File size: 10,522 Bytes
e5725a3
349c343
 
a1430d2
 
 
e3a164f
 
 
349c343
a1430d2
 
 
 
 
 
 
 
 
 
 
1d804c6
 
 
 
c651799
 
 
 
7700f9c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a1430d2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
349c343
86c245c
349c343
a1430d2
 
 
349c343
 
 
 
 
 
 
 
 
7dca0e3
349c343
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
76fdaaa
 
 
 
 
 
 
 
349c343
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a1430d2
 
324f5ff
a1430d2
 
 
6f7615d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
349c343
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
import os
import gradio as gr

from llama2 import GradioLLaMA2ChatPPManager
from llama2 import gen_text

from styles import MODEL_SELECTION_CSS
from js import GET_LOCAL_STORAGE, UPDATE_LEFT_BTNS_STATE, UPDATE_PLACEHOLDERS
from templates import templates

from pingpong.context import CtxLastWindowStrategy

TOKEN = os.getenv('HF_TOKEN')
MODEL_ID = 'meta-llama/Llama-2-70b-chat-hf'

def build_prompts(ppmanager, global_context, win_size=3):
    dummy_ppm = copy.deepcopy(ppmanager)
    dummy_ppm.ctx = global_context
    lws = CtxLastWindowStrategy(win_size)
    return lws(dummy_ppm)

ex_file = open("examples.txt", "r")
examples = ex_file.read().split("\n")
ex_btns = []

chl_file = open("channels.txt", "r")
channels = chl_file.read().split("\n")
channel_btns = []

def fill_up_placeholders(txt):
    placeholders = get_placeholders(txt)
    highlighted_txt = txt

    return (
        gr.update(
            visible=True,
            value=highlighted_txt
        ),
        gr.update(
            visible=True if len(placeholders) >= 1 else False,
            placeholder=placeholders[0] if len(placeholders) >= 1 else ""
        ),
        gr.update(
            visible=True if len(placeholders) >= 2 else False,
            placeholder=placeholders[1] if len(placeholders) >= 2 else ""
        ),
        gr.update(
            visible=True if len(placeholders) >= 3 else False,
            placeholder=placeholders[2] if len(placeholders) >= 3 else ""
        ),
        "" if len(placeholders) >= 1 else txt
    )

async def chat_stream(idx, local_data, instruction_txtbox, chat_state):
    res = [
      chat_state["ppmanager_type"].from_json(json.dumps(ppm))
      for ppm in local_data
    ]

    ppm = res[idx]
    ppm.add_pingpong(
        PingPong(instruction_txtbox, "")
    )
    prompt = build_prompts(ppm, "global context", 3)
    for result in await gen_text(prompt, hf_model=MODEL_ID, hf_token=TOKEN):
        ppm.append_pong(result)
        yield ppm.build_uis(), str(res)
        

with gr.Blocks(css=MODEL_SELECTION_CSS, theme='gradio/soft') as demo:
    with gr.Column() as chat_view:
        idx = gr.State(0)
        chat_state = gr.State({
            "ppmanager_type": GradioLLaMA2ChatPPManager
        })
        local_data = gr.JSON({}, visible=False)

        with gr.Row():
            with gr.Column(scale=1, min_width=180):
                gr.Markdown("GradioChat", elem_id="left-top")

                with gr.Column(elem_id="left-pane"):
                    chat_back_btn = gr.Button("Back", elem_id="chat-back-btn")
                    
                    with gr.Accordion("Histories", elem_id="chat-history-accordion", open=True):
                        channel_btns.append(gr.Button(channels[0], elem_classes=["custom-btn-highlight"]))

                        for channel in channels[1:]:
                            channel_btns.append(gr.Button(channel, elem_classes=["custom-btn"]))

            with gr.Column(scale=8, elem_id="right-pane"):
                with gr.Column(
                    elem_id="initial-popup", visible=False
                ) as example_block:
                    with gr.Row(scale=1):
                        with gr.Column(elem_id="initial-popup-left-pane"):
                            gr.Markdown("GradioChat", elem_id="initial-popup-title")
                            gr.Markdown("Making the community's best AI chat models available to everyone.")
                        with gr.Column(elem_id="initial-popup-right-pane"):
                            gr.Markdown("Chat UI is now open sourced on Hugging Face Hub")
                            gr.Markdown("check out the [↗ repository](https://huggingface.co/spaces/chansung/test-multi-conv)")

                    with gr.Column(scale=1):
                        gr.Markdown("Examples")
                        with gr.Row():
                            for example in examples:
                                ex_btns.append(gr.Button(example, elem_classes=["example-btn"]))

                with gr.Column(elem_id="aux-btns-popup", visible=True):
                    with gr.Row():
                        stop = gr.Button("Stop", elem_classes=["aux-btn"])
                        regenerate = gr.Button("Regen", interactive=False, elem_classes=["aux-btn"])
                        clean = gr.Button("Clean", elem_classes=["aux-btn"])

                with gr.Accordion("Context Inspector", elem_id="aux-viewer", open=False):
                    context_inspector = gr.Textbox(
                        "",
                        elem_id="aux-viewer-inspector",
                        label="",
                        lines=30,
                        max_lines=50,
                    )                        
                        
