): Transition(ParserState.CONTENT), |
| 144 | (ParserState.TOOL_PREAMBLE, "EOT"): Transition(ParserState.CONTENT), |
| 145 | (ParserState.TOOL_NAME, "EOT"): Transition(ParserState.CONTENT), |
| 146 | (ParserState.TOOL_ARGS, "EOT"): Transition(ParserState.CONTENT), |
| 147 | (ParserState.TOOL_BETWEEN, "EOT"): Transition(ParserState.CONTENT), |
| 148 | }, |
| 149 | token_id_terminals={"EOT": EOT, **reasoning_terminals}, |
| 150 | drop_whitespace_only_content_before_tools=True, |
| 151 | stream_arg_deltas=True, |
| 152 | strip_content_whitespace_with_tools=False, |
| 153 | strip_trailing_reasoning_whitespace=False, |
| 154 | preserve_tokens=PLAMO_MARKER_TOKENS, |
| 155 | tool_args_json=True, |
| 156 | ) |
| 157 | |
| 158 | |
| 159 | class Plamo3Parser(ParserEngine): |
| 160 | """PLaMo3 reasoning and tool-call parser backed by ParserEngine.""" |
| 161 | |
| 162 | def __init__( |
| 163 | self, |
| 164 | tokenizer: TokenizerLike, |
| 165 | tools: list[Tool] | None = None, |
| 166 | **kwargs, |
| 167 | ) -> None: |
| 168 | # Configure PLaMo parsing and cache its multi-token reasoning markers. |
| 169 | chat_kwargs = kwargs.get("chat_template_kwargs", {}) or {} |
| 170 | self.thinking_enabled = chat_kwargs.get("enable_thinking", True) |
| 171 | kwargs.setdefault( |
| 172 | "parser_engine_config", |
| 173 | plamo3_config(thinking=self.thinking_enabled), |
| 174 | ) |
| 175 | super().__init__(tokenizer, tools, **kwargs) |
| 176 | |
| 177 | self._reasoning_start_token_id_sequence: list[int] = list( |
| 178 | tokenizer.encode(BEGIN_THINK, add_special_tokens=False) |
| 179 | ) |
| 180 | self._reasoning_end_token_id_sequence: list[int] = list( |
| 181 | tokenizer.encode(END_THINK, add_special_tokens=False) |
| 182 | ) |
| 183 | # Token-ID sequences of every terminal literal, for token-level |
| 184 | # unfinished-marker detection. |
| 185 | self._terminal_token_id_sequences: dict[str, tuple[int, ...]] = { |
| 186 | literal: tuple(tokenizer.encode(literal, add_special_tokens=False)) |
| 187 | for literal in self.parser_engine_config.terminal_literals |
| 188 | } |
| 189 | |
| 190 | def is_reasoning_end(self, input_ids: list[int]) -> bool: |
| 191 | # Detect PLaMo reasoning boundaries that span multiple token IDs. |
| 192 | if not self.thinking_enabled: |
| 193 | return True |
| 194 | |
| 195 | start_ids = self._reasoning_start_token_id_sequence |
| 196 | end_ids = self._reasoning_end_token_id_sequence |
| 197 | for i in range(len(input_ids) - 1, -1, -1): |
| 198 | if input_ids[i : i + len(end_ids)] == end_ids: |
| 199 | return True |
| 200 | if input_ids[i : i + len(start_ids)] == start_ids: |
| 201 | return False |
| 202 | if input_ids[i] == self.vocab.get(EOT): |
| 203 | return False |
| 204 | return False |
| 205 | |
| 206 | def extract_content_ids(self, input_ids: list[int]) -> list[int]: |
| 207 | # Extract content after PLaMo's multi-token reasoning end marker. |
| 208 | if not self.thinking_enabled: |
| 209 | return input_ids |
| 210 | end_ids = self._reasoning_end_token_id_sequence |
| 211 | for i in range(len(input_ids) - len(end_ids), -1, -1): |
| 212 | if input_ids[i : i + len(end_ids)] == end_ids: |
| 213 | return input_ids[i + len(end_ids) :] |
| 214 | return input_ids |
| 215 | |
| 216 | def _strip_unfinished_marker(self, value: str | None) -> str | None: |
| 217 | if not value: |
| 218 | return value |
| 219 | marker_start = value.rfind("<|plamo:") |
| 220 | if marker_start < 0: |
| 221 | return value or None |
| 222 | suffix = value[marker_start:] |
| 223 | # Strip a trailing partial marker at the token-ID level. |
| 224 | suffix_ids = tuple( |
| 225 | self.model_tokenizer.encode(suffix, add_special_tokens=False) |
| 226 | ) |
| 227 | if any( |
