shared_utils/generate.py at main · GeodesicResearch/shared_utils · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
from typing import Optional, List, Dict

import torch
from torch import Tensor as _T

from transformers import AutoModelForCausalLM, AutoTokenizer


Conversations = List[List[Dict[str, str]]]


def format_conversation(user_prompts: List[str], system_prompt: Optional[str | List[str]] = None) -> Conversations:

    if isinstance(system_prompt, str):
        batch_messages = [[{"role": "system", "content": system_prompt}] for _ in user_prompts]
    elif isinstance(system_prompt, list):
        assert len(system_prompt) == len(user_prompts)
        batch_messages = [[{"role": "system", "content": sp}] for sp in system_prompt]
    else:
        assert system_prompt is None
        batch_messages = [[] for _ in user_prompts]

    for bm, up in zip(batch_messages, user_prompts):
        bm.append({"role": "user", "content": up})

    return batch_messages


def transform_conversations(conversations: Conversations, prefiller: Optional[str | List[str]] = None) -> List[str]:
    print('transform_conversations currently only for Deepseek models!')

    output_conversations = []

    if isinstance(prefiller, list):
        assert len(prefiller) == len(conversations)

    for c_i, conversation in enumerate(conversations):

        output_conversation = ""

        for message in conversation:
            role = message["role"]
            content = message["content"]
            if role == "user":
                output_conversation += f"<｜User｜> {content}\n"
            elif role == "assistant" or role == "system":                   # Doing system prompts in first person, to make them look like OCR
                output_conversation += f"<｜Assistant｜> {content}\n"

        if isinstance(prefiller, list):
            output_conversation += f"<｜Assistant｜> {prefiller[c_i]}\n"
        elif isinstance(prefiller, str):
            output_conversation += f"<｜Assistant｜> {prefiller}\n"
        else:
            assert output_conversation is None

        output_conversations.append(output_conversation)

    return output_conversations


def full_tokenize(prompts: List[str], tokenizer: AutoTokenizer, device: str, add_special_tokens: bool = True) -> List[Dict[str, torch.TensorType]]:
    print('full_tokenize currently only for Deepseek models!')
    return tokenizer(
        prompts,
        return_tensors="pt",
        padding=True,
        truncation=True,
        add_special_tokens=add_special_tokens
    ).to(device)


def generate_text(model: AutoModelForCausalLM, prompt: str | List[str], system_prompt: str| List[str], prefiller: Optional[str| List[str]], tokenizer: AutoTokenizer, generation_config: dict, device: str):
    if isinstance(prompt, str):
        prompt = [prompt]
    pre_transformed_conversation = format_conversation(user_prompts = prompt, system_prompt=system_prompt)
    full_prompts = tokenizer.apply_chat_template(pre_transformed_conversation,
                                                 prefiller=prefiller,
                                                 tokenize=False,
                                                 add_generation_prompt=True # adds the <｜Assistant｜> token at the end
                                                 )
    inputs = full_tokenize(prompts=full_prompts, tokenizer=tokenizer, device = device, add_special_tokens=False)
    print('prompt tokens shape:', inputs['input_ids'].shape)
    all_model_outputs = model.generate(**inputs, **generation_config)
    decoded_responses = tokenizer.decode(all_model_outputs[0].squeeze())
    return decoded_responses, all_model_outputs