Update app.py
Browse files
app.py
CHANGED
|
@@ -1,65 +1,82 @@
|
|
| 1 |
-
import sys
|
| 2 |
-
import subprocess
|
| 3 |
-
import os
|
| 4 |
-
|
| 5 |
-
# First, try to install all dependencies
|
| 6 |
-
packages_to_install = [
|
| 7 |
-
"gradio",
|
| 8 |
-
"torch",
|
| 9 |
-
"transformers",
|
| 10 |
-
"accelerate",
|
| 11 |
-
"einops",
|
| 12 |
-
"timm",
|
| 13 |
-
"av",
|
| 14 |
-
"opencv-python-headless" # Using headless version for better compatibility
|
| 15 |
-
]
|
| 16 |
-
|
| 17 |
-
for package in packages_to_install:
|
| 18 |
-
print(f"Installing {package}...")
|
| 19 |
-
try:
|
| 20 |
-
subprocess.check_call([sys.executable, "-m", "pip", "install", package])
|
| 21 |
-
print(f"Successfully installed {package}")
|
| 22 |
-
except Exception as e:
|
| 23 |
-
print(f"Error installing {package}: {e}")
|
| 24 |
-
|
| 25 |
-
# Now proceed with the actual application
|
| 26 |
import gradio as gr
|
| 27 |
import torch
|
| 28 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 29 |
import gc
|
|
|
|
| 30 |
import datetime
|
| 31 |
import time
|
| 32 |
-
import spaces
|
| 33 |
|
| 34 |
-
# ---
|
| 35 |
-
MODEL_ID = "naver-hyperclovax/HyperCLOVAX-SEED-
|
| 36 |
MAX_NEW_TOKENS = 512
|
|
|
|
| 37 |
|
| 38 |
# Hugging Face 토큰 설정 - 환경 변수에서 가져오기
|
| 39 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
| 40 |
if not HF_TOKEN:
|
| 41 |
print("경고: HF_TOKEN 환경 변수가 설정되지 않았습니다. 비공개 모델에 접근할 수 없을 수 있습니다.")
|
| 42 |
|
| 43 |
-
# ---
|
| 44 |
-
print("---
|
| 45 |
-
|
| 46 |
-
print(f"
|
|
|
|
|
|
|
| 47 |
print(f"HF_TOKEN 설정 여부: {'있음' if HF_TOKEN else '없음'}")
|
| 48 |
|
| 49 |
-
#
|
| 50 |
-
|
| 51 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 52 |
|
| 53 |
model = None
|
| 54 |
tokenizer = None
|
| 55 |
load_successful = False
|
| 56 |
-
stop_token_ids_list = []
|
| 57 |
|
| 58 |
try:
|
| 59 |
start_load_time = time.time()
|
| 60 |
-
# 자원에 따라 device_map 설정
|
| 61 |
-
device_map = "auto" if torch.cuda.is_available() else "cpu"
|
| 62 |
-
dtype = torch.float16 if torch.cuda.is_available() else torch.float32
|
| 63 |
|
| 64 |
# 토크나이저 로딩
|
| 65 |
tokenizer_kwargs = {
|
|
@@ -69,7 +86,7 @@ try:
|
|
| 69 |
# HF_TOKEN이 설정되어 있으면 추가
|
| 70 |
if HF_TOKEN:
|
| 71 |
tokenizer_kwargs["token"] = HF_TOKEN
|
| 72 |
-
|
| 73 |
tokenizer = AutoTokenizer.from_pretrained(
|
| 74 |
MODEL_ID,
|
| 75 |
**tokenizer_kwargs
|
|
@@ -77,15 +94,15 @@ try:
|
|
| 77 |
|
| 78 |
# 모델 로딩
|
| 79 |
model_kwargs = {
|
| 80 |
-
"
|
| 81 |
-
"device_map":
|
| 82 |
-
"
|
| 83 |
}
|
| 84 |
|
| 85 |
# HF_TOKEN이 설정되어 있으면 추가
|
| 86 |
if HF_TOKEN:
|
| 87 |
model_kwargs["token"] = HF_TOKEN
|
| 88 |
-
|
| 89 |
model = AutoModelForCausalLM.from_pretrained(
|
| 90 |
MODEL_ID,
|
| 91 |
**model_kwargs
|
|
@@ -93,74 +110,80 @@ try:
|
|
| 93 |
|
| 94 |
model.eval()
|
| 95 |
load_time = time.time() - start_load_time
|
| 96 |
-
print(f"---
|
| 97 |
load_successful = True
|
| 98 |
|
| 99 |
-
# ---
|
| 100 |
-
stop_token_strings = ["
|
| 101 |
temp_stop_ids = [tokenizer.convert_tokens_to_ids(token) for token in stop_token_strings]
|
| 102 |
|
| 103 |
if tokenizer.eos_token_id is not None and tokenizer.eos_token_id not in temp_stop_ids:
|
| 104 |
temp_stop_ids.append(tokenizer.eos_token_id)
|
| 105 |
elif tokenizer.eos_token_id is None:
|
| 106 |
-
print("
|
| 107 |
|
| 108 |
stop_token_ids_list = [tid for tid in temp_stop_ids if tid is not None]
|
| 109 |
|
| 110 |
if not stop_token_ids_list:
|
| 111 |
-
print("
|
| 112 |
if tokenizer.eos_token_id is not None:
|
| 113 |
stop_token_ids_list = [tokenizer.eos_token_id]
|
| 114 |
else:
|
| 115 |
-
print("
|
| 116 |
|
| 117 |
-
print(f"
|
| 118 |
|
| 119 |
except Exception as e:
|
| 120 |
-
print(f"!!!
