File size: 6,979 Bytes
2ca0c5e
 
40912b5
 
 
 
 
 
 
 
 
 
 
2ca0c5e
 
80e88ee
2ca0c5e
 
 
 
 
 
 
 
 
 
 
 
 
 
a2b7758
2ca0c5e
 
 
 
 
 
285adea
2ca0c5e
b637f0b
2ca0c5e
b637f0b
2ca0c5e
 
 
a2b7758
2ca0c5e
 
 
 
 
 
 
 
 
 
 
 
 
 
80e88ee
2ca0c5e
 
 
 
 
 
80e88ee
2ca0c5e
 
 
80e88ee
2ca0c5e
 
80e88ee
2ca0c5e
 
 
 
80e88ee
2ca0c5e
 
 
80e88ee
2ca0c5e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
40912b5
2ca0c5e
 
 
 
 
 
 
 
 
 
 
 
 
d78f59f
2ca0c5e
 
 
 
 
 
d78f59f
2ca0c5e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d78f59f
2ca0c5e
40912b5
2ca0c5e
 
 
 
 
a2b7758
d78f59f
 
2ca0c5e
 
 
 
 
 
 
d78f59f
 
2ca0c5e
 
 
 
 
 
 
 
 
 
 
80e88ee
2ca0c5e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
80e88ee
 
 
 
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
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
import os

# Questions for Gradio
# - Chat share button is enabled by default but thrown an error when clicked.
# - How to add local images in HTML? (https://github.com/gradio-app/gradio/issues/884)
# - How to allow Chatbot to fill the vertical space? (https://github.com/gradio-app/gradio/issues/4001)
# TODO:
# - Add the 1MB models, keras/gemma_1.1_instruct_7b_en
# - Add retry button, for each model individually
# - Add ability to route a message to a single model only.
# - log_applied_layout_map: make it work for Llama3CausalLM and LlamaCausalLM (vicuna)
# - display context length

os.environ["KERAS_BACKEND"] = "jax"

import gradio as gr
from gradio import ChatMessage
import keras_hub

from chatstate import ChatState
from models import (
    model_presets,
    load_model,
    model_labels,
    preset_to_website_url,
    get_appropriate_chat_template,
)

model_labels_list = list(model_labels)

# load and warm up (compile) all the models
models = []
for preset in model_presets:
    model = load_model(preset)
    chat_template = get_appropriate_chat_template(preset)
    chat_state = ChatState(model, "", chat_template)
    prompt, response = chat_state.send_message("Hello")
    print("model " + preset + " loaded and initialized.")
    print("The model responded: " + response)
    models.append(model)

# For local debugging
# model = keras_hub.models.Llama3CausalLM.from_preset(
#     "hf://meta-llama/Llama-3.2-1B-Instruct", dtype="bfloat16"
# )
# models = [model, model, model, model, model]


def chat_turn_assistant_1(
    model,
    message,
    history,
    system_message,
    preset,
    # max_tokens,
    # temperature,
    # top_p,
):
    chat_template = get_appropriate_chat_template(preset)
    chat_state = ChatState(model, system_message, chat_template)

    for msg in history:
        msg = ChatMessage(**msg)
        if msg.role == "user":
            chat_state.add_to_history_as_user(msg.content)
        elif msg.role == "assistant":
            chat_state.add_to_history_as_model(msg.content)

    prompt, response = chat_state.send_message(message)
    history.append(ChatMessage(role="assistant", content=response))
    return history


def chat_turn_assistant(
    message,
    sel1,
    history1,
    sel2,
    history2,
    system_message,
    # max_tokens,
    # temperature,
    # top_p,
):
    history1 = chat_turn_assistant_1(
        models[sel1], message, history1, system_message, model_presets[sel1]
    )
    history2 = chat_turn_assistant_1(
        models[sel2], message, history2, system_message, model_presets[sel2]
    )
    return "", history1, history2


def chat_turn_user_1(message, history):
    history.append(ChatMessage(role="user", content=message))
    return history


def chat_turn_user(message, history1, history2):
    history1 = chat_turn_user_1(message, history1)
    history2 = chat_turn_user_1(message, history2)
    return "", history1, history2


