[Tokenizer] Fix TokenizerFast missing clean_up_tokenization_spaces

[dynamicheart]

PR types

Bug Fixes

PR changes

Models

Description

问题：使用LLamaTokenizerFast出现问题

复现代码

from paddlenlp.transformers import LlamaTokenizerFast
tokenizer = LlamaTokenizerFast.from_pretrained('meta-llama/Llama-2-13b')
output = tokenizer._decode([0, 3, 2, 1])

结果：

Traceback (most recent call last):
  File "test_tokenizer.py", line 6, in <module>
    output = tokenizer._decode([0, 3, 2, 1])
  File "/workspace/mnt/PaddleNLP/paddlenlp/transformers/tokenizer_utils_fast.py", line 653, in _decode
    else self.clean_up_tokenization_spaces
AttributeError: 'LlamaTokenizerFast' object has no attribute 'clean_up_tokenization_spaces'

修复参考链接：https://github.com/huggingface/transformers/blob/e5d14f39ad82475f238cad41279a5d61ac5db287/src/transformers/tokenization_utils_base.py/#L1616

[paddle-bot]

Thanks for your contribution!

[CLAassistant]

All committers have signed the CLA.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 52.74%. Comparing base (3443b9f) to head (ee6d354).
Report is 3 commits behind head on develop.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #9304      +/-   ##
===========================================
+ Coverage    52.62%   52.74%   +0.12%     
===========================================
  Files          661      661              
  Lines       107365   107366       +1     
===========================================
+ Hits         56499    56631     +132     
+ Misses       50866    50735     -131     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.