[Mistral Tokenizer] allow more leniency in apply_chat_template (#41658)

Signed-off-by: juliendenize <julien.denize@mistral.ai>
Co-authored-by: Roger Wang <hey@rogerw.io>
This commit is contained in:
Julien Denize
2026-05-05 19:56:15 -07:00
committed by GitHub
co-authored by Roger Wang
parent c7aa186d67
commit 16e336491e
10 changed files with 144 additions and 397 deletions
+4 -51
View File
@@ -13,7 +13,6 @@ from mistral_common.protocol.instruct.request import (
from mistral_common.protocol.instruct.request import (
ReasoningEffort,
)
from mistral_common.protocol.instruct.tool_calls import Function, Tool
from mistral_common.protocol.instruct.validator import ValidationMode
from mistral_common.tokens.tokenizers.base import (
SpecialTokenPolicy,
@@ -68,36 +67,6 @@ def _pop_unallowed_keys_and_warn(
)
# TODO(juliendenize): remove this once OpenAI API is better supported by
# `mistral-common`.
def adapt_inplace_to_mistral_tool(
tool: dict[str, Any],
) -> dict[str, Any]:
tools_fields = set(Tool.model_fields.keys())
function_fields = set(Function.model_fields.keys())
# The Mistral client, in comparison to the OpenAI client, requires the
# "parameters" dict and the "description" string to be present
# even if they are empty.
if function := tool.get("function"):
if function.get("parameters") is None:
function["parameters"] = {}
if function.get("description") is None:
function["description"] = ""
_pop_unallowed_keys_and_warn(
dictionary=function,
allowed_keys=function_fields,
err_dict_name="function",
)
_pop_unallowed_keys_and_warn(
dictionary=tool, allowed_keys=tools_fields, err_dict_name="tools"
)
return tool
def maybe_serialize_tool_calls(request: "MistralChatCompletionRequest"):
# SEE: https://github.com/vllm-project/vllm/pull/9951
# Credits go to: @gcalmettes
@@ -167,12 +136,11 @@ def truncate_tool_call_ids(request: "MistralChatCompletionRequest"):
request.messages[i]["tool_call_id"] = tool_call_id
def _prepare_apply_chat_template_tools_and_messages(
def _validate_apply_chat_template_args(
messages: list["ChatCompletionMessageParam"],
tools: list[dict[str, Any]] | None = None,
continue_final_message: bool = False,
add_generation_prompt: bool = False,
) -> tuple[list["ChatCompletionMessageParam"], list[dict[str, Any]] | None]:
) -> None:
if add_generation_prompt and continue_final_message:
raise ValueError(
"Cannot set both `add_generation_prompt` and "
@@ -196,21 +164,6 @@ def _prepare_apply_chat_template_tools_and_messages(
"the last message is not from the assistant."
)
# mistral-common requires AssistantMessage content to be string [1].
#
# [1]: https://github.com/mistralai/mistral-common/blob/f4a06998b75ed78bbf5aaf569590b772ea26c9f6/src/mistral_common/protocol/instruct/messages.py#L80
for message in messages:
# Remove reasoning as unsupported by Mistral
_ = message.pop("reasoning", None) # type: ignore
tools = (
[adapt_inplace_to_mistral_tool(tool=tool) for tool in tools]
if tools is not None
else None
)
return messages, tools
def validate_request_params(request: "ChatCompletionRequest"):
if request.chat_template is not None or request.chat_template_kwargs is not None:
@@ -449,8 +402,8 @@ class MistralTokenizer(TokenizerLike):
if self.version >= 15:
version_kwargs["reasoning_effort"] = kwargs.get("reasoning_effort")
messages, tools = _prepare_apply_chat_template_tools_and_messages(
messages, tools, continue_final_message, add_generation_prompt
_validate_apply_chat_template_args(
messages, continue_final_message, add_generation_prompt
)
return self.transformers_tokenizer.apply_chat_template(
+2 -7
View File
@@ -43,7 +43,7 @@ from vllm.logger import init_logger
from vllm.reasoning.mistral_reasoning_parser import MistralReasoningParser
from vllm.sampling_params import StructuredOutputsParams
from vllm.tokenizers import TokenizerLike
from vllm.tokenizers.mistral import MistralTokenizer, adapt_inplace_to_mistral_tool
from vllm.tokenizers.mistral import MistralTokenizer
from vllm.tool_parsers.abstract_tool_parser import (
Tool,
ToolParser,
@@ -241,12 +241,7 @@ class MistralToolParser(ToolParser):
)
mistral_tools = (
[
MistralTool.model_validate(
adapt_inplace_to_mistral_tool(tool.model_dump())
)
for tool in request.tools
]
[MistralTool.from_openai(tool.model_dump()) for tool in request.tools]
if request.tools is not None
else None
)