| 228 | len(seq) > len(suffix_ids) and seq[: len(suffix_ids)] == suffix_ids |
| 229 | for seq in self._terminal_token_id_sequences.values() |
| 230 | ): |
| 231 | return value[:marker_start] or None |
| 232 | return value or None |
| 233 | |
| 234 | def _coalesce_finished_tool_events( |
| 235 | self, events: list[SemanticEvent] |
| 236 | ) -> list[SemanticEvent]: |
| 237 | combined = [] |
| 238 | for (kind, _), group in groupby( |
| 239 | events, key=lambda event: (event.type, event.tool_index) |
| 240 | ): |
| 241 | chunks = list(group) |
| 242 | if kind not in (EventType.TOOL_NAME, EventType.ARG_VALUE_CHUNK): |
| 243 | combined.extend(chunks) |
| 244 | continue |
| 245 | if kind == EventType.TOOL_NAME and chunks[0].tool_index < 0: |
| 246 | continue |
| 247 | value = "".join(chunk.value for chunk in chunks) |
| 248 | if kind == EventType.TOOL_NAME: |
| 249 | value = self._strip_unfinished_marker(value) or "" |
| 250 | if value: |
| 251 | combined.append(replace(chunks[0], value=value)) |
| 252 | return combined |
| 253 | |
| 254 | def _events_to_delta( |
| 255 | self, |
| 256 | events: list[SemanticEvent], |
| 257 | finished: bool = False, |
| 258 | ) -> DeltaMessage | None: |
| 259 | # Coalesce flushed tool fragments before stripping truncated markers. |
| 260 | if not finished: |
| 261 | return super()._events_to_delta(events) |
| 262 | |
| 263 | events = self._coalesce_finished_tool_events(events) |
| 264 | delta = super()._events_to_delta(events, finished=True) |
| 265 | if delta is None: |
| 266 | return None |
| 267 | |
| 268 | delta.reasoning = self._strip_unfinished_marker(delta.reasoning) |
| 269 | if not self.skip_tool_parsing: |
| 270 | delta.content = self._strip_unfinished_marker(delta.content) |
| 271 | # The flush emits the remaining (never-streamed) tool-call |
| 272 | # arguments; an empty remainder must match non-streaming |
| 273 | # extraction (_extract_args_json), which normalizes it to '{}'. |
| 274 | for tc in delta.tool_calls or []: |
| 275 | if tc.function is not None and not (tc.function.arguments or "").strip(): |
| 276 | tc.function.arguments = "{}" |
| 277 | if delta.reasoning is None and delta.content is None and not delta.tool_calls: |
| 278 | return None |
| 279 | return delta |
| 280 | |
| 281 | def _handle_arg_chunk( |
| 282 | self, |
| 283 | event: SemanticEvent, |
| 284 | deltas: list[DeltaToolCall], |
| 285 | ) -> None: |
| 286 | # Strip truncated marker suffixes and preserve the first argument delta. |
| 287 | idx = event.tool_index |
| 288 | slot = self._tool_slots[idx] |
| 289 | if (marker_pos := event.value.rfind("<|plamo:")) >= 0: |
| 290 | stripped = event.value[:marker_pos] |
| 291 | try: |
| 292 | json.loads(slot.args + stripped) |
| 293 | except ValueError: |
| 294 | pass |
| 295 | else: |
| 296 | event = replace(event, value=stripped) |
| 297 | name_sent_before = slot.name_sent |
| 298 | super()._handle_arg_chunk(event, deltas) |
| 299 | if event.value and not name_sent_before and slot.name_sent: |
| 300 | deltas.append( |
| 301 | DeltaToolCall( |
| 302 | index=idx, |
| 303 | function=DeltaFunctionCall(arguments=event.value), |
| 304 | ) |
| 305 | ) |
| 306 | |
| 307 | def _extract_args_json(self, raw_args: str, func_name: str) -> str: |
| 308 | # Return the raw arguments as JSON |
| 309 | # since PLaMo3 generates function names and arguments separately. |
| 310 | return raw_args.strip() or "{}" |
| 311 | |
| 312 | def extract_reasoning( |
| 313 | self, |
| 314 | model_output: str, |
| 315 | request: ChatCompletionRequest | ResponsesRequest, |
| 316 | ) -> tuple[str | None, str | None]: |
| 317 | reasoning, content = super().extract_reasoning(model_output, request) |
| 318 | return self._strip_unfinished_marker(reasoning), content |