File size: 17,096 Bytes
e5725a3
ae17679
a3a9516
9ca5709
349c343
 
a1430d2
 
 
e3a164f
 
 
13272b6
349c343
fcc28fc
a1430d2
 
 
 
 
 
 
 
 
 
 
1d804c6
 
 
 
c651799
 
 
 
ae17679
 
 
 
 
 
7700f9c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1df330b
b3bde95
 
 
 
 
 
 
 
 
73d048f
 
 
b3bde95
 
73d048f
b3bde95
 
 
 
 
 
 
 
 
 
 
 
 
55d784d
af7cd47
55d784d
73d048f
 
 
 
 
 
 
 
 
 
 
 
 
629e8c6
 
 
 
a1430d2
 
 
 
 
 
 
 
 
629e8c6
 
 
 
 
 
 
 
 
 
 
 
a1430d2
eba85a0
 
a4f9cd0
0512804
3ffe660
 
 
 
 
 
 
 
 
0512804
 
 
 
 
 
a1430d2
c1d72c2
 
 
de4d793
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
349c343
86c245c
349c343
a1430d2
 
 
349c343
 
71e0dac
 
 
 
 
 
 
9e14ade
93e5809
9e14ade
 
 
 
 
 
 
71e0dac
349c343
 
 
 
 
 
 
7dca0e3
349c343
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eba85a0
349c343
 
 
 
 
 
 
 
 
 
 
 
8248004
349c343
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
629e8c6
349c343
13272b6
349c343
 
 
 
 
 
76fdaaa
 
 
 
 
 
 
 
349c343
 
 
629e8c6
 
 
 
d7c22fb
629e8c6
349c343
 
 
 
 
73d048f
6917255
 
 
 
 
 
 
a1430d2
629e8c6
 
eba85a0
4495d82
23d7269
 
 
a1430d2
b3bde95
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
23d7269
 
 
 
 
4495d82
23d7269
 
 
73d048f
4495d82
 
 
 
73d048f
6f7615d
 
 
 
 
 
 
 
 
 
fcc28fc
 
 
 
 
 
73d048f
 
 
 
 
 
 
 
 
 
fcc28fc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
73d048f
6f7615d
 
 
 
 
 
 
93e5809
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
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
import os
import re
import json
import copy
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 constants import DEFAULT_GLOBAL_CTX

from pingpong import PingPong
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 get_placeholders(text):
    """Returns all substrings in between <placeholder> and </placeholder>."""
    pattern = r"\[([^\]]*)\]"
    matches = re.findall(pattern, text)
    return matches

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 rollback_last(
    idx, local_data, chat_state,
    global_context, res_temp, res_topk, res_rpen, res_mnts, res_sample, ctx_num_lconv
):
    res = [
      chat_state["ppmanager_type"].from_json(json.dumps(ppm))
      for ppm in local_data
    ]

    ppm = res[idx]
    last_user_message = res[idx].pingpongs[-1].ping
    res[idx].pingpongs = res[idx].pingpongs[:-1]
    
    ppm.add_pingpong(
        PingPong(last_user_message, "")
    )
    prompt = build_prompts(ppm, global_context, ctx_num_lconv)
    async for result in gen_text(
        prompt, hf_model=MODEL_ID, hf_token=TOKEN,
        parameters={
            'max_new_tokens': res_mnts,
            'do_sample': res_sample,
            'return_full_text': False,
            'temperature': res_temp,
            'top_k': res_topk,
            'repetition_penalty': res_rpen           
        }
    ):
        ppm.append_pong(result)
        yield prompt, ppm.build_uis(), str(res), gr.update(interactive=False)
        
    yield prompt, ppm.build_uis(), str(res), gr.update(interactive=True)

def reset_chat(idx, ld, state):
    res = [state["ppmanager_type"].from_json(json.dumps(ppm_str)) for ppm_str in ld]
    res[idx].pingpongs = []
        
    return (
        "",
        [],
        str(res),
        gr.update(visible=True),
        gr.update(interactive=False),
    )