|
| 121 |
if 'model' in locals() and model is not None: del model
|
| 122 |
if 'tokenizer' in locals() and tokenizer is not None: del tokenizer
|
| 123 |
gc.collect()
|
| 124 |
-
|
|
|
|
| 125 |
|
| 126 |
-
# ---
|
| 127 |
def get_system_prompt():
|
| 128 |
current_date = datetime.datetime.now().strftime("%Y-%m-%d (%A)")
|
| 129 |
return (
|
| 130 |
-
f"-
|
| 131 |
-
f"-
|
|
|
|
| 132 |
)
|
| 133 |
|
| 134 |
-
# ---
|
| 135 |
def warmup_model():
|
| 136 |
if not load_successful or model is None or tokenizer is None:
|
| 137 |
-
print("
|
| 138 |
return
|
| 139 |
|
| 140 |
-
print("---
|
| 141 |
try:
|
| 142 |
start_warmup_time = time.time()
|
| 143 |
warmup_message = "안녕하세요"
|
| 144 |
-
|
| 145 |
-
# 모델에 맞는 형식으로 입력 구성
|
| 146 |
system_prompt = get_system_prompt()
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 154 |
gen_kwargs = {
|
| 155 |
"max_new_tokens": 10,
|
| 156 |
"pad_token_id": tokenizer.eos_token_id if tokenizer.eos_token_id is not None else tokenizer.pad_token_id,
|
| 157 |
"do_sample": False
|
| 158 |
}
|
| 159 |
-
|
| 160 |
if stop_token_ids_list:
|
| 161 |
gen_kwargs["eos_token_id"] = stop_token_ids_list
|
| 162 |
else:
|
| 163 |
-
print("
|
| 164 |
|
| 165 |
with torch.no_grad():
|
| 166 |
output_ids = model.generate(**inputs, **gen_kwargs)
|
|
@@ -169,132 +192,176 @@ def warmup_model():
|
|
| 169 |
del output_ids
|
| 170 |
gc.collect()
|
| 171 |
warmup_time = time.time() - start_warmup_time
|
| 172 |
-
print(f"---
|
| 173 |
|
| 174 |
except Exception as e:
|
| 175 |
-
print(f"!!!
|
| 176 |
finally:
|
| 177 |
gc.collect()
|
| 178 |
|
| 179 |
-
# ---
|
| 180 |
-
@spaces.GPU()
|
| 181 |
def predict(message, history):
|
| 182 |
"""
|
| 183 |
-
|
| 184 |
-
'history'
|
| 185 |
"""
|
| 186 |
if model is None or tokenizer is None:
|
| 187 |
return "오류: 모델이 로드되지 않았습니다."
|
| 188 |
|
| 189 |
-
|
| 190 |
-
history_text = ""
|
| 191 |
-
if isinstance(history, list):
|
| 192 |
-
for turn in history:
|
| 193 |
-
if isinstance(turn, tuple) and len(turn) == 2:
|
| 194 |
-
history_text += f"Human: {turn[0]}\nAssistant: {turn[1]}\n"
|
| 195 |
|
| 196 |
-
#
|
| 197 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 198 |
|
| 199 |
inputs = None
|
| 200 |
output_ids = None
|
| 201 |
|
| 202 |
try:
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 207 |
|
| 208 |
except Exception as e:
|
| 209 |
-
print(f"!!!
