File size: 8,439 Bytes
42f4126
 
aea4015
42f4126
 
 
aea4015
 
 
04fa68c
aea4015
 
 
 
 
 
 
89817e2
42f4126
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
aea4015
42f4126
 
 
 
 
 
 
 
aea4015
 
 
 
42f4126
 
 
 
aea4015
46ef1e4
42f4126
46ef1e4
 
 
42f4126
 
 
aea4015
42f4126
 
 
 
 
 
 
 
 
 
 
 
 
 
070daeb
42f4126
46ef1e4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
42f4126
 
 
 
 
 
46ef1e4
42f4126
aea4015
 
 
 
 
 
42f4126
aea4015
 
42f4126
aea4015
 
 
 
 
 
42f4126
aea4015
42f4126
aea4015
 
 
 
 
 
42f4126
aea4015
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
42f4126
 
 
 
aea4015
42f4126
46ef1e4
 
 
42f4126
 
 
46ef1e4
42f4126
46ef1e4
 
 
 
 
 
 
 
42f4126
46ef1e4
 
 
 
 
 
 
 
 
 
 
 
42f4126
aea4015
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
46ef1e4
42f4126
 
46ef1e4
 
 
 
 
 
 
 
 
 
 
42f4126
46ef1e4
42f4126
46ef1e4
42f4126
 
 
 
46ef1e4
 
42f4126
 
 
aea4015
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
import re
import threading

import gradio as gr
import spaces
import transformers
from transformers import pipeline

# λͺ¨λΈκ³Ό ν† ν¬λ‚˜μ΄μ € λ‘œλ”©
model_name = "CohereForAI/c4ai-command-r7b-arabic-02-2025"
if gr.NO_RELOAD:
    pipe = pipeline(
        "text-generation",
        model=model_name,
        device_map="auto",
        torch_dtype="auto",
    )

# μ΅œμ’… 닡변을 κ°μ§€ν•˜κΈ° μœ„ν•œ 마컀
ANSWER_MARKER = "**λ‹΅λ³€**"

# 단계별 좔둠을 μ‹œμž‘ν•˜λŠ” λ¬Έμž₯λ“€
rethink_prepends = [
    "자, 이제 λ‹€μŒμ„ νŒŒμ•…ν•΄μ•Ό ν•©λ‹ˆλ‹€ ",
    "제 μƒκ°μ—λŠ” ",
    "μž μ‹œλ§Œμš”, 제 μƒκ°μ—λŠ” ",
    "λ‹€μŒ 사항이 λ§žλŠ”μ§€ 확인해 λ³΄κ² μŠ΅λ‹ˆλ‹€ ",
    "λ˜ν•œ κΈ°μ–΅ν•΄μ•Ό ν•  것은 ",
    "또 λ‹€λ₯Έ μ£Όλͺ©ν•  점은 ",
    "그리고 μ €λŠ” λ‹€μŒκ³Ό 같은 사싀도 κΈ°μ–΅ν•©λ‹ˆλ‹€ ",
    "이제 μΆ©λΆ„νžˆ μ΄ν•΄ν–ˆλ‹€κ³  μƒκ°ν•©λ‹ˆλ‹€ ",
    "μ§€κΈˆκΉŒμ§€μ˜ 정보λ₯Ό λ°”νƒ•μœΌλ‘œ, μ›λž˜ μ§ˆλ¬Έμ— μ‚¬μš©λœ μ–Έμ–΄λ‘œ λ‹΅λ³€ν•˜κ² μŠ΅λ‹ˆλ‹€:"
    "\n{question}\n"
    f"\n{ANSWER_MARKER}\n",
]


# μˆ˜μ‹ ν‘œμ‹œ 문제 해결을 μœ„ν•œ μ„€μ •
latex_delimiters = [
    {"left": "$$", "right": "$$", "display": True},
    {"left": "$", "right": "$", "display": False},
]


def reformat_math(text):
    """Gradio ꡬ문(Katex)을 μ‚¬μš©ν•˜λ„λ‘ MathJax ꡬ뢄 기호 μˆ˜μ •.
    이것은 Gradioμ—μ„œ μˆ˜ν•™ 곡식을 ν‘œμ‹œν•˜κΈ° μœ„ν•œ μž„μ‹œ ν•΄κ²°μ±…μž…λ‹ˆλ‹€. ν˜„μž¬λ‘œμ„œλŠ”
    λ‹€λ₯Έ latex_delimitersλ₯Ό μ‚¬μš©ν•˜μ—¬ μ˜ˆμƒλŒ€λ‘œ μž‘λ™ν•˜κ²Œ ν•˜λŠ” 방법을 μ°Ύμ§€ λͺ»ν–ˆμŠ΅λ‹ˆλ‹€...
    """
    text = re.sub(r"\\\[\s*(.*?)\s*\\\]", r"$$\1$$", text, flags=re.DOTALL)
    text = re.sub(r"\\\(\s*(.*?)\s*\\\)", r"$\1$", text, flags=re.DOTALL)
    return text


