File size: 22,034 Bytes
a824a18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5df3ede
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a824a18
5df3ede
 
a824a18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0b8daad
 
c6dfdac
230a504
 
c6dfdac
 
230a504
 
 
c6dfdac
a824a18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5df3ede
 
a824a18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5df3ede
 
 
a824a18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5df3ede
 
 
a824a18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c6dfdac
a824a18
5df3ede
c6dfdac
 
5df3ede
a824a18
 
 
 
 
 
 
 
 
 
c6dfdac
 
a824a18
c6dfdac
5df3ede
a824a18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c6dfdac
 
a824a18
 
c6dfdac
a824a18
 
 
 
5df3ede
a824a18
 
 
 
5df3ede
 
 
 
 
 
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
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
import argparse
import datetime
import hashlib
import json
import os
import subprocess
import sys
import time

import gradio as gr
import requests

from llava.constants import LOGDIR
from llava.conversation import SeparatorStyle, conv_templates, default_conversation
from llava.utils import (
    build_logger,
    moderation_msg,
    server_error_msg,
    violates_moderation,
)

logger = build_logger("gradio_web_server", "gradio_web_server.log")

headers = {"User-Agent": "LLaVA Client"}

no_change_btn = gr.Button.update()
enable_btn = gr.Button.update(interactive=True)
disable_btn = gr.Button.update(interactive=False)

priority = {
    "vicuna-13b": "aaaaaaa",
    "koala-13b": "aaaaaab",
}


def get_conv_log_filename():
    t = datetime.datetime.now()
    name = os.path.join(LOGDIR, f"{t.year}-{t.month:02d}-{t.day:02d}-conv.json")
    return name


def get_model_list():
    ret = requests.post(args.controller_url + "/refresh_all_workers")
    assert ret.status_code == 200
    ret = requests.post(args.controller_url + "/list_models")
    models = ret.json()["models"]
    models.sort(key=lambda x: priority.get(x, x))
    logger.info(f"Models: {models}")
    return models


get_window_url_params = """
function() {
    const params = new URLSearchParams(window.location.search);
    url_params = Object.fromEntries(params);
    console.log(url_params);
    return url_params;
    }
"""


def load_demo(url_params, request: gr.Request):
    logger.info(f"load_demo. ip: {request.client.host}. params: {url_params}")

    dropdown_update = gr.Dropdown.update(visible=True)
    if "model" in url_params:
        model = url_params["model"]
        if model in models:
            dropdown_update = gr.Dropdown.update(value=model, visible=True)

    state = default_conversation.copy()
    return state, dropdown_update


def load_demo_refresh_model_list(request: gr.Request):
    logger.info(f"load_demo. ip: {request.client.host}")
    models = get_model_list()
    state = default_conversation.copy()

    models_downloaded = True if models else False

    model_dropdown_kwargs = {
        "choices": [],
        "value": "Downloading the models...",
        "interactive": models_downloaded,
    }

    if models_downloaded:
        model_dropdown_kwargs["choices"] = models
        model_dropdown_kwargs["value"] = models[0]

    models_dropdown_update = gr.Dropdown.update(**model_dropdown_kwargs)

    send_button_update = gr.Button.update(
        interactive=models_downloaded,
    )

    return state, models_dropdown_update, send_button_update


def vote_last_response(state, vote_type, model_selector, request: gr.Request):
    with open(get_conv_log_filename(), "a") as fout:
        data = {
            "tstamp": round(time.time(), 4),
            "type": vote_type,
            "model": model_selector,
            "state": state.dict(),
            "ip": request.client.host,
        }
        fout.write(json.dumps(data) + "\n")


def upvote_last_response(state, model_selector, request: gr.Request):
    logger.info(f"upvote. ip: {request.client.host}")
    vote_last_response(state, "upvote", model_selector, request)
    return ("",) + (disable_btn,) * 3


def downvote_last_response(state, model_selector, request: gr.Request):
    logger.info(f"downvote. ip: {request.client.host}")
    vote_last_response(state, "downvote", model_selector, request)
    return ("",) + (disable_btn,) * 3


