biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

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

Browse Source 
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
Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/sampling_params.py
View File

@@ -16,6 +16,7 @@ from vllm.config import ModelConfig, SpeculativeConfig, StructuredOutputsConfig
from vllm.exceptions import VLLMValidationError
from vllm.logger import init_logger
from vllm.tokenizers import TokenizerLike
from vllm.utils.mistral import is_mistral_tokenizer
from vllm.v1.serial_utils import PydanticMsgspecMixin
logger = init_logger(__name__)
@@ -731,7 +732,6 @@ class SamplingParams(
):
raise ValueError("structured_outputs.grammar cannot be an empty string")
from vllm.tokenizers.mistral import MistralTokenizer
from vllm.v1.structured_output.backend_guidance import (
has_guidance_unsupported_json_features,
validate_guidance_grammar,
@@ -752,7 +752,7 @@ class SamplingParams(
# allows <|special_token|> and similar, see
# https://github.com/guidance-ai/llguidance/blob/main/docs/syntax.md#special-tokens
# Without tokenizer these are disallowed in grammars.
if isinstance(tokenizer, MistralTokenizer):
if is_mistral_tokenizer(tokenizer):
raise ValueError(
"Mistral tokenizer is not supported for the 'guidance' "
"structured output backend. Please use ['xgrammar', 'outlines'] "
@@ -764,7 +764,7 @@ class SamplingParams(
validate_structured_output_request_outlines(self)
elif backend == "lm-format-enforcer":
# lm format enforcer backend
if isinstance(tokenizer, MistralTokenizer):
if is_mistral_tokenizer(tokenizer):
raise ValueError(
"Mistral tokenizer is not supported for the 'lm-format-enforcer' "
"structured output backend. Please use ['xgrammar', 'outlines'] "
@@ -796,7 +796,7 @@ class SamplingParams(
schema = so_params.json
skip_guidance = has_guidance_unsupported_json_features(schema)
if isinstance(tokenizer, MistralTokenizer) or skip_guidance:
if is_mistral_tokenizer(tokenizer) or skip_guidance:
# Fall back to outlines if the tokenizer is Mistral
# or if schema contains features unsupported by guidance
validate_structured_output_request_outlines(self)