File size: 17,080 Bytes
558edd9
 
 
 
 
 
 
 
 
 
 
 
a0303a2
adb2c74
558edd9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6721fd3
 
 
 
adb2c74
 
 
6721fd3
558edd9
6721fd3
 
 
558edd9
 
 
adb2c74
6721fd3
 
558edd9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
adb2c74
 
 
 
 
 
 
 
 
558edd9
 
 
 
 
 
 
 
 
 
adb2c74
 
 
 
558edd9
 
 
 
adb2c74
 
 
e12262f
558edd9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a296f33
558edd9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a296f33
558edd9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Run codes."""
# pylint: disable=line-too-long, broad-exception-caught, invalid-name, missing-function-docstring, too-many-instance-attributes, missing-class-docstring
import os
import time
from dataclasses import asdict, dataclass
from pathlib import Path
from types import SimpleNamespace
from urllib.parse import urlparse

import gradio as gr
import psutil
from about_time import about_time
from ctransformers import AutoModelForCausalLM
from huggingface_hub import hf_hub_download, snapshot_download
from loguru import logger

filename_list = [
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q2_K.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q3_K_L.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q3_K_M.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q3_K_S.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_0.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_1.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_K_M.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_K_S.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q5_0.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q5_1.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q5_K_M.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q5_K_S.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q6_K.bin",
    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q8_0.bin",
]

URL = "https://huggingface.co/TheBloke/Wizard-Vicuna-7B-Uncensored-GGML/raw/main/Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_K_M.bin"  # 4.05G

URL = "https://huggingface.co/TheBloke/30B-Lazarus-GGML/blob/main/30b-Lazarus.ggmlv3.q4_0.bin"
URL = "https://huggingface.co/TheBloke/30B-Lazarus-GGML/blob/main/30b-Lazarus.ggmlv3.q4_1.bin"
URL = "https://huggingface.co/TheBloke/30B-Lazarus-GGML/resolve/main/30b-Lazarus.ggmlv3.q4_K_M.bin"
URL = "https://huggingface.co/TheBloke/30B-Lazarus-GGML/resolve/main/30b-Lazarus.ggmlv3.q4_K_S.bin"  # 18GB

URL = "https://huggingface.co/TheBloke/30B-Lazarus-GGML/blob/main/30b-Lazarus.ggmlv3.q3_K_S.bin"  # 14GB

MODEL_FILENAME = Path(URL).name

# MODEL_FILENAME = filename_list[0]  # q2_K 4.05G
# MODEL_FILENAME = filename_list[5]  # q4_1 4.21

REPO_ID = "/".join(
    urlparse(URL).path.strip("/").split("/")[:2]
)
# TheBloke/30B-Lazarus-GGML
# # TheBloke/Wizard-Vicuna-7B-Uncensored-GGML

DESTINATION_FOLDER = "models"

os.environ["TZ"] = "Asia/Shanghai"
try:
    time.tzset()  # type: ignore # pylint: disable=no-member
except Exception:
    # Windows
    logger.warning("Windows, cant run time.tzset()")

ns = SimpleNamespace(
    response="",
    generator=[],
)

default_system_prompt = "A conversation between a user and an LLM-based AI assistant named Local Assistant. Local Assistant gives helpful and honest answers."

user_prefix = "[user]: "
assistant_prefix = "[assistant]: "


def predict_str(prompt, bot):  # bot is in fact bot_history
    # logger.debug(f"{prompt=}, {bot=}, {timeout=}")

    if bot is None:
        bot = []

    logger.debug(f"{prompt=}, {bot=}")

    try:
        # user_prompt = prompt
        generator = generate(
            LLM,
            GENERATION_CONFIG,
            system_prompt=default_system_prompt,
            user_prompt=prompt.strip(),
        )

        ns.generator = generator  # for .then

    except Exception as exc:
        logger.error(exc)

    # bot.append([prompt, f"{response} {_}"])
    # return prompt, bot

    _ = bot + [[prompt, None]]
    logger.debug(f"{prompt=}, {_=}")

    return prompt, _


def bot_str(bot):
    if bot:
        bot[-1][1] = ""
    else:
        bot = [["Something is wrong", ""]]

    print(assistant_prefix, end=" ", flush=True)

    response = ""

    flag = 1
    then = time.time()
    for word in ns.generator:
        # record first response time
        if flag:
            logger.debug(f"\t {time.time() - then:.1f}s")
            flag = 0
        print(word, end="", flush=True)
        # print(word, flush=True)  # vertical stream
        response += word
        bot[-1][1] = response
        yield bot


def predict(prompt, bot):
    # logger.debug(f"{prompt=}, {bot=}, {timeout=}")
    logger.debug(f"{prompt=}, {bot=}")

    ns.response = ""
    then = time.time()
    with about_time() as atime:  # type: ignore
        try:
            # user_prompt = prompt
            generator = generate(
                LLM,
                GENERATION_CONFIG,
                system_prompt=default_system_prompt,
                user_prompt=prompt.strip(),
            )

            ns.generator = generator  # for .then

            print(assistant_prefix, end=" ", flush=True)

            response = ""
            buff.update(value="diggin...")