def flag_last_response(state, model_selector, request: gr.Request):
    logger.info(f"flag. ip: {request.client.host}")
    vote_last_response(state, "flag", model_selector, request)
    return ("",) + (disable_btn,) * 3


def regenerate(state, image_process_mode, request: gr.Request):
    logger.info(f"regenerate. ip: {request.client.host}")
    state.messages[-1][-1] = None
    prev_human_msg = state.messages[-2]
    if type(prev_human_msg[1]) in (tuple, list):
        prev_human_msg[1] = (*prev_human_msg[1][:2], image_process_mode)
    state.skip_next = False
    return (state, state.to_gradio_chatbot(), "", None) + (disable_btn,) * 5


def clear_history(request: gr.Request):
    logger.info(f"clear_history. ip: {request.client.host}")
    state = default_conversation.copy()
    return (state, state.to_gradio_chatbot(), "", None) + (disable_btn,) * 5


def add_text(state, text, image, image_process_mode, request: gr.Request):
    logger.info(f"add_text. ip: {request.client.host}. len: {len(text)}")
    if len(text) <= 0 and image is None:
        state.skip_next = True
        return (state, state.to_gradio_chatbot(), "", None) + (no_change_btn,) * 5
    if args.moderate:
        flagged = violates_moderation(text)
        if flagged:
            state.skip_next = True
            return (state, state.to_gradio_chatbot(), moderation_msg, None) + (
                no_change_btn,
            ) * 5

    text = text[:1536]  # Hard cut-off
    if image is not None:
        text = text[:1200]  # Hard cut-off for images
        if "<image>" not in text:
            # text = '<Image><image></Image>' + text
            text = text + "\n<image>"
        text = (text, image, image_process_mode)
        if len(state.get_images(return_pil=True)) > 0:
            state = default_conversation.copy()
    state.append_message(state.roles[0], text)
    state.append_message(state.roles[1], None)
    state.skip_next = False
    return (state, state.to_gradio_chatbot(), "", None) + (disable_btn,) * 5


def http_bot(
    state, model_selector, temperature, top_p, max_new_tokens, request: gr.Request
):
    logger.info(f"http_bot. ip: {request.client.host}")
    start_tstamp = time.time()
    model_name = model_selector

    if state.skip_next:
        # This generate call is skipped due to invalid inputs
        yield (state, state.to_gradio_chatbot()) + (no_change_btn,) * 5
        return

    if len(state.messages) == state.offset + 2:
        # First round of conversation
        if "llava" in model_name.lower():
            if "llama-2" in model_name.lower():
                template_name = "llava_llama_2"
            elif "v1" in model_name.lower():
                if "mmtag" in model_name.lower():
                    template_name = "v1_mmtag"
                elif (
                    "plain" in model_name.lower()
                    and "finetune" not in model_name.lower()
                ):
                    template_name = "v1_mmtag"
                else:
                    template_name = "llava_v1"
            elif "mpt" in model_name.lower():
                template_name = "mpt"
            else:
                if "mmtag" in model_name.lower():
                    template_name = "v0_mmtag"
                elif (
                    "plain" in model_name.lower()
                    and "finetune" not in model_name.lower()
                ):
                    template_name = "v0_mmtag"
                else:
                    template_name = "llava_v0"
        elif "mpt" in model_name:
            template_name = "mpt_text"
        elif "llama-2" in model_name:
            template_name = "llama_2"
        else:
            template_name = "vicuna_v1"
        new_state = conv_templates[template_name].copy()
        new_state.append_message(new_state.roles[0], state.messages[-2][1])
        new_state.append_message(new_state.roles[1], None)
        state = new_state

    # Query worker address
    controller_url = args.controller_url
    ret = requests.post(
        controller_url + "/get_worker_address", json={"model": model_name}
    )
    worker_addr = ret.json()["address"]
    logger.info(f"model_name: {model_name}, worker_addr: {worker_addr}")

    # No available worker
    if worker_addr == "":
        state.messages[-1][-1] = server_error_msg
        yield (
            state,
            state.to_gradio_chatbot(),
            disable_btn,
            disable_btn,
            disable_btn,
            enable_btn,
            enable_btn,
        )
        return

