| from __future__ import annotations |
|
|
| from dataclasses import dataclass |
| from typing import Any |
|
|
|
|
| @dataclass |
| class LocalHFArenaModel: |
| """Lazy, GPU-only Transformers backend for the frozen arena evaluator.""" |
|
|
| model_id: str |
| adapter_path: str | None = None |
| max_new_tokens: int = 224 |
|
|
| def __post_init__(self) -> None: |
| import torch |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
| self.name = self.adapter_path or self.model_id |
| self.tokenizer = AutoTokenizer.from_pretrained(self.model_id) |
| self.model = AutoModelForCausalLM.from_pretrained( |
| self.model_id, |
| dtype=torch.bfloat16, |
| attn_implementation="sdpa", |
| ).to("cuda") |
| if self.adapter_path: |
| from peft import PeftModel |
|
|
| self.model = PeftModel.from_pretrained(self.model, self.adapter_path) |
| self.model.eval() |
|
|
| def respond(self, messages: list[dict[str, str]], oracle_target: str) -> str: |
| del oracle_target |
| import torch |
|
|
| kwargs: dict[str, Any] = { |
| "tokenize": True, |
| "add_generation_prompt": True, |
| "return_tensors": "pt", |
| } |
| try: |
| input_ids = self.tokenizer.apply_chat_template( |
| messages, enable_thinking=False, **kwargs |
| ) |
| except TypeError: |
| input_ids = self.tokenizer.apply_chat_template(messages, **kwargs) |
| input_ids = input_ids.to(self.model.device) |
| attention_mask = torch.ones_like(input_ids) |
| with torch.inference_mode(): |
| output = self.model.generate( |
| input_ids=input_ids, |
| attention_mask=attention_mask, |
| max_new_tokens=self.max_new_tokens, |
| do_sample=False, |
| pad_token_id=self.tokenizer.eos_token_id, |
| ) |
| return self.tokenizer.decode( |
| output[0, input_ids.shape[1] :], skip_special_tokens=True |
| ).strip() |
|
|