|
| 210 |
return f"오류: 입력 형식을 처리하는 중 문제가 발생했습니다. ({e})"
|
| 211 |
|
| 212 |
try:
|
| 213 |
-
print("
|
| 214 |
generation_start_time = time.time()
|
| 215 |
|
| 216 |
-
#
|
| 217 |
gen_kwargs = {
|
| 218 |
"max_new_tokens": MAX_NEW_TOKENS,
|
| 219 |
"pad_token_id": tokenizer.eos_token_id if tokenizer.eos_token_id is not None else tokenizer.pad_token_id,
|
| 220 |
"do_sample": True,
|
| 221 |
"temperature": 0.7,
|
| 222 |
"top_p": 0.9,
|
| 223 |
-
"repetition_penalty": 1.1
|
| 224 |
}
|
| 225 |
-
|
| 226 |
if stop_token_ids_list:
|
| 227 |
gen_kwargs["eos_token_id"] = stop_token_ids_list
|
| 228 |
else:
|
| 229 |
-
print("
|
| 230 |
|
| 231 |
with torch.no_grad():
|
| 232 |
output_ids = model.generate(**inputs, **gen_kwargs)
|
| 233 |
|
| 234 |
generation_time = time.time() - generation_start_time
|
| 235 |
-
print(f"
|
| 236 |
|
| 237 |
except Exception as e:
|
| 238 |
-
print(f"!!!
|
| 239 |
if inputs is not None: del inputs
|
| 240 |
if output_ids is not None: del output_ids
|
| 241 |
gc.collect()
|
| 242 |
return f"오류: 응답을 생성하는 중 문제가 발생했습니다. ({e})"
|
| 243 |
|
| 244 |
-
#
|
| 245 |
response = "오류: 응답 생성에 실패했습니다."
|
| 246 |
if output_ids is not None:
|
| 247 |
try:
|
| 248 |
new_tokens = output_ids[0, input_length:]
|
| 249 |
response = tokenizer.decode(new_tokens, skip_special_tokens=True)
|
| 250 |
-
print(f"
|
| 251 |
del new_tokens
|
| 252 |
except Exception as e:
|
| 253 |
-
print(f"!!!
|
| 254 |
response = "오류: 응답을 디코딩하는 중 문제가 발생했습니다."
|
| 255 |
|
| 256 |
-
#
|
| 257 |
if inputs is not None: del inputs
|
| 258 |
if output_ids is not None: del output_ids
|
| 259 |
gc.collect()
|
| 260 |
-
print("
|
| 261 |
-
|
| 262 |
-
return response
|
| 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 |
if __name__ == "__main__":
|
| 291 |
if load_successful:
|
| 292 |
warmup_model()
|
| 293 |
else:
|
| 294 |
-
print("
|
| 295 |
|
| 296 |
-
print("--- Gradio
|
| 297 |
demo.queue().launch(
|
| 298 |
-
# share=True #
|
| 299 |
-
# server_name="0.0.0.0" #
|
| 300 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
import torch
|
| 3 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 4 |
import gc
|
| 5 |
+
import os
|
| 6 |
import datetime
|
| 7 |
import time
|
|
|
|
| 8 |
|
| 9 |
+
# --- Configuration ---
|
| 10 |
+
MODEL_ID = "naver-hyperclovax/HyperCLOVAX-SEED-Text-Instruct-0.5B"
|
| 11 |
MAX_NEW_TOKENS = 512
|
| 12 |
+
USE_GPU = True # Enable GPU usage
|
| 13 |
|
| 14 |
# Hugging Face 토큰 설정 - 환경 변수에서 가져오기
|
| 15 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
| 16 |
if not HF_TOKEN:
|
| 17 |
print("경고: HF_TOKEN 환경 변수가 설정되지 않았습니다. 비공개 모델에 접근할 수 없을 수 있습니다.")
