[Feature] Lazy import for the "mistral" tokenizer module. (#34651)

Signed-off-by: Neil Schemenauer <nas@arctrix.com>
This commit is contained in:
Neil Schemenauer
2026-02-23 00:43:01 -08:00
committed by GitHub
parent e631f8e78e
commit 54e2f83d0a
14 changed files with 68 additions and 48 deletions
+3 -3
View File
@@ -83,9 +83,9 @@ from vllm.renderers.inputs.preprocess import (
from vllm.sampling_params import BeamSearchParams, RequestOutputKind, SamplingParams
from vllm.tasks import PoolingTask
from vllm.tokenizers import TokenizerLike
from vllm.tokenizers.mistral import MistralTokenizer
from vllm.usage.usage_lib import UsageContext
from vllm.utils.counter import Counter
from vllm.utils.mistral import is_mistral_tokenizer
from vllm.utils.tqdm_utils import maybe_tqdm
from vllm.v1.engine.llm_engine import LLMEngine
from vllm.v1.sample.logits_processor import LogitsProcessor
@@ -891,7 +891,7 @@ class LLM:
add_generation_prompt=add_generation_prompt,
continue_final_message=continue_final_message,
tools=tools,
tokenize=isinstance(renderer.tokenizer, MistralTokenizer),
tokenize=is_mistral_tokenizer(renderer.tokenizer),
),
),
)
@@ -1458,7 +1458,7 @@ class LLM:
model_config = self.model_config
tokenizer = self.get_tokenizer()
if isinstance(tokenizer, MistralTokenizer):
if is_mistral_tokenizer(tokenizer):
raise ValueError("Score API is not supported for Mistral tokenizer")
if len(data_1) == 1:
@@ -75,16 +75,12 @@ from vllm.parser import ParserManager
from vllm.reasoning import ReasoningParser
from vllm.sampling_params import BeamSearchParams, SamplingParams
from vllm.tokenizers import TokenizerLike
from vllm.tokenizers.mistral import (
MistralTokenizer,
maybe_serialize_tool_calls,
truncate_tool_call_ids,
validate_request_params,
)
from vllm.tool_parsers import ToolParser
from vllm.tool_parsers.mistral_tool_parser import MistralToolCall
from vllm.tool_parsers.utils import partial_json_loads
from vllm.utils.collection_utils import as_list
from vllm.utils.mistral import is_mistral_tokenizer
from vllm.utils.mistral import mt as _mt
logger = init_logger(__name__)
@@ -244,18 +240,18 @@ class OpenAIServingChat(OpenAIServing):
tool_parser = self.tool_parser
if isinstance(tokenizer, MistralTokenizer):
if is_mistral_tokenizer(tokenizer):
# because of issues with pydantic we need to potentially
# re-serialize the tool_calls field of the request
# for more info: see comment in `maybe_serialize_tool_calls`
maybe_serialize_tool_calls(request) # type: ignore[arg-type]
truncate_tool_call_ids(request) # type: ignore[arg-type]
validate_request_params(request)
_mt.maybe_serialize_tool_calls(request) # type: ignore[arg-type]
_mt.truncate_tool_call_ids(request) # type: ignore[arg-type]
_mt.validate_request_params(request)
# Check if tool parsing is unavailable (common condition)
tool_parsing_unavailable = (
tool_parser is None
and not isinstance(tokenizer, MistralTokenizer)
and not is_mistral_tokenizer(tokenizer)
and not self.use_harmony
)
@@ -639,8 +635,6 @@ class OpenAIServingChat(OpenAIServing):
request_metadata: RequestResponseMetadata,
reasoning_parser: ReasoningParser | None = None,
) -> AsyncGenerator[str, None]:
from vllm.tokenizers.mistral import MistralTokenizer
created_time = int(time.time())
chunk_object_type: Final = "chat.completion.chunk"
first_iteration = True
@@ -955,7 +949,7 @@ class OpenAIServingChat(OpenAIServing):
)
else:
# Generate ID based on tokenizer type
if isinstance(tokenizer, MistralTokenizer):
if is_mistral_tokenizer(tokenizer):
tool_call_id = MistralToolCall.generate_random_id()
else:
tool_call_id = make_tool_call_id(
@@ -1516,7 +1510,7 @@ class OpenAIServingChat(OpenAIServing):
tool_parser_cls=self.tool_parser,
)
tool_call_class = (
MistralToolCall if isinstance(tokenizer, MistralTokenizer) else ToolCall
MistralToolCall if is_mistral_tokenizer(tokenizer) else ToolCall
)
if self.use_harmony:
# Harmony models already have parsed content and tool_calls
@@ -1951,7 +1945,7 @@ class OpenAIServingChat(OpenAIServing):
# because of issues with pydantic we need to potentially
# re-serialize the tool_calls field of the request
# for more info: see comment in `maybe_serialize_tool_calls`
maybe_serialize_tool_calls(request) # type: ignore[arg-type]
_mt.maybe_serialize_tool_calls(request) # type: ignore[arg-type]
# Add system message.
# NOTE: In Chat Completion API, browsing is enabled by default
+2 -3
View File
@@ -128,6 +128,7 @@ from vllm.utils.async_utils import (
collect_from_async_generator,
merge_async_iterators,
)
from vllm.utils.mistral import is_mistral_tokenizer
class GenerationError(Exception):
@@ -976,15 +977,13 @@ class OpenAIServing:
tool_dicts: list[dict[str, Any]] | None = None,
tool_parser: Callable[[TokenizerLike], ToolParser] | None = None,
) -> tuple[list[ConversationMessage], list[ProcessorInputs]]:
from vllm.tokenizers.mistral import MistralTokenizer
renderer = self.renderer
default_template_kwargs = merge_kwargs(
default_template_kwargs,
dict(
tools=tool_dicts,
tokenize=isinstance(renderer.tokenizer, MistralTokenizer),
tokenize=is_mistral_tokenizer(renderer.tokenizer),
),
)
+2 -2
View File
@@ -41,8 +41,8 @@ from vllm.logger import init_logger
from vllm.lora.request import LoRARequest
from vllm.outputs import PoolingRequestOutput, ScoringRequestOutput
from vllm.tokenizers import TokenizerLike
from vllm.tokenizers.mistral import MistralTokenizer
from vllm.utils.async_utils import make_async, merge_async_iterators
from vllm.utils.mistral import is_mistral_tokenizer
logger = init_logger(__name__)
@@ -348,7 +348,7 @@ class ServingScores(OpenAIServing):
trace_headers: Mapping[str, str] | None = None,
) -> list[PoolingRequestOutput] | ErrorResponse:
tokenizer = self.renderer.get_tokenizer()
if isinstance(tokenizer, MistralTokenizer):
if is_mistral_tokenizer(tokenizer):
raise ValueError("MistralTokenizer not supported for cross-encoding")
model_config = self.model_config