| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from __future__ import annotations |
| 5 | |
| 6 | from typing import TYPE_CHECKING |
| 7 | |
| 8 | from vllm.parser.engine.registered_adapters import MistralParserToolAdapter |
| 9 | |
| 10 | if TYPE_CHECKING: |
| 11 | from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest |
| 12 | from vllm.entrypoints.openai.responses.protocol import ResponsesRequest |
| 13 | |
| 14 | |
| 15 | class MistralToolParser(MistralParserToolAdapter): # type: ignore[valid-type, misc] |
| 16 | # Marked so is_mistral_tool_parser() in vllm.utils.mistral recognises |
| 17 | # this adapter and lets adjust_request() fire for tool_choice="none" |
| 18 | # on grammar-capable Mistral tokenizers. |
| 19 | IS_MISTRAL_TOOL_PARSER = True |
| 20 | |
| 21 | # Mistral emits tool calls in the [TOOL_CALLS]name[ARGS]{...} format, not the |
| 22 | # plain JSON array / bare-args that the serving layer's generic required and |
| 23 | # named handlers expect. Route required/named through this parser's own |
| 24 | # extraction (treated like "auto") so the [TOOL_CALLS] format is parsed |
| 25 | # correctly instead of crashing or dumping the raw envelope into arguments. |
| 26 | supports_required_and_named = False |
| 27 | |
| 28 | @property |
| 29 | def bot_token(self) -> str: |
| 30 | """Return the Mistral tool-call marker exposed by the legacy parser.""" |
| 31 | return self._parser_engine.bot_token |
| 32 | |
| 33 | @property |
| 34 | def bot_token_id(self) -> int | None: |
| 35 | """Return the token id for :attr:`bot_token`.""" |
| 36 | return self._parser_engine.bot_token_id |
| 37 | |
| 38 | def adjust_request( |
| 39 | self, |
| 40 | request: ChatCompletionRequest | ResponsesRequest, |
| 41 | ) -> ChatCompletionRequest | ResponsesRequest: |
| 42 | # Skip the base ToolParser tool_choice -> structured_outputs |
| 43 | # conversion. Mistral enforces tool_choice via the grammar ``mode`` |
| 44 | # (auto/none/required/named); forwarding a tool-derived json_schema |
| 45 | # would be unioned into the grammar by mistral-common, allowing raw |
| 46 | # JSON in place of a real [TOOL_CALLS] call. Genuine user-provided |
| 47 | # structured output on the request is still honored by the engine's |
| 48 | # adjust_request. |
| 49 | return self._parser_engine.adjust_request(request) |