File size: 6,592 Bytes
b874271
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
import os
import pickle as pkl
from pathlib import Path
from threading import Thread
from typing import List, Optional, Tuple, Iterator

import gradio as gr
import numpy as np
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer


MAX_MAX_NEW_TOKENS = 2048
DEFAULT_MAX_NEW_TOKENS = 1024
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))

DESCRIPTION = """\
# Llama-2 7B Chat with Streamable Semantic Uncertainty Probe
This Space demonstrates the Llama-2-7b-chat model with an added semantic uncertainty probe. 
The highlighted text shows the model's uncertainty in real-time, with more intense yellow indicating higher uncertainty.
"""

if torch.cuda.is_available():
    model_id = "meta-llama/Llama-2-7b-chat-hf"
    # TODO load the full model?
    model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", load_in_8bit=True)
    tokenizer = AutoTokenizer.from_pretrained(model_id)
    tokenizer.use_default_system_prompt = False

    # load the probe data
    # TODO load accuracy and SE probe and compare in different tabs
    with open("./model/20240625-131035_demo.pkl", "rb") as f:
        probe_data = pkl.load(f)
    # take the NQ open one
    probe_data = probe_data[-2]
    model = probe_data['t_bmodel']
    layer_range = probe_data['sep_layer_range']
    acc_model = probe_data['t_amodel']
    acc_layer_range = probe_data['ap_layer_range']

def generate(
    message: str,
    chat_history: List[Tuple[str, str]],
    system_prompt: str,
    max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,
    temperature: float = 0.6,
    top_p: float = 0.9,
    top_k: int = 50,
    repetition_penalty: float = 1.2,
) -> Iterator[str]:
    conversation = []
    if system_prompt:
        conversation.append({"role": "system", "content": system_prompt})
    for user, assistant in chat_history:
        conversation.extend([{"role": "user", "content": user}, {"role": "assistant", "content": assistant}])
    conversation.append({"role": "user", "content": message})

    input_ids = tokenizer.apply_chat_template(conversation, return_tensors="pt")
    if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
        input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
        gr.Warning(f"Trimmed input from conversation as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.")
    input_ids = input_ids.to(model.device)

    streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
    generation_kwargs = dict(
        input_ids=input_ids,
        max_new_tokens=max_new_tokens,
        do_sample=True,
        top_p=top_p,
        top_k=top_k,
        temperature=temperature,
        repetition_penalty=repetition_penalty,
        streamer=streamer,
        output_hidden_states=True,
        return_dict_in_generate=True,
    )

    thread = Thread(target=model.generate, kwargs=generation_kwargs)
    thread.start()

    generated_text = ""
    highlighted_text = ""
    for output in streamer:
        print(output)
        generated_text += output

        yield generated_text
    
    # this is doing it twice... just do autoregressive generation instead
    for new_text in streamer:
        generated_text += new_text
        current_input_ids = tokenizer.encode(generated_text, return_tensors="pt").to(model.device)
        
        with torch.no_grad():
            outputs = model(current_input_ids, output_hidden_states=True)
            hidden = outputs.hidden_states
            # Stack second last token embeddings from all layers 
            # if len(hidden) == 1:  # FIX: runtime error for mistral-7b on bioasq
            #     sec_last_input = hidden[0]
            # elif ((n_generated - 2) >= len(hidden)):
            #     sec_last_input = hidden[-2]
            # else:
            #     sec_last_input = hidden[n_generated - 2]
            last_hidden_state = torch.stack([layer[:, -1, :].cpu() for layer in hidden[-1]]).cpu().numpy()
            # print(sec_last_token_embedding.shape)
        # last_hidden_state = outputs.hidden_states[-1][:, -1, :].cpu().numpy()
        print(last_hidden_state.shape)  
        # TODO potentially need to only compute uncertainty for the last token in sentence?

        # concatenate the hidden states from the specified layers
        probe_input = np.concatenate(last_hidden_state[layer_range], axis=1) 
        print(probe_input.shape)
        uncertainty_score = model.predict(probe_input)
        print(uncertainty_score)
        new_highlighted_text = highlight_text(new_text, uncertainty_score[0])
        print(new_highlighted_text)
        highlighted_text += new_highlighted_text

        yield highlighted_text


def highlight_text(text: str, uncertainty_score: float) -> str:
    if uncertainty_score > 0:
        html_color = "#%02X%02X%02X" % (
            255,
            int(255 * (1 - uncertainty_score)),
            int(255 * (1 - uncertainty_score)),
        )
    else:
        html_color = "#%02X%02X%02X" % (
            int(255 * (1 + uncertainty_score)),
            255,
            int(255 * (1 + uncertainty_score)),
        )
    return '<span style="background-color: {}; color: black">{}</span>'.format(
        html_color, text
    )
chat_interface = gr.ChatInterface(
    fn=generate,
    additional_inputs=[
        gr.Textbox(label="System prompt", lines=6),
        gr.Slider(
            label="Max new tokens",
            minimum=1,
            maximum=MAX_MAX_NEW_TOKENS,
            step=1,
            value=DEFAULT_MAX_NEW_TOKENS,
        ),
        gr.Slider(
            label="Temperature",
            minimum=0.1,
            maximum=4.0,
            step=0.1,
            value=0.6,
        ),
        gr.Slider(
            label="Top-p (nucleus sampling)",
            minimum=0.05,
            maximum=1.0,
            step=0.05,
            value=0.9,
        ),
        gr.Slider(
            label="Top-k",
            minimum=1,
            maximum=1000,
            step=1,
            value=50,
        ),
        gr.Slider(
            label="Repetition penalty",
            minimum=1.0,
            maximum=2.0,
            step=0.05,
            value=1.2,
        ),
    ],
    stop_btn=None,
    examples=[
        ["What is the capital of France?"],
        ["Explain the theory of relativity in simple terms."],
        ["Write a short poem about artificial intelligence."]
    ],
    title="Llama-2 7B Chat with Streamable Semantic Uncertainty Probe",
    description=DESCRIPTION,
)

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