    # Construct prompt
    prompt = state.get_prompt()

    all_images = state.get_images(return_pil=True)
    all_image_hash = [hashlib.md5(image.tobytes()).hexdigest() for image in all_images]
    for image, hash in zip(all_images, all_image_hash):
        t = datetime.datetime.now()
        filename = os.path.join(
            LOGDIR, "serve_images", f"{t.year}-{t.month:02d}-{t.day:02d}", f"{hash}.jpg"
        )
        if not os.path.isfile(filename):
            os.makedirs(os.path.dirname(filename), exist_ok=True)
            image.save(filename)

    # Make requests
    pload = {
        "model": model_name,
        "prompt": prompt,
        "temperature": float(temperature),
        "top_p": float(top_p),
        "max_new_tokens": min(int(max_new_tokens), 1536),
        "stop": state.sep
        if state.sep_style in [SeparatorStyle.SINGLE, SeparatorStyle.MPT]
        else state.sep2,
        "images": f"List of {len(state.get_images())} images: {all_image_hash}",
    }
    logger.info(f"==== request ====\n{pload}")

    pload["images"] = state.get_images()

    state.messages[-1][-1] = "β–Œ"
    yield (state, state.to_gradio_chatbot()) + (disable_btn,) * 5

    try:
        # Stream output
        response = requests.post(
            worker_addr + "/worker_generate_stream",
            headers=headers,
            json=pload,
            stream=True,
            timeout=10,
        )
        for chunk in response.iter_lines(decode_unicode=False, delimiter=b"\0"):
            if chunk:
                data = json.loads(chunk.decode())
                if data["error_code"] == 0:
                    output = data["text"][len(prompt) :].strip()
                    state.messages[-1][-1] = output + "β–Œ"
                    yield (state, state.to_gradio_chatbot()) + (disable_btn,) * 5
                else:
                    output = data["text"] + f" (error_code: {data['error_code']})"
                    state.messages[-1][-1] = output
                    yield (state, state.to_gradio_chatbot()) + (
                        disable_btn,
                        disable_btn,
                        disable_btn,
                        enable_btn,
                        enable_btn,
                    )
                    return
                time.sleep(0.03)
    except requests.exceptions.RequestException as e:
        state.messages[-1][-1] = server_error_msg
        yield (state, state.to_gradio_chatbot()) + (
            disable_btn,
            disable_btn,
            disable_btn,
            enable_btn,
            enable_btn,
        )
        return

    state.messages[-1][-1] = state.messages[-1][-1][:-1]
    yield (state, state.to_gradio_chatbot()) + (enable_btn,) * 5

    finish_tstamp = time.time()
    logger.info(f"{output}")

    with open(get_conv_log_filename(), "a") as fout:
        data = {
            "tstamp": round(finish_tstamp, 4),
            "type": "chat",
            "model": model_name,
            "start": round(start_tstamp, 4),
            "finish": round(start_tstamp, 4),
            "state": state.dict(),
            "images": all_image_hash,
            "ip": request.client.host,
        }
        fout.write(json.dumps(data) + "\n")


title_markdown = """
# πŸŒ‹ LLaVA: Large Language and Vision Assistant
[[Project Page]](https://llava-vl.github.io) [[Paper]](https://arxiv.org/abs/2304.08485) [[Code]](https://github.com/haotian-liu/LLaVA) [[Model]](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)

ONLY WORKS WITH GPU!

You can load the model with 4-bit or 8-bit quantization to make it fit in smaller hardwares. Setting the environment variable `bits` to control the quantization.
*Note: 8-bit seems to be slower than both 4-bit/16-bit. Although it has enough VRAM to support 8-bit, until we figure out the inference speed issue, we recommend 4-bit for A10G for the best efficiency.*

Recommended configurations:
| Hardware          | T4-Small (16G)  | A10G-Small (24G) | A100-Large (40G) |
|-------------------|-----------------|------------------|------------------|
| **Bits**          | 4 (default)     | 4                | 16               |