def user_input(message, history_original, history_thinking):
    """μ‚¬μš©μž μž…λ ₯을 νžˆμŠ€ν† λ¦¬μ— μΆ”κ°€ν•˜κ³  μž…λ ₯ ν…μŠ€νŠΈ μƒμž λΉ„μš°κΈ°"""
    return "", history_original + [
        gr.ChatMessage(role="user", content=message.replace(ANSWER_MARKER, ""))
    ], history_thinking + [
        gr.ChatMessage(role="user", content=message.replace(ANSWER_MARKER, ""))
    ]


def rebuild_messages(history: list):
    """쀑간 생각 κ³Όμ • 없이 λͺ¨λΈμ΄ μ‚¬μš©ν•  νžˆμŠ€ν† λ¦¬μ—μ„œ λ©”μ‹œμ§€ μž¬κ΅¬μ„±"""
    messages = []
    for h in history:
        if isinstance(h, dict) and not h.get("metadata", {}).get("title", False):
            messages.append(h)
        elif (
            isinstance(h, gr.ChatMessage)
            and h.metadata.get("title")
            and isinstance(h.content, str)
        ):
            messages.append({"role": h.role, "content": h.content})
    return messages


@spaces.GPU
def bot_original(
    history: list,
    max_num_tokens: int,
    do_sample: bool,
    temperature: float,
):
    """원본 λͺ¨λΈμ΄ μ§ˆλ¬Έμ— λ‹΅λ³€ν•˜λ„λ‘ ν•˜κΈ° (μΆ”λ‘  κ³Όμ • 없이)"""

    # λ‚˜μ€‘μ— μŠ€λ ˆλ“œμ—μ„œ 토큰을 슀트림으둜 κ°€μ Έμ˜€κΈ° μœ„ν•¨
    streamer = transformers.TextIteratorStreamer(
        pipe.tokenizer,  # pyright: ignore
        skip_special_tokens=True,
        skip_prompt=True,
    )

    # 보쑰자 λ©”μ‹œμ§€ μ€€λΉ„
    history.append(
        gr.ChatMessage(
            role="assistant",
            content=str(""),
        )
    )

    # ν˜„μž¬ μ±„νŒ…μ— ν‘œμ‹œλ  λ©”μ‹œμ§€
    messages = rebuild_messages(history[:-1])  # λ§ˆμ§€λ§‰ 빈 λ©”μ‹œμ§€ μ œμ™Έ
    
    # 원본 λͺ¨λΈμ€ μΆ”λ‘  없이 λ°”λ‘œ λ‹΅λ³€
    t = threading.Thread(
        target=pipe,
        args=(messages,),
        kwargs=dict(
            max_new_tokens=max_num_tokens,
            streamer=streamer,
            do_sample=do_sample,
            temperature=temperature,
        ),
    )
    t.start()

    for token in streamer:
        history[-1].content += token
        history[-1].content = reformat_math(history[-1].content)
        yield history
    t.join()

    yield history


@spaces.GPU
def bot_thinking(
    history: list,
    max_num_tokens: int,
    final_num_tokens: int,
    do_sample: bool,
    temperature: float,
):
    """μΆ”λ‘  과정을 ν¬ν•¨ν•˜μ—¬ λͺ¨λΈμ΄ μ§ˆλ¬Έμ— λ‹΅λ³€ν•˜λ„λ‘ ν•˜κΈ°"""

    # λ‚˜μ€‘μ— μŠ€λ ˆλ“œμ—μ„œ 토큰을 슀트림으둜 κ°€μ Έμ˜€κΈ° μœ„ν•¨
    streamer = transformers.TextIteratorStreamer(
        pipe.tokenizer,  # pyright: ignore
        skip_special_tokens=True,
        skip_prompt=True,
    )

    # ν•„μš”ν•œ 경우 좔둠에 μ§ˆλ¬Έμ„ λ‹€μ‹œ μ‚½μž…ν•˜κΈ° μœ„ν•¨
    question = history[-1]["content"]

    # 보쑰자 λ©”μ‹œμ§€ μ€€λΉ„
    history.append(
        gr.ChatMessage(
            role="assistant",
            content=str(""),
            metadata={"title": "🧠 생각 쀑...", "status": "pending"},
        )
    )

