File size: 21,400 Bytes
42f4126
 
03434f6
89817e2
03434f6
ccc2ed2
 
42f4126
 
 
03434f6
89817e2
42f4126
03434f6
 
 
ccc2ed2
03434f6
ccc2ed2
070daeb
ccc2ed2
 
0df81f7
070daeb
e94718a
070daeb
 
 
ccc2ed2
2198075
 
 
03434f6
070daeb
03434f6
ccc2ed2
42f4126
89817e2
 
 
 
 
 
 
 
 
 
 
42f4126
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
03434f6
 
 
ccc2ed2
03434f6
 
 
 
ccc2ed2
03434f6
ccc2ed2
03434f6
 
ccc2ed2
03434f6
ccc2ed2
03434f6
 
 
 
 
ccc2ed2
03434f6
ccc2ed2
03434f6
 
 
 
ccc2ed2
 
03434f6
 
 
 
 
 
 
 
 
 
 
 
 
 
42f4126
 
03434f6
42f4126
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ccc2ed2
03434f6
ccc2ed2
070daeb
ccc2ed2
 
 
070daeb
ccc2ed2
070daeb
03434f6
ccc2ed2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
89817e2
 
 
 
 
 
 
 
ccc2ed2
5d8d623
 
 
 
 
ccc2ed2
 
 
 
 
 
 
 
 
 
 
 
 
5d8d623
 
 
03434f6
 
 
 
 
 
 
ccc2ed2
 
 
03434f6
 
 
 
 
5d8d623
03434f6
89817e2
03434f6
89817e2
03434f6
 
 
 
 
 
 
 
 
 
ccc2ed2
 
 
03434f6
 
 
 
 
89817e2
03434f6
 
ccc2ed2
 
 
 
 
 
 
03434f6
ccc2ed2
 
03434f6
 
ccc2ed2
03434f6
070daeb
42f4126
 
 
 
 
 
 
 
 
2198075
070daeb
 
 
 
 
 
5969407
070daeb
 
 
 
42f4126
5969407
 
 
 
 
 
 
 
 
 
 
 
 
 
 
42f4126
5969407
 
42f4126
5969407
 
 
 
 
 
 
42f4126
 
5969407
 
 
 
 
 
 
 
 
 
 
 
 
03434f6
 
 
 
42f4126
03434f6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5969407
03434f6
42f4126
03434f6
 
 
 
 
 
5969407
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
03434f6
5969407
03434f6
 
5969407
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
03434f6
 
 
 
 
 
 
5969407
 
 
 
 
 
 
 
 
42f4126
 
 
 
03434f6
 
 
 
 
 
 
 
 
 
 
 
 
ccc2ed2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
03434f6
070daeb
 
 
 
 
 
 
42f4126
 
070daeb
42f4126
 
 
 
03434f6
42f4126
 
 
 
 
 
 
 
070daeb
42f4126
03434f6
 
 
070daeb
 
03434f6
070daeb
ccc2ed2
03434f6
070daeb
 
 
03434f6
070daeb
 
03434f6
 
 
42f4126
03434f6
 
 
 
ccc2ed2
03434f6
 
 
 
 
 
 
ccc2ed2
03434f6
 
 
 
 
 
070daeb
ccc2ed2
5969407
 
ccc2ed2
 
 
070daeb
03434f6
070daeb
 
03434f6
070daeb
 
 
 
 
 
03434f6
 
 
 
 
 
 
42f4126
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
03434f6
 
 
 
 
 
 
89817e2
 
 
 
 
 
 
03434f6
 
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
import re
import threading
import gc
import os
import torch
import time
import signal
import gradio as gr
import spaces
import transformers
from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer
from huggingface_hub import login

# ๋ชจ๋ธ ๋ฉ”๋ชจ๋ฆฌ ๊ด€๋ฆฌ ๋ฐ ์ตœ์ ํ™”๋ฅผ ์œ„ํ•œ ์„ค์ •
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DTYPE = torch.bfloat16 if torch.cuda.is_available() else torch.float32
MAX_GPU_MEMORY = 80 * 1024 * 1024 * 1024  # 80GB A100 ๊ธฐ์ค€

# ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ ๋ชฉ๋ก - ๋” ์ž‘์€ ๋ชจ๋ธ๋ถ€ํ„ฐ ์‹œ์ž‘ํ•˜๋„๋ก ๋ณ€๊ฒฝ
available_models = {
    "google/gemma-2b": "Google Gemma (2B)",  # ๋” ์ž‘์€ ๋ชจ๋ธ์„ ๊ธฐ๋ณธ์œผ๋กœ ์„ค์ •
    "mistralai/Mistral-7B-Instruct-v0.2": "Mistral 7B Instruct v0.2",
    "mistralai/Mistral-Small-3.1-24B-Base-2503": "Mistral Small 3.1 (24B)",
    "google/gemma-3-27b-it": "Google Gemma 3 (27B)",
    "Qwen/Qwen2.5-Coder-32B-Instruct": "Qwen 2.5 Coder (32B)",
    "open-r1/OlympicCoder-32B": "Olympic Coder (32B)"
}