            flag = 1
            for word in generator:
                # record first response time
                if flag:
                    logger.debug(f"\t {time.time() - then:.1f}s")
                    flag = 0
                # print(word, end="", flush=True)
                print(word, flush=True)  # vertical stream
                response += word
                ns.response = response
                buff.update(value=response)
            print("")
            logger.debug(f"{response=}")
        except Exception as exc:
            logger.error(exc)
            response = f"{exc=}"

    # bot = {"inputs": [response]}
    _ = (
        f"(time elapsed: {atime.duration_human}, "  # type: ignore
        f"{atime.duration/(len(prompt) + len(response)):.1f}s/char)"  # type: ignore
    )

    bot.append([prompt, f"{response} {_}"])

    return prompt, bot


def predict_api(prompt):
    logger.debug(f"{prompt=}")
    ns.response = ""
    try:
        # user_prompt = prompt
        _ = GenerationConfig(
            temperature=0.2,
            top_k=0,
            top_p=0.9,
            repetition_penalty=1.0,
            max_new_tokens=512,  # adjust as needed
            seed=42,
            reset=False,  # reset history (cache)
            stream=True,  # TODO stream=False and generator
            threads=os.cpu_count() // 2,  # type: ignore  # adjust for your CPU
            stop=["<|im_end|>", "|<"],
        )

        # TODO: stream does not make sense in api?
        generator = generate(
            LLM, _, system_prompt=default_system_prompt, user_prompt=prompt.strip()
        )
        print(assistant_prefix, end=" ", flush=True)

        response = ""
        buff.update(value="diggin...")
        for word in generator:
            print(word, end="", flush=True)
            response += word
            ns.response = response
            buff.update(value=response)
        print("")
        logger.debug(f"{response=}")
    except Exception as exc:
        logger.error(exc)
        response = f"{exc=}"
    # bot = {"inputs": [response]}
    # bot = [(prompt, response)]

    return response


def download_quant(destination_folder: str, repo_id: str, model_filename: str):
    local_path = os.path.abspath(destination_folder)
    return hf_hub_download(
        repo_id=repo_id,
        filename=model_filename,
        local_dir=local_path,
        local_dir_use_symlinks=True,
    )


@dataclass
class GenerationConfig:
    temperature: float
    top_k: int
    top_p: float
    repetition_penalty: float
    max_new_tokens: int
    seed: int
    reset: bool
    stream: bool
    threads: int
    stop: list[str]


def format_prompt(system_prompt: str, user_prompt: str):
    """Format prompt based on: https://huggingface.co/spaces/mosaicml/mpt-30b-chat/blob/main/app.py."""
    # TODO: fix prompts

    system_prompt = f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
    user_prompt = f"<|im_start|>user\n{user_prompt}<|im_end|>\n"
    assistant_prompt = "<|im_start|>assistant\n"

    return f"{system_prompt}{user_prompt}{assistant_prompt}"


def generate(
    llm: AutoModelForCausalLM,
    generation_config: GenerationConfig,
    system_prompt: str = default_system_prompt,
    user_prompt: str = "",
):
    """Run model inference, will return a Generator if streaming is true."""
    # if not user_prompt.strip():
    return llm(
        format_prompt(
            system_prompt,
            user_prompt,
        ),
        **asdict(generation_config),
    )


# if "mpt" in model_filename:
#     config = AutoConfig.from_pretrained("mosaicml/mpt-30b-cha t", context_length=8192)
#     llm = AutoModelForCausalLM.from_pretrained(
#         os.path.abspath(f"models/{model_filename}"),
#         model_type="mpt",
#         config=config,
#     )

# https://huggingface.co/spaces/matthoffner/wizardcoder-ggml/blob/main/main.py
_ = """
llm = AutoModelForCausalLM.from_pretrained(
    "TheBloke/WizardCoder-15B-1.0-GGML",
    model_file="WizardCoder-15B-1.0.ggmlv3.q4_0.bin",
    model_type="starcoder",
    threads=8
)
# """

logger.info(f"start dl, {REPO_ID=}, {MODEL_FILENAME=}, {DESTINATION_FOLDER=}")