"""

tos_markdown = """
### Terms of use
By using this service, users are required to agree to the following terms:
The service is a research preview intended for non-commercial use only. It only provides limited safety measures and may generate offensive content. It must not be used for any illegal, harmful, violent, racist, or sexual purposes. The service may collect user dialogue data for future research.
Please click the "Flag" button if you get any inappropriate answer! We will collect those to keep improving our moderator.
For an optimal experience, please use desktop computers for this demo, as mobile devices may compromise its quality.
"""


learn_more_markdown = """
### License
The service is a research preview intended for non-commercial use only, subject to the model [License](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) of LLaMA, [Terms of Use](https://openai.com/policies/terms-of-use) of the data generated by OpenAI, and [Privacy Practices](https://chrome.google.com/webstore/detail/sharegpt-share-your-chatg/daiacboceoaocpibfodeljbdfacokfjb) of ShareGPT. Please contact us if you find any potential violation.
"""

block_css = """

#buttons button {
    min-width: min(120px,100%);
}

"""


def build_demo(embed_mode):
    models = get_model_list()

    textbox = gr.Textbox(
        show_label=False, placeholder="Enter text and press ENTER", container=False
    )
    with gr.Blocks(title="LLaVA", theme=gr.themes.Default(), css=block_css) as demo:
        state = gr.State(default_conversation.copy())

        if not embed_mode:
            gr.Markdown(title_markdown)

        with gr.Row():
            with gr.Column(scale=3):
                with gr.Row(elem_id="model_selector_row"):
                    model_selector = gr.Dropdown(
                        choices=models,
                        value=models[0] if models else "Downloading the models...",
                        interactive=True if models else False,
                        show_label=False,
                        container=False,
                    )

                imagebox = gr.Image(type="pil")
                image_process_mode = gr.Radio(
                    ["Crop", "Resize", "Pad", "Default"],
                    value="Default",
                    label="Preprocess for non-square image",
                    visible=False,
                )

                cur_dir = os.path.dirname(os.path.abspath(__file__))
                gr.Examples(
                    examples=[
                        [
                            f"{cur_dir}/examples/extreme_ironing.jpg",
                            "What is unusual about this image?",
                        ],
                        [
                            f"{cur_dir}/examples/waterview.jpg",
                            "What are the things I should be cautious about when I visit here?",
                        ],
                    ],
                    inputs=[imagebox, textbox],
                )

                with gr.Accordion("Parameters", open=False) as parameter_row:
                    temperature = gr.Slider(
                        minimum=0.0,
                        maximum=1.0,
                        value=0.2,
                        step=0.1,
                        interactive=True,
                        label="Temperature",
                    )
                    top_p = gr.Slider(
                        minimum=0.0,
                        maximum=1.0,
                        value=0.7,
                        step=0.1,
                        interactive=True,
                        label="Top P",
                    )
                    max_output_tokens = gr.Slider(
                        minimum=0,
                        maximum=1024,
                        value=512,
                        step=64,
                        interactive=True,
                        label="Max output tokens",
                    )

            with gr.Column(scale=8):
                chatbot = gr.Chatbot(
                    elem_id="chatbot", label="LLaVA Chatbot", height=550
                )
                with gr.Row():
                    with gr.Column(scale=8):
                        textbox.render()
                    with gr.Column(scale=1, min_width=50):
                        submit_btn = gr.Button(
                            value="Send", variant="primary", interactive=False
                        )
                with gr.Row(elem_id="buttons") as button_row:
                    upvote_btn = gr.Button(value="πŸ‘  Upvote", interactive=False)
                    downvote_btn = gr.Button(value="πŸ‘Ž  Downvote", interactive=False)
                    flag_btn = gr.Button(value="⚠️  Flag", interactive=False)
                    # stop_btn = gr.Button(value="⏹️  Stop Generation", interactive=False)
                    regenerate_btn = gr.Button(value="πŸ”„  Regenerate", interactive=False)
                    clear_btn = gr.Button(value="πŸ—‘οΈ  Clear history", interactive=False)

        if not embed_mode:
            gr.Markdown(tos_markdown)
            gr.Markdown(learn_more_markdown)
        url_params = gr.JSON(visible=False)

