Skip to content

vllm.tokenizers.mistral

Functions:

tekken_convert_ids_to_tokens(tokenizer, ids)

Convert ids to pieces, using raw bytes for byte-fallback tokens.

Source code in vllm/tokenizers/mistral.py
def tekken_convert_ids_to_tokens(
    tokenizer: "Tekkenizer", ids: Sequence[int]
) -> list[str | bytes]:
    """Convert ids to pieces, using raw `bytes` for byte-fallback tokens."""
    tokens: list[str | bytes] = [tokenizer.id_to_piece(i) for i in ids]
    if any("�" in t for t in tokens):
        tokens = [
            tokenizer.id_to_byte_piece(i, SpecialTokenPolicy.KEEP)
            if i >= tokenizer.num_special_tokens
            else tokenizer.decode([i], SpecialTokenPolicy.KEEP)
            for i in ids
        ]
    return tokens

tekken_convert_tokens_to_string(tokenizer, tokens)

Reassemble pieces from tekken_convert_ids_to_tokens into text.

Source code in vllm/tokenizers/mistral.py
def tekken_convert_tokens_to_string(
    tokenizer: "Tekkenizer", tokens: Sequence[str | bytes]
) -> str:
    """Reassemble pieces from `tekken_convert_ids_to_tokens` into text."""
    if any(isinstance(t, bytes) for t in tokens):
        ids = [_tekken_token_to_id(tokenizer, t) for t in tokens]
        return tokenizer.decode(ids, SpecialTokenPolicy.KEEP)
    return "".join(cast(Sequence[str], tokens))

to_mistral_response_format(response_format)

Convert an OpenAI response format to its mistral-common equivalent.

Returns:

  • ResponseFormat | None –

    The converted format, or None for formats that do not affect Mistral

  • ResponseFormat | None –

    prompt rendering (text, structural tags, schema-less json_schema).

Source code in vllm/tokenizers/mistral.py
def to_mistral_response_format(
    response_format: AnyResponseFormat | None,
) -> MistralResponseFormat | None:
    """Convert an OpenAI response format to its mistral-common equivalent.

    Returns:
        The converted format, or `None` for formats that do not affect Mistral
        prompt rendering (`text`, structural tags, schema-less `json_schema`).

    """
    if response_format is None:
        return None

    if response_format.type == "json_object":
        return MistralResponseFormat(type=MistralResponseFormats.json_object)

    if response_format.type != "json_schema":
        return None

    json_schema = response_format.json_schema
    if json_schema is None or json_schema.json_schema is None:
        return None

    return MistralResponseFormat(
        type=MistralResponseFormats.json_schema,
        json_schema=MistralJsonSchema(
            name=json_schema.name,
            schema=json_schema.json_schema,
            strict=bool(json_schema.strict),
        ),
    )

truncate_tool_call_ids(request)

Truncates tool call IDs for Mistral's ID requirements.

Source code in vllm/tokenizers/mistral.py
def truncate_tool_call_ids(request: "MistralChatCompletionRequest"):
    """Truncates tool call IDs for Mistral's ID requirements."""
    for i, message in enumerate(request.messages):
        if message.get("role") == "assistant":
            tool_calls = message.get("tool_calls", [])
            for tool_call in tool_calls:
                if len(tool_call["id"]) > 9:
                    logger.warning(
                        "Truncating tool call ID: %s to %s",
                        tool_call["id"],
                        tool_call["id"][-9:],
                    )
                    tool_call["id"] = tool_call["id"][-9:]

            request.messages[i]["tool_calls"] = tool_calls

        elif message.get("role") in {"tool_results", "tool"}:
            if "tool_call_id" in message:
                tool_call_id = message["tool_call_id"]

                if len(tool_call_id) > 9:
                    logger.warning(
                        "Truncating tool_call_id: %s to %s",
                        tool_call_id,
                        tool_call_id[-9:],
                    )
                    tool_call_id = tool_call_id[-9:]
                request.messages[i]["tool_call_id"] = tool_call_id