                chatbot = gr.Chatbot(elem_id='chatbot')
                instruction_txtbox = gr.Textbox(placeholder="Ask anything", label="", elem_id="prompt-txt")

        with gr.Accordion("Example Templates", open=False):
            template_txt = gr.Textbox(visible=False)
            template_md = gr.Markdown(label="Chosen Template", visible=False, elem_classes="template-txt")

            with gr.Row():
                placeholder_txt1 = gr.Textbox(label="placeholder #1", visible=False, interactive=True)
                placeholder_txt2 = gr.Textbox(label="placeholder #2", visible=False, interactive=True)
                placeholder_txt3 = gr.Textbox(label="placeholder #3", visible=False, interactive=True)

            for template in templates:
                with gr.Tab(template['title']):
                    gr.Examples(
                        template['template'],
                        inputs=[template_txt],
                        outputs=[template_md, placeholder_txt1, placeholder_txt2, placeholder_txt3, instruction_txtbox],
                        run_on_click=True,
                        fn=fill_up_placeholders,          
                    )

        with gr.Accordion("Control Panel", open=False) as control_panel:
            with gr.Column():
                with gr.Column():
                    gr.Markdown("#### Global context")
                    with gr.Accordion("global context will persist during conversation, and it is placed at the top of the prompt", open=False):
                        global_context = gr.Textbox(
                            "global context",
                            lines=5,
                            max_lines=10,
                            interactive=True,
                            elem_id="global-context"
                        )
                    
                    # gr.Markdown("#### Internet search")
                    # with gr.Row():
                    #     internet_option = gr.Radio(choices=["on", "off"], value="off", label="mode")
                    #     serper_api_key = gr.Textbox(
                    #         value= "" if args.serper_api_key is None else args.serper_api_key,
                    #         placeholder="Get one by visiting serper.dev", 
                    #         label="Serper api key"
                    #     )
                    
                    gr.Markdown("#### GenConfig for **response** text generation")
                    with gr.Row():
                        res_temp = gr.Slider(0.0, 2.0, 0, step=0.1, label="temp", interactive=True)
                        res_topp = gr.Slider(0.0, 2.0, 0, step=0.1, label="top_p", interactive=True)
                        res_topk = gr.Slider(20, 1000, 0, step=1, label="top_k", interactive=True)
                        res_rpen = gr.Slider(0.0, 2.0, 0, step=0.1, label="rep_penalty", interactive=True)
                        res_mnts = gr.Slider(64, 8192, 0, step=1, label="new_tokens", interactive=True)                            
                        res_beams = gr.Slider(1, 4, 0, step=1, label="beams")
                        res_cache = gr.Radio([True, False], value=0, label="cache", interactive=True)
                        res_sample = gr.Radio([True, False], value=0, label="sample", interactive=True)
                        res_eosid = gr.Number(value=0, visible=False, precision=0)
                        res_padid = gr.Number(value=0, visible=False, precision=0)

                with gr.Column(visible=False):
                    gr.Markdown("#### GenConfig for **summary** text generation")
                    with gr.Row():
                        sum_temp = gr.Slider(0.0, 2.0, 0, step=0.1, label="temp", interactive=True)
                        sum_topp = gr.Slider(0.0, 2.0, 0, step=0.1, label="top_p", interactive=True)
                        sum_topk = gr.Slider(20, 1000, 0, step=1, label="top_k", interactive=True)
                        sum_rpen = gr.Slider(0.0, 2.0, 0, step=0.1, label="rep_penalty", interactive=True)
                        sum_mnts = gr.Slider(64, 8192, 0, step=1, label="new_tokens", interactive=True)
                        sum_beams = gr.Slider(1, 8, 0, step=1, label="beams", interactive=True)
                        sum_cache = gr.Radio([True, False], value=0, label="cache", interactive=True)
                        sum_sample = gr.Radio([True, False], value=0, label="sample", interactive=True)
                        sum_eosid = gr.Number(value=0, visible=False, precision=0)
                        sum_padid = gr.Number(value=0, visible=False, precision=0)

                with gr.Column():
                    gr.Markdown("#### Context managements")
                    with gr.Row():
                        ctx_num_lconv = gr.Slider(2, 10, 3, step=1, label="number of recent talks to keep", interactive=True)
                        ctx_sum_prompt = gr.Textbox(
                            "summarize our conversations. what have we discussed about so far?",
                            label="design a prompt to summarize the conversations",
                            visible=False
                        )

    instruction_txtbox.submit(
        chat_stream,
        [idx, local_data, instruction_txtbox, chat_state],
        [chatbot, local_data]
    )

    for btn in channel_btns:
        btn.click(
            set_chatbot,
            [btn, local_data, chat_state],
            [chatbot, idx, example_block, regenerate]
        ).then(
            None, btn, None, 
            _js=UPDATE_LEFT_BTNS_STATE        
        )

    demo.load(
        None,
        inputs=None,
        outputs=[chatbot, local_data],
        _js=GET_LOCAL_STORAGE,
    )     

demo.launch()