# download_quant(DESTINATION_FOLDER, REPO_ID, MODEL_FILENAME)
snapshot_download(
  repo_id=REPO_ID,  # TheBloke/30B-Lazarus-GGML
  allow_patterns=MODEL_FILENAME,  # 30b-Lazarus.ggmlv3.q4_K_S.bin 18.3G
  #   revision="ggmlv3",
  local_dir="models",
)

logger.info("done dl")

logger.debug(f"{os.cpu_count()=} {psutil.cpu_count(logical=False)=}")
cpu_count = os.cpu_count() // 2  # type: ignore
cpu_count = psutil.cpu_count(logical=False)

logger.debug(f"{cpu_count=}")

logger.info("load llm")

# from ctransformers import AutoConfig
# AutoConfig(REPO_ID)
# AutoConfig(config='TheBloke/30B-Lazarus-GGML', model_type=None)

_ = Path("models", MODEL_FILENAME).absolute().as_posix()
logger.debug(f"model_file: {_}, exists: {Path(_).exists()}")
LLM = AutoModelForCausalLM.from_pretrained(
    # "TheBloke/WizardCoder-15B-1.0-GGML",
    # REPO_ID,  # DESTINATION_FOLDER,  # model_path_or_repo_id: str required
    # model_file=_,
    _,
    model_type="llama",  #AutoConfig.from_pretrained(REPO_ID).model_type,
    threads=cpu_count,
)

logger.info("done load llm")

GENERATION_CONFIG = GenerationConfig(
    temperature=0.2,
    top_k=0,
    top_p=0.9,
    repetition_penalty=1.0,
    max_new_tokens=512,  # adjust as needed
    seed=42,
    reset=False,  # reset history (cache)
    stream=True,  # streaming per word/token
    threads=cpu_count,
    stop=["<|im_end|>", "|<"],  # TODO possible fix of stop
)

css = """
    .importantButton {
        background: linear-gradient(45deg, #7e0570,#5d1c99, #6e00ff) !important;
        border: none !important;
    }
    .importantButton:hover {
        background: linear-gradient(45deg, #ff00e0,#8500ff, #6e00ff) !important;
        border: none !important;
    }
    .disclaimer {font-variant-caps: all-small-caps; font-size: xx-small;}
    .xsmall {font-size: x-small;}
"""
etext = """In America, where cars are an important part of the national psyche, a decade ago people had suddenly started to drive less, which had not happened since the oil shocks of the 1970s. """
examples = [
    ["How to pick a lock? Provide detailed steps."],
    ["Explain the plot of Cinderella in a sentence."],
    [
        "How long does it take to become proficient in French, and what are the best methods for retaining information?"
    ],
    ["What are some common mistakes to avoid when writing code?"],
    ["Build a prompt to generate a beautiful portrait of a horse"],
    ["Suggest four metaphors to describe the benefits of AI"],
    ["Write a pop song about leaving home for the sandy beaches."],
    ["Write a summary demonstrating my ability to tame lions"],
    ["鲁迅和周树人什么关系 说中文"],
    ["鲁迅和周树人什么关系"],
    ["鲁迅和周树人什么关系 用英文回答"],
    ["从前有一头牛,这头牛后面有什么?"],
    ["正无穷大加一大于正无穷大吗?"],
    ["正无穷大加正无穷大大于正无穷大吗?"],
    ["-2的平方根等于什么"],
    ["树上有5只鸟,猎人开枪打死了一只。树上还有几只鸟?"],
    ["树上有11只鸟,猎人开枪打死了一只。树上还有几只鸟?提示:需考虑鸟可能受惊吓飞走。"],
    ["以红楼梦的行文风格写一张委婉的请假条。不少于320字。"],
    [f"{etext} 翻成中文,列出3个版本"],
    [f"{etext} \n 翻成中文,保留原意,但使用文学性的语言。不要写解释。列出3个版本"],
    ["假定 1 + 2 = 4, 试求 7 + 8"],
    ["判断一个数是不是质数的 javascript 码"],
    ["实现python 里 range(10)的 javascript 码"],
    ["实现python 里 [*(range(10)]的 javascript 码"],
    ["Erkläre die Handlung von Cinderella in einem Satz."],
    ["Erkläre die Handlung von Cinderella in einem Satz. Auf Deutsch"],
]

