| from transformers import AutoTokenizer | |
| import os | |
| from typing import Union | |
| class TokenizerWrapper: | |
| def __init__(self, tokenizer_name_or_path, tokenizer_revision, trust_remote_code): | |
| print(f"tokenizer_name_or_path: {tokenizer_name_or_path}, tokenizer_revision: {tokenizer_revision}, trust_remote_code: {trust_remote_code}") | |
| self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path, revision=tokenizer_revision or "main", trust_remote_code=trust_remote_code) | |
| self.custom_chat_template = os.getenv("CUSTOM_CHAT_TEMPLATE") | |
| self.has_chat_template = bool(self.tokenizer.chat_template) or bool(self.custom_chat_template) | |
| if self.custom_chat_template and isinstance(self.custom_chat_template, str): | |
| self.tokenizer.chat_template = self.custom_chat_template | |
| def apply_chat_template(self, input: Union[str, list[dict[str, str]]]) -> str: | |
| if isinstance(input, list): | |
| if not self.has_chat_template: | |
| raise ValueError( | |
| "Chat template does not exist for this model, you must provide a single string input instead of a list of messages" | |
| ) | |
| elif isinstance(input, str): | |
| input = [{"role": "user", "content": input}] | |
| else: | |
| raise ValueError("Input must be a string or a list of messages") | |
| return self.tokenizer.apply_chat_template( | |
| input, tokenize=False, add_generation_prompt=True | |
| ) | |