# ๊ธฐ๋ณธ ๋ชจ๋ธ - ๊ฐ€์žฅ ์ž‘์€ ๋ชจ๋ธ๋กœ ์„ค์ •
DEFAULT_MODEL_KEY = list(available_models.keys())[0]
DEFAULT_MODEL_VALUE = available_models[DEFAULT_MODEL_KEY]

# ๋ชจ๋ธ ๋กœ๋“œ์— ์‚ฌ์šฉ๋˜๋Š” ์ „์—ญ ๋ณ€์ˆ˜
pipe = None
current_model_name = None
loading_in_progress = False

# Hugging Face ํ† ํฐ์œผ๋กœ ๋กœ๊ทธ์ธ ์‹œ๋„
try:
    hf_token = os.getenv("HF_TOKEN")
    if hf_token:
        login(token=hf_token)
        print("Hugging Face์— ์„ฑ๊ณต์ ์œผ๋กœ ๋กœ๊ทธ์ธํ–ˆ์Šต๋‹ˆ๋‹ค.")
    else:
        print("๊ฒฝ๊ณ : HF_TOKEN ํ™˜๊ฒฝ ๋ณ€์ˆ˜๊ฐ€ ์„ค์ •๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.")
except Exception as e:
    print(f"Hugging Face ๋กœ๊ทธ์ธ ์—๋Ÿฌ: {str(e)}")

# ์ตœ์ข… ๋‹ต๋ณ€์„ ๊ฐ์ง€ํ•˜๊ธฐ ์œ„ํ•œ ๋งˆ์ปค
ANSWER_MARKER = "**๋‹ต๋ณ€**"

# ๋‹จ๊ณ„๋ณ„ ์ถ”๋ก ์„ ์‹œ์ž‘ํ•˜๋Š” ๋ฌธ์žฅ๋“ค
rethink_prepends = [
    "์ž, ์ด์ œ ๋‹ค์Œ์„ ํŒŒ์•…ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค ",
    "์ œ ์ƒ๊ฐ์—๋Š” ",
    "์ž ์‹œ๋งŒ์š”, ์ œ ์ƒ๊ฐ์—๋Š” ",
    "๋‹ค์Œ ์‚ฌํ•ญ์ด ๋งž๋Š”์ง€ ํ™•์ธํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค ",
    "๋˜ํ•œ ๊ธฐ์–ตํ•ด์•ผ ํ•  ๊ฒƒ์€ ",
    "๋˜ ๋‹ค๋ฅธ ์ฃผ๋ชฉํ•  ์ ์€ ",
    "๊ทธ๋ฆฌ๊ณ  ์ €๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์‚ฌ์‹ค๋„ ๊ธฐ์–ตํ•ฉ๋‹ˆ๋‹ค ",
    "์ด์ œ ์ถฉ๋ถ„ํžˆ ์ดํ•ดํ–ˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค ",
    "์ง€๊ธˆ๊นŒ์ง€์˜ ์ •๋ณด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ, ์›๋ž˜ ์งˆ๋ฌธ์— ์‚ฌ์šฉ๋œ ์–ธ์–ด๋กœ ๋‹ต๋ณ€ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค:"
    "\n{question}\n"
    f"\n{ANSWER_MARKER}\n",
]

# ์ˆ˜์‹ ํ‘œ์‹œ ๋ฌธ์ œ ํ•ด๊ฒฐ์„ ์œ„ํ•œ ์„ค์ •
latex_delimiters = [
    {"left": "$$", "right": "$$", "display": True},
    {"left": "$", "right": "$", "display": False},
]

# ๋ชจ๋ธ ํฌ๊ธฐ ๊ธฐ๋ฐ˜ ๊ตฌ์„ฑ - ๋ชจ๋ธ ํฌ๊ธฐ์— ๋”ฐ๋ฅธ ์ตœ์  ์„ค์ • ์ •์˜
MODEL_CONFIG = {
    "small": {  # <10B
        "max_memory": {0: "10GiB"},
        "offload": False,
        "quantization": None
    },
    "medium": {  # 10B-30B
        "max_memory": {0: "30GiB"},
        "offload": False,
        "quantization": None
    },
    "large": {  # >30B
        "max_memory": {0: "60GiB"},
        "offload": True,
        "quantization": None
    }
}

