[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
@@ -23,7 +23,7 @@ from vllm.multimodal.cache import MultiModalProcessorOnlyCache
from vllm.multimodal.inputs import MultiModalInputs, batched_tensors_equal
from vllm.multimodal.processing import BaseMultiModalProcessor, InputProcessingContext
from vllm.tokenizers import TokenizerLike, cached_tokenizer_from_config
from vllm.tokenizers.mistral import MistralTokenizer
from vllm.utils.mistral import is_mistral_tokenizer
from ....multimodal.utils import random_audio, random_image, random_video
from ...registry import (
@@ -183,7 +183,7 @@ def get_text_token_prompts(
text_prompt: str | None
token_prompt: list[int]
if isinstance(tokenizer, MistralTokenizer):
if is_mistral_tokenizer(tokenizer):
# ChatCompletionRequest only supports ImageChunk natively;
# for other modalities (e.g. audio), fall back to the model's
# own dummy inputs builder which knows the right placeholders.
+3 -3
View File
@@ -4,7 +4,7 @@
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
from vllm.entrypoints.openai.engine.protocol import DeltaMessage
from vllm.reasoning import ReasoningParser
from vllm.tokenizers.mistral import MistralTokenizer
from vllm.utils.mistral import is_mistral_tokenizer
class StreamingReasoningReconstructor:
@@ -59,7 +59,7 @@ def run_reasoning_extraction_mistral(
request: ChatCompletionRequest | None = None,
streaming: bool = False,
) -> tuple[str | None, str | None]:
assert isinstance(reasoning_parser.model_tokenizer, MistralTokenizer), type(
assert is_mistral_tokenizer(reasoning_parser.model_tokenizer), type(
reasoning_parser.model_tokenizer
)
if streaming:
@@ -130,7 +130,7 @@ def run_reasoning_extraction_streaming_mistral(
model_deltas: list[int],
request: ChatCompletionRequest | None = None,
) -> StreamingReasoningReconstructor:
assert isinstance(reasoning_parser.model_tokenizer, MistralTokenizer), type(
assert is_mistral_tokenizer(reasoning_parser.model_tokenizer), type(
reasoning_parser.model_tokenizer
)
request = request or ChatCompletionRequest(messages=[], model="test-model")