1from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
2import torch
3from peft import PeftModel, PeftConfig
4
5
6class GoralConversation:
7 def __init__(
8 self,
9 message_template=" <s> {role}\n{content} </s>\n",
10 system_prompt="Ты — Горал, русскоязычный автоматический ассистент. Ты разговариваешь с людьми и помогаешь им.",
11 start_token_id=1,
12 bot_token_id=9225,
13 ):
14 self.message_template = message_template
15 self.start_token_id = start_token_id
16 self.bot_token_id = bot_token_id
17 self.messages = [{"role": "system", "content": system_prompt}]
18
19 def get_start_token_id(self):
20 return self.start_token_id
21
22 def get_bot_token_id(self):
23 return self.bot_token_id
24
25 def add_user_message(self, message):
26 self.messages.append({"role": "user", "content": message})
27
28 def add_bot_message(self, message):
29 self.messages.append({"role": "bot", "content": message})
30
31 def get_prompt(self, tokenizer):
32 final_text = ""
33 for message in self.messages:
34 message_text = self.message_template.format(**message)
35 final_text += message_text
36 final_text += tokenizer.decode(
37 [
38 self.start_token_id,
39 ]
40 )
41 final_text += " "
42 final_text += tokenizer.decode([self.bot_token_id])
43 return final_text.strip()
44
45
46def generate(model, tokenizer, prompt, generation_config):
47 data = tokenizer(
48 prompt,
49 return_tensors="pt",
50 truncation=True,
51 max_length=2048,
52 )
53 data = {k: v.to(model.device) for k, v in data.items()}
54 output_ids = model.generate(**data, generation_config=generation_config)[0]
55 output_ids = output_ids[len(data["input_ids"][0]) :]
56 output = tokenizer.decode(output_ids, skip_special_tokens=True)
57 return output.strip()
58
59
60weights_path = "dim/llama2_7b_dolly_oasst1_chip2"
61access_token = ""
62
63config = PeftConfig.from_pretrained(weights_path)
64model = AutoModelForCausalLM.from_pretrained(
65 config.base_model_name_or_path,
66 load_in_8bit=True,
67 torch_dtype=torch.float16,
68 device_map={"": 0},
69 token=access_token,
70)
71model = PeftModel.from_pretrained(
72 model,
73 weights_path,
74 torch_dtype=torch.float16,
75)
76model.eval()
77
78tokenizer = AutoTokenizer.from_pretrained(weights_path)
79generation_config = GenerationConfig.from_pretrained(weights_path)
80generation_config.do_sample = False
81
82
83inp = "Напишите интересный пост в блоге о недавней поездке на Гавайи, рассказывая о культурном опыте и достопримечательностях, которые обязательно нужно увидеть."
84conversation = GoralConversation(
85 start_token_id=1,
86 bot_token_id=9225,
87)
88conversation.add_user_message(inp)
89prompt = conversation.get_prompt(tokenizer)
90
91output = generate(model, tokenizer, prompt, generation_config)
92print(inp)
93print(output)
94