def get_model_size_category(model_name):
    """๋ชจ๋ธ ํฌ๊ธฐ ์นดํ…Œ๊ณ ๋ฆฌ ๊ฒฐ์ •"""
    if "2B" in model_name or "3B" in model_name or "7B" in model_name or "8B" in model_name:
        return "small"
    elif "15B" in model_name or "24B" in model_name or "27B" in model_name:
        return "medium"
    elif "32B" in model_name or "70B" in model_name:
        return "large"
    else:
        # ๊ธฐ๋ณธ๊ฐ’์œผ๋กœ small ๋ฐ˜ํ™˜ (์•ˆ์ „์„ ์œ„ํ•ด)
        return "small"

def clear_gpu_memory():
    """GPU ๋ฉ”๋ชจ๋ฆฌ ์ •๋ฆฌ"""
    global pipe
    
    if pipe is not None:
        del pipe
        pipe = None
    
    # CUDA ์บ์‹œ ์ •๋ฆฌ
    gc.collect()
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
        torch.cuda.synchronize()

def reformat_math(text):
    """Gradio ๊ตฌ๋ฌธ(Katex)์„ ์‚ฌ์šฉํ•˜๋„๋ก MathJax ๊ตฌ๋ถ„ ๊ธฐํ˜ธ ์ˆ˜์ •."""
    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: list):
    """์‚ฌ์šฉ์ž ์ž…๋ ฅ์„ ํžˆ์Šคํ† ๋ฆฌ์— ์ถ”๊ฐ€ํ•˜๊ณ  ์ž…๋ ฅ ํ…์ŠคํŠธ ์ƒ์ž ๋น„์šฐ๊ธฐ"""
    return "", history + [
        gr.ChatMessage(role="user", content=message.replace(ANSWER_MARKER, ""))
    ]

def rebuild_messages(history: list):
    """์ค‘๊ฐ„ ์ƒ๊ฐ ๊ณผ์ • ์—†์ด ๋ชจ๋ธ์ด ์‚ฌ์šฉํ•  ํžˆ์Šคํ† ๋ฆฌ์—์„œ ๋ฉ”์‹œ์ง€ ์žฌ๊ตฌ์„ฑ"""
    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

def load_model(model_names, status_callback=None):
    """์„ ํƒ๋œ ๋ชจ๋ธ ์ด๋ฆ„์— ๋”ฐ๋ผ ๋ชจ๋ธ ๋กœ๋“œ (A100์— ์ตœ์ ํ™”๋œ ์„ค์ • ์‚ฌ์šฉ)"""
    global pipe, current_model_name, loading_in_progress
    
    # ์ด๋ฏธ ๋กœ๋”ฉ ์ค‘์ธ ๊ฒฝ์šฐ
    if loading_in_progress:
        return "๋‹ค๋ฅธ ๋ชจ๋ธ์ด ์ด๋ฏธ ๋กœ๋“œ ์ค‘์ž…๋‹ˆ๋‹ค. ์ž ์‹œ ๊ธฐ๋‹ค๋ ค์ฃผ์„ธ์š”."
    
    loading_in_progress = True
    
    try:
        # ๊ธฐ์กด ๋ชจ๋ธ ์ •๋ฆฌ
        clear_gpu_memory()
        
        # ๋ชจ๋ธ์ด ์„ ํƒ๋˜์ง€ ์•Š์•˜์„ ๊ฒฝ์šฐ ๊ธฐ๋ณธ๊ฐ’ ์ง€์ •
        if not model_names:
            model_name = DEFAULT_MODEL_KEY
        else:
            # ์ฒซ ๋ฒˆ์งธ ์„ ํƒ๋œ ๋ชจ๋ธ ์‚ฌ์šฉ
            model_name = model_names[0]
        
        # ๋ชจ๋ธ ํฌ๊ธฐ ์นดํ…Œ๊ณ ๋ฆฌ ํ™•์ธ
        size_category = get_model_size_category(model_name)
        config = MODEL_CONFIG[size_category]
        
        # ๋กœ๋”ฉ ์ƒํƒœ ์—…๋ฐ์ดํŠธ
        if status_callback:
            status_callback(f"๋ชจ๋ธ '{model_name}' ๋กœ๋“œ ์ค‘... (ํฌ๊ธฐ: {size_category})")
        
        # ๋ชจ๋ธ ๋กœ๋“œ (ํฌ๊ธฐ์— ๋”ฐ๋ผ ์ตœ์ ํ™”๋œ ์„ค์ • ์ ์šฉ)
        # HF_TOKEN ํ™˜๊ฒฝ ๋ณ€์ˆ˜ ํ™•์ธ
        hf_token = os.getenv("HF_TOKEN")
        # ๊ณตํ†ต ๋งค๊ฐœ๋ณ€์ˆ˜
        common_params = {
            "token": hf_token,  # ์ ‘๊ทผ ์ œํ•œ ๋ชจ๋ธ์„ ์œ„ํ•œ ํ† ํฐ
            "trust_remote_code": True,
        }
        
        # BitsAndBytes ์‚ฌ์šฉ ์—ฌ๋ถ€ ํ™•์ธ
        try:
            import bitsandbytes
            has_bitsandbytes = True
        except ImportError:
            has_bitsandbytes = False
            if status_callback:
                status_callback(f"BitsAndBytes ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ์ฐพ์„ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์–‘์žํ™” ์—†์ด ๋กœ๋“œํ•ฉ๋‹ˆ๋‹ค.")
        
