[Bugfix] Support MistralCommonBackend tokenizers in structured output (#52720)
Signed-off-by: Thanh Phan <thanhphantuan1110@gmail.com>
This commit is contained in:
@@ -14,6 +14,7 @@ from typing import TYPE_CHECKING, TypeVar
|
||||
import regex as re
|
||||
import torch
|
||||
from cachetools import LRUCache
|
||||
from transformers import MistralCommonBackend
|
||||
|
||||
import vllm.envs as envs
|
||||
from vllm.logger import init_logger
|
||||
@@ -305,6 +306,18 @@ re_llama_byte_token = re.compile(r"^<0x[0-9A-F]{2}>$")
|
||||
re_replacement_seq = re.compile(r"^.{0,6}�+.{0,6}$")
|
||||
|
||||
|
||||
def maybe_wrap_mistral_common_tokenizer(tokenizer: TokenizerLike) -> TokenizerLike:
|
||||
"""The grammar backends cannot consume a `MistralCommonBackend` directly."""
|
||||
if not isinstance(tokenizer, MistralCommonBackend):
|
||||
return tokenizer
|
||||
|
||||
# Deferred: `vllm.tokenizers.mistral` pulls in a large dependency tree, and
|
||||
# this module is imported by `vllm.v1.worker.gpu_model_runner`.
|
||||
from vllm.tokenizers.mistral import MistralTokenizer
|
||||
|
||||
return MistralTokenizer(tokenizer)
|
||||
|
||||
|
||||
def _reduced_vocabulary(tokenizer: TokenizerLike) -> dict[bytes, list[int]]:
|
||||
"""Create a map from vocabulary tokens to lists of equivalent token ids.
|
||||
|
||||
|
||||
Reference in New Issue
Block a user