async def chat_stream(
    idx, local_data, instruction_txtbox, chat_state,
    global_context, res_temp, res_topk, res_rpen, res_mnts, res_sample, ctx_num_lconv
):
    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, ctx_num_lconv)
    async for result in gen_text(
        prompt, hf_model=MODEL_ID, hf_token=TOKEN,
        parameters={
            'max_new_tokens': res_mnts,
            'do_sample': res_sample,
            'return_full_text': False,
            'temperature': res_temp,
            'top_k': res_topk,
            'repetition_penalty': res_rpen           
        }
    ):
        ppm.append_pong(result)
        yield "", prompt, ppm.build_uis(), str(res), gr.update(interactive=False)

    yield "", prompt, ppm.build_uis(), str(res), gr.update(interactive=True)

def channel_num(btn_title):
    choice = 0

    for idx, channel in enumerate(channels):
        if channel == btn_title:
            choice = idx

    return choice

def set_chatbot(btn, ld, state):
    choice = channel_num(btn)

    res = [state["ppmanager_type"].from_json(json.dumps(ppm_str)) for ppm_str in ld]
    empty = len(res[choice].pingpongs) == 0
    return (res[choice].build_uis(), choice, gr.update(visible=empty), gr.update(interactive=not empty))

def set_example(btn):
    return btn, gr.update(visible=False)

def get_final_template(
    txt, placeholder_txt1, placeholder_txt2, placeholder_txt3
):
    placeholders = get_placeholders(txt)
    example_prompt = txt    

    if len(placeholders) >= 1:
        if placeholder_txt1 != "":
            example_prompt = example_prompt.replace(f"[{placeholders[0]}]", placeholder_txt1)
    if len(placeholders) >= 2:
        if placeholder_txt2 != "":
            example_prompt = example_prompt.replace(f"[{placeholders[1]}]", placeholder_txt2)
    if len(placeholders) >= 3:
        if placeholder_txt3 != "":
            example_prompt = example_prompt.replace(f"[{placeholders[2]}]", placeholder_txt3)

    return (
        example_prompt,
        "",
        "",
        ""
    )

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)

        gr.Markdown("## LLaMA2 70B with Gradio Chat and Hugging Face Inference API", elem_classes=["center"])
        gr.Markdown(
            "This space demonstrates how to build feature rich chatbot UI in [Gradio](https://www.gradio.app/). Supported features "
            "include • multiple chatting channels, • chat history save/restoration, • stop generating text response, • regenerate the "
            "last conversation, • clean the chat history, • dynamic kick-starting prompt templates, • adjusting text generation parameters, "
            "• inspecting the actual prompt that the model sees. The underlying Large Language Model is the [Meta AI](https://ai.meta.com/)'s "
            "[LLaMA2-70B](https://huggingface.co/meta-llama/Llama-2-70b-chat-hf) which is hosted as [Hugging Face Inference API](https://huggingface.co/inference-api), "
            "and [Text Generation Inference](https://github.com/huggingface/text-generation-inference) is the underlying serving framework. ",
            elem_classes=["center"]
        )
        gr.Markdown(
            "***NOTE:*** If you are subscribing [PRO](https://huggingface.co/pricing#pro), you can simply duplicate this space and use your "
            "Hugging Face Access Token to run the same application. Just add `HF_TOKEN` secret with the Token value accorindg to [this guide](https://huggingface.co/docs/hub/spaces-overview#managing-secrets-and-environment-variables)",
            elem_classes=["center"]
        )
            
        
        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', label="LLaMA2-70B-Chat")
                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=True):
                        global_context = gr.Textbox(
                            DEFAULT_GLOBAL_CTX,
                            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, 1.0, step=0.1, label="temp", interactive=True)
                        res_topk = gr.Slider(20, 1000, 50, step=1, label="top_k", interactive=True)
                        res_rpen = gr.Slider(0.0, 2.0, 1.2, step=0.1, label="rep_penalty", interactive=True)
                        res_mnts = gr.Slider(64, 8192, 512, step=1, label="new_tokens", interactive=True)
                        res_sample = gr.Radio([True, False], value=True, label="sample", interactive=True)
                