def bot_icon_select(model_name):
    if "gemma" in model_name:
        return "img/gemma.png"
    elif "llama" in model_name:
        return "img/meta.png"
    elif "vicuna" in model_name:
        return "img/vicuna.png"
    elif "mistral" in model_name:
        return "img/mistral.png"
    # default
    return "img/bot.png"


def instantiate_chatbots(sel1, sel2):
    model_name1 = model_presets[sel1]
    chatbot1 = gr.Chatbot(
        type="messages",
        show_label=False,
        show_share_button=False,
        avatar_images=("img/usr.png", bot_icon_select(model_name1)),
    )
    model_name2 = model_presets[sel2]
    chatbot2 = gr.Chatbot(
        type="messages",
        show_label=False,
        show_share_button=False,
        avatar_images=("img/usr.png", bot_icon_select(model_name2)),
    )
    return chatbot1, chatbot2


def instantiate_select_boxes(sel1, sel2, model_labels):
    sel1 = gr.Dropdown(
        choices=[(name, i) for i, name in enumerate(model_labels)],
        show_label=False,
        info="<span style='color:black'>Selected model 1:</span> "
        + "<a href='"
        + preset_to_website_url(model_presets[sel1])
        + "'>"
        + preset_to_website_url(model_presets[sel1])
        + "</a>",
        value=sel1,
    )
    sel2 = gr.Dropdown(
        choices=[(name, i) for i, name in enumerate(model_labels)],
        show_label=False,
        info="<span style='color:black'>Selected model 2:</span> "
        + "<a href='"
        + preset_to_website_url(model_presets[sel2])
        + "'>"
        + preset_to_website_url(model_presets[sel2])
        + "</a>",
        value=sel2,
    )
    return sel1, sel2


def instantiate_chatbots_and_select_boxes(sel1, sel2, model_labels):
    chatbot1, chatbot2 = instantiate_chatbots(sel1, sel2)
    sel1, sel2 = instantiate_select_boxes(sel1, sel2, model_labels)
    return sel1, chatbot1, sel2, chatbot2


with gr.Blocks(fill_width=True, title="Keras demo") as demo:

    with gr.Row():
        gr.Image(
            "img/keras_logo_k.png",
            width=80,
            height=80,
            min_width=80,
            show_label=False,
            show_download_button=False,
            show_fullscreen_button=False,
            show_share_button=False,
            interactive=False,
            scale=0,
            container=False,
        )
        gr.HTML(
            "<H2> Battle of the Keras chatbots on TPU</H2>"
            + "All the models are loaded into the TPU memory. "
            + "You can call any of them and compare their answers. <br/>"
            + "The entire chat history is fed to the models at every submission. "
            + "This demo is runnig on a Google TPU v5e 2x4 (8 cores).",
        )
    with gr.Row():
        sel1, sel2 = instantiate_select_boxes(0, 1, model_labels_list)

    with gr.Row():
        chatbot1, chatbot2 = instantiate_chatbots(sel1.value, sel2.value)

    msg = gr.Textbox(label="Your message:", submit_btn=True)

    with gr.Row():
        gr.ClearButton([msg, chatbot1, chatbot2])
        with gr.Accordion("Additional settings", open=False):
            system_message = gr.Textbox(
                label="Sytem prompt",
                value="You are a helpful assistant and your name is Eliza.",
            )

    sel1.select(
        lambda sel1, sel2: instantiate_chatbots_and_select_boxes(
            sel1, sel2, model_labels_list
        ),
        inputs=[sel1, sel2],
        outputs=[sel1, chatbot1, sel2, chatbot2],
    )

    sel2.select(
        lambda sel1, sel2: instantiate_chatbots_and_select_boxes(
            sel1, sel2, model_labels_list
        ),
        inputs=[sel1, sel2],
        outputs=[sel1, chatbot1, sel2, chatbot2],
    )

    msg.submit(
        chat_turn_user,
        inputs=[msg, chatbot1, chatbot2],
        outputs=[msg, chatbot1, chatbot2],
    ).then(
        chat_turn_assistant,
        [msg, sel1, chatbot1, sel2, chatbot2, system_message],
        outputs=[msg, chatbot1, chatbot2],
    )


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