File size: 5,068 Bytes
21a5563
c8ae30a
21a5563
c48bfa8
21a5563
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c48bfa8
 
 
 
 
21a5563
 
 
 
 
c48bfa8
 
21a5563
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c48bfa8
 
21a5563
 
 
 
 
 
 
 
 
c48bfa8
 
a74faaa
c48bfa8
 
 
 
 
 
21a5563
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c48bfa8
 
 
 
21a5563
 
 
 
 
c48bfa8
 
21a5563
 
 
 
c48bfa8
 
21a5563
 
c8ae30a
 
 
 
 
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
import torch
import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
from huggingface_hub import login


choices_base_models = {
    'groloch/Llama-3.2-3B-Instruct-PromptEnhancing': 'meta-llama/Llama-3.2-3B-Instruct', 
    'groloch/gemma-2-2b-it-PromptEnhancing': 'google/gemma-2-2b-it',
    'groloch/Qwen2.5-3B-Instruct-PromptEnhancing': 'Qwen/Qwen2.5-3B-Instruct',
    'groloch/Ministral-3b-instruct-PromptEnhancing': 'ministral/Ministral-3b-instruct'
}

choices_gen_token = {
    'groloch/Llama-3.2-3B-Instruct-PromptEnhancing': 'assistant', 
    'groloch/gemma-2-2b-it-PromptEnhancing': 'model',
    'groloch/Qwen2.5-3B-Instruct-PromptEnhancing': 'assistant',
    'groloch/Ministral-3b-instruct-PromptEnhancing': 'ministral/Ministral-3b-instruct'
}

gated_models = [
    'groloch/Llama-3.2-3B-Instruct-PromptEnhancing',
    'groloch/gemma-2-2b-it-PromptEnhancing'
]

previous_choice = ''

model = None
tokenizer = None

logged_in = False


def load_model(adapter_repo_id: str):
    global model, tokenizer
    base_repo_id = choices_base_models[adapter_repo_id]
    
    tokenizer = AutoTokenizer.from_pretrained(base_repo_id)
    model = AutoModelForCausalLM.from_pretrained(base_repo_id, torch_dtype=torch.bfloat16)
    
    model.load_adapter(adapter_repo_id)

def generate(prompt_to_enhance: str, 
             choice: str,
             max_tokens: float,
             temperature: float, 
             top_p: float, 
             repetition_penalty: float,
             access_token: str
             ):
    if prompt_to_enhance is None or prompt_to_enhance == '':
        raise gr.Error('Please enter a prompt')
    global previous_choice
    
    if choice != previous_choice:
        previous_choice = choice
        load_model(choice)
        
    if choice in gated_models and access_token == '':
        raise gr.Error(f'Please enter your access token (in Additional inputs) if youre using one of the following \
            models: {", ".join(gated_models)}. Make sure you have access to those models.')
        
    global logged_in
    if not logged_in and choice in gated_models:
        login(access_token)
        logged_in = True
        
    chat = [
        {'role' : 'user', 'content': prompt_to_enhance}
    ]

    prompt = tokenizer.apply_chat_template(chat, 
                                        tokenize=False, 
                                        add_generation_prompt=True,
                                        return_tensors='pt')

    encoding = tokenizer(prompt, return_tensors="pt")

    generation_config = model.generation_config
    generation_config.do_sample = True
    generation_config.max_new_tokens = int(max_tokens)
    generation_config.temperature = float(temperature)
    generation_config.top_p = float(top_p)
    generation_config.num_return_sequences = 1
    generation_config.pad_token_id = tokenizer.eos_token_id
    generation_config.eos_token_id = tokenizer.eos_token_id
    generation_config.repetition_penalty = float(repetition_penalty)

    with torch.inference_mode():
        outputs = model.generate(
            input_ids=encoding.input_ids,
            attention_mask=encoding.attention_mask,
            generation_config=generation_config
        )
    
    return tokenizer.decode(outputs[0], skip_special_tokens=True).split(choices_gen_token[choice])[-1]


#
# Inputs
#
model_choice = gr.Dropdown(
    label='Model choice',
    choices=['groloch/Llama-3.2-3B-Instruct-PromptEnhancing', 
             'groloch/gemma-2-2b-it-PromptEnhancing',
             'groloch/Qwen2.5-3B-Instruct-PromptEnhancing',
             'groloch/Ministral-3b-instruct-PromptEnhancing'
             ],
    value='groloch/Llama-3.2-3B-Instruct-PromptEnhancing'
)
input_prompt = gr.Text(
    label='Prompt to enhance'
)

#
# Additional inputs
#
input_max_tokens = gr.Number(
    label='Max generated tokens',
    value=64,
    minimum=16,
    maximum=128
)
input_temperature = gr.Number(
    label='Temperature',
    value=0.3,
    minimum=0.0,
    maximum=1.5,
    step=0.05
)
input_top_p = gr.Number(
    label='Top p',
    value=0.9,
    minimum=0.0,
    maximum=1.0,
    step=0.05
)
input_repetition_penalty = gr.Number(
    label='Repetition penalty',
    value=2.0,
    minimum=0.0,
    maximum=5.0,
    step=0.1
)
input_access_token = gr.Text(
    label='Access token for gated models',
    value=''
)

demo = gr.Interface(
    generate,
    title='Prompt Enhancing Playground',
    description='This space is a tool to compare the different prompt enhancing model I have finetuned. \
            Feel free to experiment as you want ! \n\
            If you want to use this locally, you can download the gpu version (see in files)',
    inputs=[input_prompt, model_choice],
    additional_inputs=[input_max_tokens, 
                       input_temperature, 
                       input_top_p, 
                       input_repetition_penalty,
                       input_access_token
                       ],
    outputs=['text']
)


if __name__ == "__main__":
    demo.launch()