        # ์‹œ๊ฐ„ ์ œํ•œ ์„ค์ • (๋ชจ๋ธ ํฌ๊ธฐ์— ๋”ฐ๋ผ ๋‹ค๋ฅด๊ฒŒ)
        if size_category == "small":
            load_timeout = 180  # 3๋ถ„
        elif size_category == "medium":
            load_timeout = 300  # 5๋ถ„
        else:
            load_timeout = 600  # 10๋ถ„
        
        # ๋กœ๋”ฉ ์‹œ์ž‘ ์‹œ๊ฐ„
        start_time = time.time()
        
        # ์–‘์žํ™” ์„ค์ •์ด ํ•„์š”ํ•˜๊ณ  BitsAndBytes๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒฝ์šฐ
        if config["quantization"] and has_bitsandbytes:
            # ์–‘์žํ™” ์ ์šฉ
            from transformers import BitsAndBytesConfig
            quantization_config = BitsAndBytesConfig(
                load_in_4bit=config["quantization"] == "4bit",
                bnb_4bit_compute_dtype=DTYPE
            )
            
            if status_callback:
                status_callback(f"๋ชจ๋ธ '{model_name}' ๋กœ๋“œ ์ค‘... (์–‘์žํ™” ์ ์šฉ)")
            
            model = AutoModelForCausalLM.from_pretrained(
                model_name,
                device_map="auto",
                max_memory=config["max_memory"],
                torch_dtype=DTYPE,
                quantization_config=quantization_config,
                offload_folder="offload" if config["offload"] else None,
                **common_params
            )
            tokenizer = AutoTokenizer.from_pretrained(model_name, **common_params)
            
            pipe = pipeline(
                "text-generation",
                model=model,
                tokenizer=tokenizer,
                torch_dtype=DTYPE,
                device_map="auto"
            )
        else:
            # ์–‘์žํ™” ์—†์ด ๋กœ๋“œ
            if status_callback:
                status_callback(f"๋ชจ๋ธ '{model_name}' ๋กœ๋“œ ์ค‘... (ํ‘œ์ค€ ๋ฐฉ์‹)")
            
            pipe = pipeline(
                "text-generation",
                model=model_name,
                device_map="auto",
                torch_dtype=DTYPE,
                **common_params
            )
        
        # ์‹œ๊ฐ„ ์ œํ•œ ์ดˆ๊ณผ ํ™•์ธ
        elapsed_time = time.time() - start_time
        if elapsed_time > load_timeout:
            clear_gpu_memory()
            loading_in_progress = False
            return f"๋ชจ๋ธ ๋กœ๋“œ ์‹œ๊ฐ„ ์ดˆ๊ณผ: {load_timeout}์ดˆ๊ฐ€ ์ง€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋‹ค์‹œ ์‹œ๋„ํ•˜์„ธ์š”."
        
        current_model_name = model_name
        loading_in_progress = False
        return f"๋ชจ๋ธ '{model_name}'์ด(๊ฐ€) ์„ฑ๊ณต์ ์œผ๋กœ ๋กœ๋“œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. (์ตœ์ ํ™”: {size_category}, ์†Œ์š”์‹œ๊ฐ„: {elapsed_time:.1f}์ดˆ)"
    
    except Exception as e:
        loading_in_progress = False
        return f"๋ชจ๋ธ ๋กœ๋“œ ์‹คํŒจ: {str(e)}"

@spaces.GPU
def bot(
    history: list,
    max_num_tokens: int,
    final_num_tokens: int,
    do_sample: bool,
    temperature: float,
):
    """๋ชจ๋ธ์ด ์งˆ๋ฌธ์— ๋‹ต๋ณ€ํ•˜๋„๋ก ํ•˜๊ธฐ"""
    global pipe, current_model_name
    