    # ν˜„μž¬ μ±„νŒ…μ— ν‘œμ‹œλ  μΆ”λ‘  κ³Όμ •
    messages = rebuild_messages(history)
    for i, prepend in enumerate(rethink_prepends):
        if i > 0:
            messages[-1]["content"] += "\n\n"
        messages[-1]["content"] += prepend.format(question=question)

        num_tokens = int(
            max_num_tokens if ANSWER_MARKER not in prepend else final_num_tokens
        )
        t = threading.Thread(
            target=pipe,
            args=(messages,),
            kwargs=dict(
                max_new_tokens=num_tokens,
                streamer=streamer,
                do_sample=do_sample,
                temperature=temperature,
            ),
        )
        t.start()

        # μƒˆ λ‚΄μš©μœΌλ‘œ νžˆμŠ€ν† λ¦¬ μž¬κ΅¬μ„±
        history[-1].content += prepend.format(question=question)
        if ANSWER_MARKER in prepend:
            history[-1].metadata = {"title": "πŸ’­ 사고 κ³Όμ •", "status": "done"}
            # 생각 μ’…λ£Œ, 이제 λ‹΅λ³€μž…λ‹ˆλ‹€ (쀑간 단계에 λŒ€ν•œ 메타데이터 μ—†μŒ)
            history.append(gr.ChatMessage(role="assistant", content=""))
        for token in streamer:
            history[-1].content += token
            history[-1].content = reformat_math(history[-1].content)
            yield history
        t.join()

    yield history


with gr.Blocks(fill_height=True, title="λͺ¨λ“  LLM λͺ¨λΈμ— μΆ”λ‘  λŠ₯λ ₯ λΆ€μ—¬ν•˜κΈ°") as demo:
    with gr.Row(scale=1):
        with gr.Column(scale=2):
            gr.Markdown("## Before (Original)")
            chatbot_original = gr.Chatbot(
                scale=1,
                type="messages",
                latex_delimiters=latex_delimiters,
                label="Original Model (No Reasoning)"
            )
        
        with gr.Column(scale=2):
            gr.Markdown("## After (Thinking)")
            chatbot_thinking = gr.Chatbot(
                scale=1,
                type="messages",
                latex_delimiters=latex_delimiters,
                label="Model with Reasoning"
            )
    
    with gr.Row():
        msg = gr.Textbox(
            submit_btn=True,
            label="",
            show_label=False,
            placeholder="여기에 μ§ˆλ¬Έμ„ μž…λ ₯ν•˜μ„Έμš”.",
            autofocus=True,
        )
    
    with gr.Row():
        with gr.Column():
            gr.Markdown("""## λ§€κ°œλ³€μˆ˜ μ‘°μ •""")
            num_tokens = gr.Slider(
                50,
                4000,
                2000,
                step=1,
                label="μΆ”λ‘  단계당 μ΅œλŒ€ 토큰 수",
                interactive=True,
            )
            final_num_tokens = gr.Slider(
                50,
                4000,
                2000,
                step=1,
                label="μ΅œμ’… λ‹΅λ³€μ˜ μ΅œλŒ€ 토큰 수",
                interactive=True,
            )
            do_sample = gr.Checkbox(True, label="μƒ˜ν”Œλ§ μ‚¬μš©")
            temperature = gr.Slider(0.1, 1.0, 0.7, step=0.1, label="μ˜¨λ„")

    # μ‚¬μš©μžκ°€ λ©”μ‹œμ§€λ₯Ό μ œμΆœν•˜λ©΄ 두 봇이 λ™μ‹œμ— μ‘λ‹΅ν•©λ‹ˆλ‹€
    msg.submit(
        user_input,
        [msg, chatbot_original, chatbot_thinking],  # μž…λ ₯
        [msg, chatbot_original, chatbot_thinking],  # 좜λ ₯
    ).then(
        bot_original,
        [
            chatbot_original,
            num_tokens,
            do_sample,
            temperature,
        ],
        chatbot_original,  # 좜λ ₯μ—μ„œ μƒˆ νžˆμŠ€ν† λ¦¬ μ €μž₯
    ).then(
        bot_thinking,
        [
            chatbot_thinking,
            num_tokens,
            final_num_tokens,
            do_sample,
            temperature,
        ],
        chatbot_thinking,  # 좜λ ₯μ—μ„œ μƒˆ νžˆμŠ€ν† λ¦¬ μ €μž₯
    )

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