|
| 18 |
|
| 19 |
+
# --- Environment setup ---
|
| 20 |
+
print("--- Environment Setup ---")
|
| 21 |
+
device = torch.device("cuda" if torch.cuda.is_available() and USE_GPU else "cpu")
|
| 22 |
+
print(f"PyTorch version: {torch.__version__}")
|
| 23 |
+
print(f"Running on device: {device}")
|
| 24 |
+
print(f"Torch Threads: {torch.get_num_threads()}")
|
| 25 |
print(f"HF_TOKEN 설정 여부: {'있음' if HF_TOKEN else '없음'}")
|
| 26 |
|
| 27 |
+
# Custom CSS for improved UI
|
| 28 |
+
custom_css = """
|
| 29 |
+
.gradio-container {
|
| 30 |
+
max-width: 850px !important;
|
| 31 |
+
margin: auto;
|
| 32 |
+
}
|
| 33 |
+
.gr-chat {
|
| 34 |
+
border-radius: 10px;
|
| 35 |
+
box-shadow: 0 4px 8px rgba(0, 0, 0, 0.1);
|
| 36 |
+
}
|
| 37 |
+
.user-message {
|
| 38 |
+
background-color: #f0f7ff !important;
|
| 39 |
+
border-radius: 8px;
|
| 40 |
+
}
|
| 41 |
+
.assistant-message {
|
| 42 |
+
background-color: #f9f9f9 !important;
|
| 43 |
+
border-radius: 8px;
|
| 44 |
+
}
|
| 45 |
+
.gr-button.primary-button {
|
| 46 |
+
background-color: #1f4e79 !important;
|
| 47 |
+
}
|
| 48 |
+
.gr-form {
|
| 49 |
+
padding: 20px;
|
| 50 |
+
border-radius: 10px;
|
| 51 |
+
box-shadow: 0 2px 6px rgba(0, 0, 0, 0.05);
|
| 52 |
+
}
|
| 53 |
+
#intro-message {
|
| 54 |
+
text-align: center;
|
| 55 |
+
margin-bottom: 20px;
|
| 56 |
+
padding: 15px;
|
| 57 |
+
background: linear-gradient(135deg, #e8f4ff 0%, #f0f7ff 100%);
|
| 58 |
+
border-radius: 10px;
|
| 59 |
+
border-left: 4px solid #1f4e79;
|
| 60 |
+
}
|
| 61 |
+
.footer {
|
| 62 |
+
text-align: center;
|
| 63 |
+
margin-top: 20px;
|
| 64 |
+
font-size: 0.8em;
|
| 65 |
+
color: #666;
|
| 66 |
+
}
|
| 67 |
+
"""
|
| 68 |
+
|
| 69 |
+
# --- Model and Tokenizer Loading ---
|
| 70 |
+
print(f"--- Loading Model: {MODEL_ID} ---")
|
| 71 |
+
print("This might take a few minutes, especially on the first launch...")
|
| 72 |
|
| 73 |
model = None
|
| 74 |
tokenizer = None
|
| 75 |
load_successful = False
|
| 76 |
+
stop_token_ids_list = [] # Initialize stop_token_ids_list
|
| 77 |
|
| 78 |
try:
|
| 79 |
start_load_time = time.time()
|
|
|
|
|
|
|
|
|
|
| 80 |
|
| 81 |
# 토크나이저 로딩
|
| 82 |
tokenizer_kwargs = {
|
|
|
|
| 86 |
# HF_TOKEN이 설정되어 있으면 추가
|
| 87 |
if HF_TOKEN:
|
| 88 |
tokenizer_kwargs["token"] = HF_TOKEN
|
| 89 |
+
|
| 90 |
tokenizer = AutoTokenizer.from_pretrained(
|
| 91 |
MODEL_ID,
|
| 92 |
**tokenizer_kwargs
|
|
|
|
| 94 |
|
| 95 |
# 모델 로딩
|
| 96 |
model_kwargs = {
|
| 97 |
+
"trust_remote_code": True,
|
| 98 |
+
"device_map": "auto" if device.type == "cuda" else "cpu",
|
| 99 |
+
"torch_dtype": torch.float16 if device.type == "cuda" else torch.float32,
|
| 100 |
}
|
| 101 |
|
| 102 |
# HF_TOKEN이 설정되어 있으면 추가
|
| 103 |
if HF_TOKEN:
|
| 104 |
model_kwargs["token"] = HF_TOKEN
|
| 105 |
+
|
| 106 |
model = AutoModelForCausalLM.from_pretrained(
|
| 107 |
MODEL_ID,
|
| 108 |
**model_kwargs
|
|
|
|
| 110 |
|
| 111 |
model.eval()
|
| 112 |
load_time = time.time() - start_load_time
|
| 113 |
+
print(f"--- Model and Tokenizer Loaded Successfully in {load_time:.2f} seconds ---")
|
| 114 |
load_successful = True
|
| 115 |
|
| 116 |
+
# --- Stop Token Configuration ---
|
| 117 |
+
stop_token_strings = ["<|endofturn|>", "<|stop|>"]
|
| 118 |
temp_stop_ids = [tokenizer.convert_tokens_to_ids(token) for token in stop_token_strings]
|
| 119 |
|
| 120 |
if tokenizer.eos_token_id is not None and tokenizer.eos_token_id not in temp_stop_ids:
|
| 121 |
temp_stop_ids.append(tokenizer.eos_token_id)
|
| 122 |
elif tokenizer.eos_token_id is None:
|
| 123 |
+
print("Warning: tokenizer.eos_token_id is None. Cannot add to stop tokens.")