    # ๋ชจ๋ธ์ด ๋กœ๋“œ๋˜์ง€ ์•Š์•˜๋‹ค๋ฉด ์˜ค๋ฅ˜ ๋ฉ”์‹œ์ง€ ํ‘œ์‹œ
    if pipe is None:
        history.append(
            gr.ChatMessage(
                role="assistant",
                content="๋ชจ๋ธ์ด ๋กœ๋“œ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ํ•˜๋‚˜ ์ด์ƒ์˜ ๋ชจ๋ธ์„ ์„ ํƒํ•˜๊ณ  '๋ชจ๋ธ ๋กœ๋“œ' ๋ฒ„ํŠผ์„ ํด๋ฆญํ•ด ์ฃผ์„ธ์š”.",
            )
        )
        yield history
        return

    try:
        # ํ† ํฐ ๊ธธ์ด ์ž๋™ ์กฐ์ • (๋ชจ๋ธ ํฌ๊ธฐ์— ๋”ฐ๋ผ)
        size_category = get_model_size_category(current_model_name)
        
        # ๋Œ€ํ˜• ๋ชจ๋ธ์€ ํ† ํฐ ์ˆ˜๋ฅผ ์ค„์—ฌ ๋ฉ”๋ชจ๋ฆฌ ํšจ์œจ์„ฑ ํ–ฅ์ƒ
        if size_category == "large":
            max_num_tokens = min(max_num_tokens, 1000)
            final_num_tokens = min(final_num_tokens, 1500)
        
        # ๋‚˜์ค‘์— ์Šค๋ ˆ๋“œ์—์„œ ํ† ํฐ์„ ์ŠคํŠธ๋ฆผ์œผ๋กœ ๊ฐ€์ ธ์˜ค๊ธฐ ์œ„ํ•จ
        streamer = transformers.TextIteratorStreamer(
            pipe.tokenizer,
            skip_special_tokens=True,
            skip_prompt=True,
        )

        # ํ•„์š”ํ•œ ๊ฒฝ์šฐ ์ถ”๋ก ์— ์งˆ๋ฌธ์„ ๋‹ค์‹œ ์‚ฝ์ž…ํ•˜๊ธฐ ์œ„ํ•จ
        question = history[-1]["content"]

        # ๋ณด์กฐ์ž ๋ฉ”์‹œ์ง€ ์ค€๋น„
        history.append(
            gr.ChatMessage(
                role="assistant",
                content=str(""),
                metadata={"title": "๐Ÿง  ์ƒ๊ฐ ์ค‘...", "status": "pending"},
            )
        )

        # ํ˜„์žฌ ์ฑ„ํŒ…์— ํ‘œ์‹œ๋  ์ถ”๋ก  ๊ณผ์ •
        messages = rebuild_messages(history)
        
        # ํƒ€์ž„์•„์›ƒ ์„ค์ •
        class TimeoutError(Exception):
            pass
        
        def timeout_handler(signum, frame):
            raise TimeoutError("์š”์ฒญ ์ฒ˜๋ฆฌ ์‹œ๊ฐ„์ด ์ดˆ๊ณผ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.")
        
        # ๊ฐ ๋‹จ๊ณ„๋งˆ๋‹ค ์ตœ๋Œ€ 120์ดˆ ํƒ€์ž„์•„์›ƒ ์„ค์ •
        timeout_seconds = 120
        
        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
            )
            
            # ์Šค๋ ˆ๋“œ์—์„œ ๋ชจ๋ธ ์‹คํ–‰
            t = threading.Thread(
                target=pipe,
                args=(messages,),
                kwargs=dict(
                    max_new_tokens=num_tokens,
                    streamer=streamer,
                    do_sample=do_sample,
                    temperature=temperature,
                    # ๋ฉ”๋ชจ๋ฆฌ ํšจ์œจ์„ฑ์„ ์œ„ํ•œ ์ถ”๊ฐ€ ํŒŒ๋ผ๋ฏธํ„ฐ
                    repetition_penalty=1.2,  # ๋ฐ˜๋ณต ๋ฐฉ์ง€
                    use_cache=True,  # KV ์บ์‹œ ์‚ฌ์šฉ
                ),
            )
            t.daemon = True  # ๋ฐ๋ชฌ ์Šค๋ ˆ๋“œ๋กœ ์„ค์ •ํ•˜์—ฌ ๋ฉ”์ธ ์Šค๋ ˆ๋“œ๊ฐ€ ์ข…๋ฃŒ๋˜๋ฉด ํ•จ๊ป˜ ์ข…๋ฃŒ
            t.start()

            # ์ƒˆ ๋‚ด์šฉ์œผ๋กœ ํžˆ์Šคํ† ๋ฆฌ ์žฌ๊ตฌ์„ฑ
            history[-1].content += prepend.format(question=question)
            if ANSWER_MARKER in prepend:
                history[-1].metadata = {"title": "๐Ÿ’ญ ์‚ฌ๊ณ  ๊ณผ์ •", "status": "done"}
                # ์ƒ๊ฐ ์ข…๋ฃŒ, ์ด์ œ ๋‹ต๋ณ€์ž…๋‹ˆ๋‹ค (์ค‘๊ฐ„ ๋‹จ๊ณ„์— ๋Œ€ํ•œ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ ์—†์Œ)
                history.append(gr.ChatMessage(role="assistant", content=""))
            