        # Register listeners
        btn_list = [upvote_btn, downvote_btn, flag_btn, regenerate_btn, clear_btn]
        upvote_btn.click(
            upvote_last_response,
            [state, model_selector],
            [textbox, upvote_btn, downvote_btn, flag_btn],
        )
        downvote_btn.click(
            downvote_last_response,
            [state, model_selector],
            [textbox, upvote_btn, downvote_btn, flag_btn],
        )
        flag_btn.click(
            flag_last_response,
            [state, model_selector],
            [textbox, upvote_btn, downvote_btn, flag_btn],
        )
        regenerate_btn.click(
            regenerate,
            [state, image_process_mode],
            [state, chatbot, textbox, imagebox] + btn_list,
        ).then(
            http_bot,
            [state, model_selector, temperature, top_p, max_output_tokens],
            [state, chatbot] + btn_list,
        )
        clear_btn.click(
            clear_history, None, [state, chatbot, textbox, imagebox] + btn_list
        )

        textbox.submit(
            add_text,
            [state, textbox, imagebox, image_process_mode],
            [state, chatbot, textbox, imagebox] + btn_list,
        ).then(
            http_bot,
            [state, model_selector, temperature, top_p, max_output_tokens],
            [state, chatbot] + btn_list,
        )
        submit_btn.click(
            add_text,
            [state, textbox, imagebox, image_process_mode],
            [state, chatbot, textbox, imagebox] + btn_list,
        ).then(
            http_bot,
            [state, model_selector, temperature, top_p, max_output_tokens],
            [state, chatbot] + btn_list,
        )

        if args.model_list_mode == "once":
            demo.load(
                load_demo,
                [url_params],
                [state, model_selector],
                _js=get_window_url_params,
            )
        elif args.model_list_mode == "reload":
            demo.load(
                load_demo_refresh_model_list, None, [state, model_selector, submit_btn]
            )
        else:
            raise ValueError(f"Unknown model list mode: {args.model_list_mode}")

    return demo


def start_controller():
    logger.info("Starting the controller")
    controller_command = [
        "python",
        "-m",
        "llava.serve.controller",
        "--host",
        "0.0.0.0",
        "--port",
        "10000",
    ]
    return subprocess.Popen(controller_command)


def start_worker(model_path: str, bits=16):
    logger.info(f"Starting the model worker for the model {model_path}")
    model_name = model_path.strip("/").split("/")[-1]
    assert bits in [4, 8, 16], "It can be only loaded with 16-bit, 8-bit, and 4-bit."
    if bits != 16:
        model_name += f"-{bits}bit"
    worker_command = [
        "python",
        "-m",
        "llava.serve.model_worker",
        "--host",
        "0.0.0.0",
        "--controller",
        "http://localhost:10000",
        "--model-path",
        model_path,
        "--model-name",
        model_name,
    ]
    if bits != 16:
        worker_command += [f"--load-{bits}bit"]
    return subprocess.Popen(worker_command)


def get_args():
    parser = argparse.ArgumentParser()
    parser.add_argument("--host", type=str, default="0.0.0.0")
    parser.add_argument("--port", type=int)
    parser.add_argument("--controller-url", type=str, default="http://localhost:10000")
    parser.add_argument("--concurrency-count", type=int, default=8)
    parser.add_argument(
        "--model-list-mode", type=str, default="reload", choices=["once", "reload"]
    )
    parser.add_argument("--share", action="store_true")
    parser.add_argument("--moderate", action="store_true")
    parser.add_argument("--embed", action="store_true")

    args = parser.parse_args()

    return args


def start_demo(args):
    demo = build_demo(args.embed)
    demo.queue(
        concurrency_count=args.concurrency_count, status_update_rate=10, api_open=False
    ).launch(server_name=args.host, server_port=args.port, share=args.share)


if __name__ == "__main__":
    args = get_args()
    logger.info(f"args: {args}")

    model_path = "liuhaotian/llava-v1.5-13b"
    bits = int(os.getenv("bits", 8))

    controller_proc = start_controller()
    worker_proc = start_worker(model_path, bits=bits)

    # Wait for worker and controller to start
    time.sleep(10)

    exit_status = 0
    try:
        start_demo(args)
    except Exception as e:
        print(e)
        exit_status = 1
    finally:
        worker_proc.kill()
        controller_proc.kill()

        sys.exit(exit_status)