File size: 7,705 Bytes
1118f77
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
beae39b
 
1118f77
 
 
 
 
 
 
 
beae39b
1118f77
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
beae39b
1118f77
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
from transformers import AutoTokenizer, AutoModelForCausalLM
from unidecode import unidecode
from collections import Counter
import torch
import os
import gradio as gr
import numpy as np
import re
import string

from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("osiria/primo")
model = AutoModelForCausalLM.from_pretrained("osiria/primo")
model = PeftModel.from_pretrained(model, "osiria/primo")

class Prime:
    
    def __init__(self, tokenizer, model):
        self.tokenizer = tokenizer
        self.model = model
        
    def _check_sublist(self, lst, sub_lst, sep = " "):
        
        l_type = type(lst[0])
        lst = sep.join(list(map(str, lst)))
        sub_lst = sep.join(list(map(str, sub_lst)))
        
        return sub_lst in lst
    
    def _exclude_sublist(self, lst, sub_lst, sep = " "):
        
        l_type = type(lst[0])
        lst = sep.join(list(map(str, lst)))
        sub_lst = sep.join(list(map(str, sub_lst)))
        lst = re.sub("\s+", " ", lst.replace(sub_lst, "")).strip().split(sep)
        lst = list(map(l_type, lst))
        
        return lst
        
    def generate(self, prompt, message = "", sep = " [AI]", max_tokens = 100, excluded = [[40, 19]], 
                 lookback = 5, resample_tokens = [27793], replace_tokens = {11302: 23318}, 
                 stop_tokens = [239], 
                 sample = False, 
                 top_k = 5):
        
        if message:
            prompt = message + ". " + prompt
        prompt = prompt.replace("β€œ", '"').replace("”", '"').replace("’", "'")
        if not sample:
            top_k = 2
        tokens = tokenizer.encode("[HUMAN] " + prompt + sep)
        tokens_generated = []
        checkpoint = 0
        while tokens[-1] not in stop_tokens and len(tokens_generated) < max_tokens:
            output = model.forward(input_ids=torch.tensor([tokens]).to(device)).logits[0,-1]
            output = torch.softmax(output, dim = 0)
            candidates = torch.topk(output, k = top_k)
            if sample:
                indices = candidates.indices
                scores = candidates.values
                next_token = indices[torch.multinomial(scores, 1)[0].item()]
            else:
                next_token = candidates.indices[0]
            next_token = next_token.item()
            sub_tokens = tokens_generated[-lookback:] + [next_token]
            if next_token in resample_tokens:
                next_token = candidates.indices[1]
                next_token = next_token.item()
            if len(tokens_generated) >= (lookback + 1) and next_token in tokens_generated[-2:]:
                next_token = candidates.indices[1]
                next_token = next_token.item()
            elif len(tokens_generated) >= lookback and self._check_sublist(tokens_generated, sub_tokens):
                if checkpoint:
                    tokens = tokens[:checkpoint]
                    break
                else:
                    next_token = candidates.indices[1]
                    next_token = next_token.item()
                    sample = True
            if next_token in replace_tokens:
                next_token = replace_tokens[next_token]
            tokens = tokens + [next_token]
            tokens_generated = tokens_generated + [next_token]
            if next_token == 5:
                checkpoint = len(tokens)
        for ex_lst in excluded:
            tokens = self._exclude_sublist(tokens, ex_lst)
        output = tokenizer.decode(tokens, skip_special_tokens=True)
        output = output.split(sep)[-1].strip()
        output = output[0].upper() + output[1:]
        if output[-1] == tokenizer.decode(stop_tokens[0]):
            output = output[:-1]
        if len(re.findall("\d\.", output)) > 1:
            output = re.sub("\d\.", "<br>β€’", output)
        return output

model.eval()
device = torch.device("cuda")
prime = Prime(tokenizer = tokenizer, model = model)

def process_input(user_input, max_tokens, sample, top_k, message):
    return prime.generate(prompt = user_input, message = message, 
                          max_tokens = max_tokens, sample = sample,
                          top_k = top_k)


header = '''--------------------------------------------------------------------------------------------------
<style>
.vertical-text {
    writing-mode: vertical-lr;
    text-orientation: upright;
    background-color:red;
}
</style>
<center>
<body>
<span class="vertical-text" style="background-color:lightgreen;border-radius: 3px;padding: 3px;"> </span>
<span class="vertical-text" style="background-color:orange;border-radius: 3px;padding: 3px;">  </span>
<span class="vertical-text" style="background-color:lightblue;border-radius: 3px;padding: 3px;">     </span>
<span class="vertical-text" style="background-color:tomato;border-radius: 3px;padding: 3px;">     </span>
<span class="vertical-text" style="background-color:lightgrey;border-radius: 3px;padding: 3px;">  </span>
<span class="vertical-text" style="background-color:#CF9FFF;border-radius: 3px;padding: 3px;"> </span>
</body>
</center>
<br>
<center><img src="file/prime.png" width="100"></center>
'''

import gradio as gr
import random
import time

with gr.Blocks(title="primo", css="footer {visibility: hidden}", theme=gr.themes.Default(text_size="md", spacing_size="md")) as interface:
    gr.Markdown(header)
    with gr.Row():
        with gr.Column(scale=1):
            gr.Markdown("<b>options</b>")
            max_tokens = gr.Slider(1, 250, value=150, label="max tokens", info="choose a limit between 1 and 250")
            sample = gr.Checkbox(label="sampling")
            top_k = gr.Slider(1, 5, value=1, label="creativity", info="choose a level between 1 and 5")
            message = gr.Textbox(label="system message", value = "")
            clear = gr.Button("clear chat")
        with gr.Column(scale=8):
            chatbot = gr.Chatbot(label = "prime").style(height=600)
            msg = gr.Textbox(label = "query")

            def user(user_message, history):
                return gr.update(value="", interactive=False), history + [[user_message, None]]

            def bot(history, message, max_tokens, sample, top_k):
                bot_message = process_input(history[-1][0], message = message, max_tokens = max_tokens,
                                            sample = sample, top_k = top_k)
                history[-1][1] = ""
                for character in bot_message:
                    history[-1][1] += character
                    time.sleep(0.05)
                    yield history

            response = msg.submit(user, [msg, chatbot], [msg, chatbot], queue=False).then(
                bot, [chatbot, message, max_tokens, sample, top_k], chatbot
            )
            response.then(lambda: gr.update(interactive=True), None, [msg], queue=False)
            clear.click(lambda: None, None, chatbot, queue=False)
        with gr.Column(scale=1):
            gr.Markdown("<b>warning</b>")
            gr.Markdown("the model might behave erratically when presented with prompts which are too far away from its pre-training or fine-tuning and, because of the probabilistic nature of its generation mechanism, it might occasionally produce biased or offensive content with respect to gender, race, ideologies, and political or religious beliefs<br><br>these limitations imply that the model and its outputs should be used with caution, and should not be involved in situations that require the generated text to be fair or true")

interface.queue()
interface.launch()