            # ํƒ€์ž„์•„์›ƒ ์„ค์ • (Unix ์‹œ์Šคํ…œ์—์„œ๋งŒ ์ž‘๋™)
            try:
                if hasattr(signal, 'SIGALRM'):
                    signal.signal(signal.SIGALRM, timeout_handler)
                    signal.alarm(timeout_seconds)
                
                # ํ† ํฐ ์ŠคํŠธ๋ฆฌ๋ฐ
                token_count = 0
                for token in streamer:
                    history[-1].content += token
                    history[-1].content = reformat_math(history[-1].content)
                    token_count += 1
                    
                    # 10๊ฐœ ํ† ํฐ๋งˆ๋‹ค yield (UI ์‘๋‹ต์„ฑ ํ–ฅ์ƒ)
                    if token_count % 10 == 0:
                        yield history
                
                # ๋‚จ์€ ๋‚ด์šฉ yield
                yield history
                
                # ํƒ€์ž„์•„์›ƒ ํ•ด์ œ
                if hasattr(signal, 'SIGALRM'):
                    signal.alarm(0)
                
            except TimeoutError:
                if hasattr(signal, 'SIGALRM'):
                    signal.alarm(0)
                history[-1].content += "\n\nโš ๏ธ ์‘๋‹ต ์ƒ์„ฑ ์‹œ๊ฐ„์ด ์ดˆ๊ณผ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ ๋‹จ๊ณ„๋กœ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค."
                yield history
                continue
            
            # ์ตœ๋Œ€ 30์ดˆ ๋Œ€๊ธฐ ํ›„ ๋‹ค์Œ ๋‹จ๊ณ„๋กœ ์ง„ํ–‰
            join_start_time = time.time()
            while t.is_alive() and (time.time() - join_start_time) < 30:
                t.join(1)  # 1์ดˆ๋งˆ๋‹ค ํ™•์ธ
            
            # ์Šค๋ ˆ๋“œ๊ฐ€ ์—ฌ์ „ํžˆ ์‹คํ–‰ ์ค‘์ด๋ฉด ๊ฐ•์ œ ์ง„ํ–‰
            if t.is_alive():
                history[-1].content += "\n\nโš ๏ธ ์‘๋‹ต ์ƒ์„ฑ์ด ์˜ˆ์ƒ๋ณด๋‹ค ์˜ค๋ž˜ ๊ฑธ๋ฆฝ๋‹ˆ๋‹ค. ๋‹ค์Œ ๋‹จ๊ณ„๋กœ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค."
                yield history
            
            # ๋Œ€ํ˜• ๋ชจ๋ธ์ธ ๊ฒฝ์šฐ ๊ฐ ๋‹จ๊ณ„ ํ›„ ๋ถ€๋ถ„์  ๋ฉ”๋ชจ๋ฆฌ ์ •๋ฆฌ
            if size_category == "large" and torch.cuda.is_available():
                torch.cuda.empty_cache()
    
    except Exception as e:
        # ์˜ค๋ฅ˜ ๋ฐœ์ƒ์‹œ ์‚ฌ์šฉ์ž์—๊ฒŒ ์•Œ๋ฆผ
        import traceback
        error_msg = f"\n\nโš ๏ธ ์ฒ˜๋ฆฌ ์ค‘ ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ–ˆ์Šต๋‹ˆ๋‹ค: {str(e)}\n{traceback.format_exc()}"
        
        if len(history) > 0 and isinstance(history[-1], gr.ChatMessage) and history[-1].role == "assistant":
            history[-1].content += error_msg
        else:
            history.append(gr.ChatMessage(role="assistant", content=error_msg))
        
        yield history

    yield history


# ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ GPU ์ •๋ณด ํ‘œ์‹œ ํ•จ์ˆ˜
def get_gpu_info():
    if not torch.cuda.is_available():
        return "GPU๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค."
    
    gpu_info = []
    for i in range(torch.cuda.device_count()):
        gpu_name = torch.cuda.get_device_name(i)
        total_memory = torch.cuda.get_device_properties(i).total_memory / 1024**3
        gpu_info.append(f"GPU {i}: {gpu_name} ({total_memory:.1f} GB)")
    
    return "\n".join(gpu_info)