with gr.Blocks(
    # title="mpt-30b-chat-ggml",
    title=f"{MODEL_FILENAME}",
    theme=gr.themes.Soft(text_size="sm", spacing_size="sm"),
    css=css,
) as block:
    with gr.Accordion("🎈 Info", open=False):
        # gr.HTML(
        #     """<center><a href="https://huggingface.co/spaces/mikeee/mpt-30b-chat?duplicate=true"><img src="https://bit.ly/3gLdBN6" alt="Duplicate"></a> and spin a CPU UPGRADE to avoid the queue</center>"""
        # )
        gr.Markdown(
            f"""<h5><center><{REPO_ID}>{MODEL_FILENAME}</center></h4>
            The bot only speaks English.

            Most examples are meant for another model.
            You probably should try to test
            some related prompts.
            """,
            elem_classes="xsmall",
        )

    # chatbot = gr.Chatbot().style(height=700)  # 500
    chatbot = gr.Chatbot(height=500)
    buff = gr.Textbox(show_label=False, visible=False)
    with gr.Row():
        with gr.Column(scale=4):
            msg = gr.Textbox(
                label="Chat Message Box",
                placeholder="Ask me anything (press Enter or click Submit to send)",
                show_label=False,
            ).style(container=False)
        with gr.Column(scale=1, min_width=50):
            with gr.Row():
                submit = gr.Button("Submit", elem_classes="xsmall")
                stop = gr.Button("Stop", visible=False)
                clear = gr.Button("Clear History", visible=True)
    with gr.Row(visible=False):
        with gr.Accordion("Advanced Options:", open=False):
            with gr.Row():
                with gr.Column(scale=2):
                    system = gr.Textbox(
                        label="System Prompt",
                        value=default_system_prompt,
                        show_label=False,
                    ).style(container=False)
                with gr.Column():
                    with gr.Row():
                        change = gr.Button("Change System Prompt")
                        reset = gr.Button("Reset System Prompt")

    with gr.Accordion("Example Inputs", open=True):
        examples = gr.Examples(
            examples=examples,
            inputs=[msg],
            examples_per_page=20,
        )

    # with gr.Row():
    with gr.Accordion("Disclaimer", open=False):
        _ = "-".join(MODEL_FILENAME.split("-")[:2])
        gr.Markdown(
            f"Disclaimer: {_} can produce factually incorrect output, and should not be relied on to produce "
            "factually accurate information. {_} was trained on various public datasets; while great efforts "
            "have been taken to clean the pretraining data, it is possible that this model could generate lewd, "
            "biased, or otherwise offensive outputs.",
            elem_classes=["disclaimer"],
        )
    _ = """
    msg.submit(
        # fn=conversation.user_turn,
        fn=predict,
        inputs=[msg, chatbot],
        outputs=[msg, chatbot],
        # queue=True,
        show_progress="full",
        api_name="predict",
    )
    submit.click(
        fn=lambda x, y: ("",) + predict(x, y)[1:],  # clear msg
        inputs=[msg, chatbot],
        outputs=[msg, chatbot],
        queue=True,
        show_progress="full",
    )
    # """
    msg.submit(
        # fn=conversation.user_turn,
        fn=predict_str,
        inputs=[msg, chatbot],
        outputs=[msg, chatbot],
        queue=True,
        show_progress="full",
        api_name="predict",
    ).then(bot_str, chatbot, chatbot)
    submit.click(
        fn=lambda x, y: ("",) + predict_str(x, y)[1:],  # clear msg
        inputs=[msg, chatbot],
        outputs=[msg, chatbot],
        queue=True,
        show_progress="full",
    ).then(bot_str, chatbot, chatbot)

    clear.click(lambda: None, None, chatbot, queue=False)

    # update buff Textbox, every: units in seconds)
    # https://huggingface.co/spaces/julien-c/nvidia-smi/discussions
    # does not work
    # AttributeError: 'Blocks' object has no attribute 'run_forever'
    # block.run_forever(lambda: ns.response, None, [buff], every=1)

    with gr.Accordion("For Chat/Translation API", open=False, visible=False):
        input_text = gr.Text()
        api_btn = gr.Button("Go", variant="primary")
        out_text = gr.Text()
    api_btn.click(
        predict_api,
        input_text,
        out_text,
        # show_progress="full",
        api_name="api",
    )

# concurrency_count=5, max_size=20
# max_size=36, concurrency_count=14
block.queue(concurrency_count=5, max_size=20).launch(debug=True)