                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)

    send_event = instruction_txtbox.submit(
        lambda: [
            gr.update(visible=False),
            gr.update(interactive=True)
        ],
        None,
        [example_block, regenerate]
    ).then(
        chat_stream,
        [idx, local_data, instruction_txtbox, chat_state,
         global_context, res_temp, res_topk, res_rpen, res_mnts, res_sample, ctx_num_lconv],
        [instruction_txtbox, context_inspector, chatbot, local_data, regenerate]
    ).then(
        None, local_data, None, 
        _js="(v)=>{ setStorage('local_data',v) }"
    )

    # regen_event1 = regenerate.click(
    #     rollback_last,
    #     [idx, local_data, chat_state],
    #     [instruction_txtbox, chatbot, local_data, regenerate]
    # )
    # regen_event2 = regen_event1.then(
    #     chat_stream,
    #     [idx, local_data, instruction_txtbox, chat_state,
    #      global_context, res_temp, res_topk, res_rpen, res_mnts, res_sample, ctx_num_lconv],
    #     [context_inspector, chatbot, local_data]
    # )
    # regen_event3 = regen_event2.then(
    #     lambda: gr.update(interactive=True),
    #     None,
    #     regenerate
    # )
    # regen_event4 = regen_event3.then(
    #     None, local_data, None, 
    #     _js="(v)=>{ setStorage('local_data',v) }"  
    # )

    regen_event = regenerate.click(
        rollback_last,
        [idx, local_data, chat_state,
         global_context, res_temp, res_topk, res_rpen, res_mnts, res_sample, ctx_num_lconv],
        [context_inspector, chatbot, local_data, regenerate]
    ).then(
        None, local_data, None, 
        _js="(v)=>{ setStorage('local_data',v) }"  
    )
    
    # stop.click(
    #     lambda: gr.update(interactive=True), None, regenerate,
    #     cancels=[send_event, regen_event]
    # )

    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        
        )

    for btn in ex_btns:
        btn.click(
            set_example,
            [btn],
            [instruction_txtbox, example_block]  
        )

    clean.click(
        reset_chat,
        [idx, local_data, chat_state],
        [instruction_txtbox, chatbot, local_data, example_block, regenerate]
    ).then(
        None, local_data, None, 
        _js="(v)=>{ setStorage('local_data',v) }"
    )

    
    placeholder_txt1.change(
        inputs=[template_txt, placeholder_txt1, placeholder_txt2, placeholder_txt3],
        outputs=[template_md],
        show_progress=False,
        _js=UPDATE_PLACEHOLDERS,
        fn=None
    )

    placeholder_txt2.change(
        inputs=[template_txt, placeholder_txt1, placeholder_txt2, placeholder_txt3],
        outputs=[template_md],
        show_progress=False,
        _js=UPDATE_PLACEHOLDERS,
        fn=None
    )

    placeholder_txt3.change(
        inputs=[template_txt, placeholder_txt1, placeholder_txt2, placeholder_txt3],
        outputs=[template_md],
        show_progress=False,
        _js=UPDATE_PLACEHOLDERS,
        fn=None
    )

    placeholder_txt1.submit(
        inputs=[template_txt, placeholder_txt1, placeholder_txt2, placeholder_txt3],
        outputs=[instruction_txtbox, placeholder_txt1, placeholder_txt2, placeholder_txt3],
        fn=get_final_template
    )

    placeholder_txt2.submit(
        inputs=[template_txt, placeholder_txt1, placeholder_txt2, placeholder_txt3],
        outputs=[instruction_txtbox, placeholder_txt1, placeholder_txt2, placeholder_txt3],
        fn=get_final_template
    )

    placeholder_txt3.submit(
        inputs=[template_txt, placeholder_txt1, placeholder_txt2, placeholder_txt3],
        outputs=[instruction_txtbox, placeholder_txt1, placeholder_txt2, placeholder_txt3],
        fn=get_final_template
    )    

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

demo.queue(concurrency_count=5, max_size=256).launch()