# ์ž๋™ ๋ชจ๋ธ ๋กœ๋“œ ํ•จ์ˆ˜ (์ƒํƒœ ์—…๋ฐ์ดํŠธ ํฌํ•จ)
def auto_load_model():
    # ์ฒซ ๋ฒˆ์งธ ๋ชจ๋ธ ์ž๋™ ๋กœ๋“œ
    model_key = DEFAULT_MODEL_KEY
    try:
        # ์ง„ํ–‰ ์ƒํƒœ ํ‘œ์‹œ๋ฅผ ์œ„ํ•œ ๋นˆ ๊ฒฐ๊ณผ ๋ฐ˜ํ™˜
        return "์ž‘์€ ๋ชจ๋ธ ์ž๋™ ๋กœ๋“œ ์ค‘... ์ž ์‹œ ๊ธฐ๋‹ค๋ ค์ฃผ์„ธ์š”."
    except Exception as e:
        return f"์ž๋™ ๋ชจ๋ธ ๋กœ๋“œ ์‹คํŒจ: {str(e)}"

# ์‹ค์ œ ๋ชจ๋ธ ๋กœ๋“œ ํ•จ์ˆ˜ (๋น„๋™๊ธฐ)
def load_model_async(model_status):
    # ๋น„๋™๊ธฐ ํ•จ์ˆ˜๋กœ ๋ชจ๋ธ ๋กœ๋“œ (์‹ค์ œ ๋กœ๋“œ๋Š” ๋ฐฑ๊ทธ๋ผ์šด๋“œ์—์„œ ์ˆ˜ํ–‰)
    model_key = DEFAULT_MODEL_KEY
    
    def update_status(status):
        model_status.update(value=status)
    
    # ๋ณ„๋„ ์Šค๋ ˆ๋“œ์—์„œ ๋กœ๋“œ
    def load_in_thread():
        try:
            result = load_model([model_key], update_status)
            model_status.update(value=result)
        except Exception as e:
            model_status.update(value=f"๋ชจ๋ธ ๋กœ๋“œ ์‹คํŒจ: {str(e)}")
    
    threading.Thread(target=load_in_thread, daemon=True).start()
    return "๋ชจ๋ธ ๋กœ๋“œ ์ค€๋น„ ์ค‘... ์ž๋™์œผ๋กœ ์ง„ํ–‰๋ฉ๋‹ˆ๋‹ค."

# Gradio ์ธํ„ฐํŽ˜์ด์Šค
with gr.Blocks(fill_height=True, title="ThinkFlow - Step-by-step Reasoning Service") as demo:
    # ์ƒ๋‹จ์— ํƒ€์ดํ‹€๊ณผ ์„ค๋ช… ์ถ”๊ฐ€
    gr.Markdown("""
    # ThinkFlow
    ## A thought amplification service that implants step-by-step reasoning abilities into LLMs without model modification
    """)
    
    with gr.Row(scale=1):
        with gr.Column(scale=5):
            # ์ฑ„ํŒ… ์ธํ„ฐํŽ˜์ด์Šค
            chatbot = gr.Chatbot(
                scale=1,
                type="messages",
                latex_delimiters=latex_delimiters,
                height=600,
            )
            msg = gr.Textbox(
                submit_btn=True,
                label="",
                show_label=False,
                placeholder="์—ฌ๊ธฐ์— ์งˆ๋ฌธ์„ ์ž…๋ ฅํ•˜์„ธ์š”.",
                autofocus=True,
            )
        
        with gr.Column(scale=1):
            # ํ•˜๋“œ์›จ์–ด ์ •๋ณด ํ‘œ์‹œ
            gpu_info = gr.Markdown(f"**์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ํ•˜๋“œ์›จ์–ด:**\n{get_gpu_info()}")
            
            # ๋ชจ๋ธ ์„ ํƒ ์„น์…˜ ์ถ”๊ฐ€
            gr.Markdown("""## ๋ชจ๋ธ ์„ ํƒ""")
            model_selector = gr.Radio(
                choices=list(available_models.values()),
                value=DEFAULT_MODEL_VALUE,
                label="์‚ฌ์šฉํ•  LLM ๋ชจ๋ธ ์„ ํƒ",
            )
            
            # ๋ชจ๋ธ ๋กœ๋“œ ๋ฒ„ํŠผ
            load_model_btn = gr.Button("๋ชจ๋ธ ๋กœ๋“œ", variant="primary")
            model_status = gr.Textbox(label="๋ชจ๋ธ ์ƒํƒœ", interactive=False)
            
            # ๋ฉ”๋ชจ๋ฆฌ ์ •๋ฆฌ ๋ฒ„ํŠผ
            clear_memory_btn = gr.Button("GPU ๋ฉ”๋ชจ๋ฆฌ ์ •๋ฆฌ", variant="secondary")
            
            gr.Markdown("""## ๋งค๊ฐœ๋ณ€์ˆ˜ ์กฐ์ •""")
            with gr.Accordion("๊ณ ๊ธ‰ ์„ค์ •", open=False):
                num_tokens = gr.Slider(
                    50,
                    2000,
                    1000,
                    step=50,
                    label="์ถ”๋ก  ๋‹จ๊ณ„๋‹น ์ตœ๋Œ€ ํ† ํฐ ์ˆ˜",
                    interactive=True,
                )
                final_num_tokens = gr.Slider(
                    50,
                    3000,
                    1500,
                    step=50,
                    label="์ตœ์ข… ๋‹ต๋ณ€์˜ ์ตœ๋Œ€ ํ† ํฐ ์ˆ˜",
                    interactive=True,
                )
                do_sample = gr.Checkbox(True, label="์ƒ˜ํ”Œ๋ง ์‚ฌ์šฉ")
                temperature = gr.Slider(0.1, 1.0, 0.7, step=0.1, label="์˜จ๋„")
    
    # ์‹œ์ž‘ ์‹œ ์ž๋™์œผ๋กœ ์ดˆ๊ธฐํ™”
    demo.load(auto_load_model, [], [model_status])
    
    # ์‹œ์ž‘ ํ›„ ๋น„๋™๊ธฐ์ ์œผ๋กœ ๋ชจ๋ธ ๋กœ๋“œ (์ดˆ๊ธฐ ํ™”๋ฉด ํ‘œ์‹œ ์ง€์—ฐ ๋ฐฉ์ง€)
    demo.load(lambda x: load_model_async(x), [model_status], [], _js="() => {}")
    
    # ์„ ํƒ๋œ ๋ชจ๋ธ ๋กœ๋“œ ์ด๋ฒคํŠธ ์—ฐ๊ฒฐ
    def get_model_names(selected_model):
        # ํ‘œ์‹œ ์ด๋ฆ„์—์„œ ์›๋ž˜ ๋ชจ๋ธ ์ด๋ฆ„์œผ๋กœ ๋ณ€ํ™˜
        inverse_map = {v: k for k, v in available_models.items()}
        return [inverse_map[selected_model]] if selected_model else []
    
    load_model_btn.click(
        lambda selected: load_model(get_model_names(selected)),
        inputs=[model_selector],
        outputs=[model_status]
    )
    
    # GPU ๋ฉ”๋ชจ๋ฆฌ ์ •๋ฆฌ ์ด๋ฒคํŠธ ์—ฐ๊ฒฐ
    clear_memory_btn.click(
        lambda: (clear_gpu_memory(), "GPU ๋ฉ”๋ชจ๋ฆฌ๊ฐ€ ์ •๋ฆฌ๋˜์—ˆ์Šต๋‹ˆ๋‹ค."),
        inputs=[],
        outputs=[model_status]
    )

    # ์‚ฌ์šฉ์ž๊ฐ€ ๋ฉ”์‹œ์ง€๋ฅผ ์ œ์ถœํ•˜๋ฉด ๋ด‡์ด ์‘๋‹ตํ•ฉ๋‹ˆ๋‹ค
    msg.submit(
        user_input,
        [msg, chatbot],  # ์ž…๋ ฅ
        [msg, chatbot],  # ์ถœ๋ ฅ
    ).then(
        bot,
        [
            chatbot,
            num_tokens,
            final_num_tokens,
            do_sample,
            temperature,
        ],  # ์‹ค์ œ๋กœ๋Š” "history" ์ž…๋ ฅ
        chatbot,  # ์ถœ๋ ฅ์—์„œ ์ƒˆ ํžˆ์Šคํ† ๋ฆฌ ์ €์žฅ
    )

if __name__ == "__main__":
    # ๋””๋ฒ„๊น… ์ •๋ณด ์ถœ๋ ฅ
    print(f"GPU ์‚ฌ์šฉ ๊ฐ€๋Šฅ: {torch.cuda.is_available()}")
    if torch.cuda.is_available():
        print(f"์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ GPU ๊ฐœ์ˆ˜: {torch.cuda.device_count()}")
        print(f"ํ˜„์žฌ GPU: {torch.cuda.current_device()}")
        print(f"GPU ์ด๋ฆ„: {torch.cuda.get_device_name(0)}")
    
    # HF_TOKEN ํ™˜๊ฒฝ ๋ณ€์ˆ˜ ํ™•์ธ
    hf_token = os.getenv("HF_TOKEN")
    if hf_token:
        print("HF_TOKEN ํ™˜๊ฒฝ ๋ณ€์ˆ˜๊ฐ€ ์„ค์ •๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.")
    else:
        print("๊ฒฝ๊ณ : HF_TOKEN ํ™˜๊ฒฝ ๋ณ€์ˆ˜๊ฐ€ ์„ค์ •๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ œํ•œ๋œ ๋ชจ๋ธ์— ์ ‘๊ทผํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.")
    
    # ํ ์‚ฌ์šฉ ๋ฐ ์•ฑ ์‹คํ–‰
    demo.queue(max_size=10).launch()