|
| 124 |
|
| 125 |
stop_token_ids_list = [tid for tid in temp_stop_ids if tid is not None]
|
| 126 |
|
| 127 |
if not stop_token_ids_list:
|
| 128 |
+
print("Warning: Could not find any stop token IDs. Using default EOS if available, otherwise generation might not stop correctly.")
|
| 129 |
if tokenizer.eos_token_id is not None:
|
| 130 |
stop_token_ids_list = [tokenizer.eos_token_id]
|
| 131 |
else:
|
| 132 |
+
print("Error: No stop tokens found, including default EOS. Generation may run indefinitely.")
|
| 133 |
|
| 134 |
+
print(f"Using Stop Token IDs: {stop_token_ids_list}")
|
| 135 |
|
| 136 |
except Exception as e:
|
| 137 |
+
print(f"!!! Error loading model: {e}")
|
| 138 |
if 'model' in locals() and model is not None: del model
|
| 139 |
if 'tokenizer' in locals() and tokenizer is not None: del tokenizer
|
| 140 |
gc.collect()
|
| 141 |
+
# Raise Gradio error to display in the Space UI if loading fails
|
| 142 |
+
raise gr.Error(f"Failed to load the model {MODEL_ID}. Cannot start the application. Error: {e}")
|
| 143 |
|
| 144 |
+
# --- System Prompt Definition ---
|
| 145 |
def get_system_prompt():
|
| 146 |
current_date = datetime.datetime.now().strftime("%Y-%m-%d (%A)")
|
| 147 |
return (
|
| 148 |
+
f"- AI 언어모델의 이름은 \"CLOVA X\" 이며 네이버에서 만들었다.\n"
|
| 149 |
+
f"- 오늘은 {current_date}이다.\n"
|
| 150 |
+
f"- 사용자의 질문에 대해 친절하고 자세하게 한국어로 답변해야 한다."
|
| 151 |
)
|
| 152 |
|
| 153 |
+
# --- Warm-up Function ---
|
| 154 |
def warmup_model():
|
| 155 |
if not load_successful or model is None or tokenizer is None:
|
| 156 |
+
print("Skipping warmup: Model not loaded successfully.")
|
| 157 |
return
|
| 158 |
|
| 159 |
+
print("--- Starting Model Warm-up ---")
|
| 160 |
try:
|
| 161 |
start_warmup_time = time.time()
|
| 162 |
warmup_message = "안녕하세요"
|
|
|
|
|
|
|
| 163 |
system_prompt = get_system_prompt()
|
| 164 |
+
warmup_chat = [
|
| 165 |
+
{"role": "tool_list", "content": ""},
|
| 166 |
+
{"role": "system", "content": system_prompt},
|
| 167 |
+
{"role": "user", "content": warmup_message}
|
| 168 |
+
]
|
| 169 |
+
|
| 170 |
+
inputs = tokenizer.apply_chat_template(
|
| 171 |
+
warmup_chat,
|
| 172 |
+
add_generation_prompt=True,
|
| 173 |
+
return_dict=True,
|
| 174 |
+
return_tensors="pt"
|
| 175 |
+
).to(device)
|
| 176 |
+
|
| 177 |
+
# Check if stop_token_ids_list is empty and handle appropriately
|
| 178 |
gen_kwargs = {
|
| 179 |
"max_new_tokens": 10,
|
| 180 |
"pad_token_id": tokenizer.eos_token_id if tokenizer.eos_token_id is not None else tokenizer.pad_token_id,
|
| 181 |
"do_sample": False
|
| 182 |
}
|
|
|
|
| 183 |
if stop_token_ids_list:
|
| 184 |
gen_kwargs["eos_token_id"] = stop_token_ids_list
|
| 185 |
else:
|
| 186 |
+
print("Warmup Warning: No stop tokens defined for generation.")
|
| 187 |
|
| 188 |
with torch.no_grad():
|
| 189 |
output_ids = model.generate(**inputs, **gen_kwargs)
|
|
|
|
| 192 |
del output_ids
|
| 193 |
gc.collect()
|
| 194 |
warmup_time = time.time() - start_warmup_time
|
| 195 |
+
print(f"--- Model Warm-up Completed in {warmup_time:.2f} seconds ---")
|
| 196 |
|
| 197 |
except Exception as e:
|
| 198 |
+
print(f"!!! Error during model warm-up: {e}")
|
| 199 |
finally:
|
| 200 |
gc.collect()
|
| 201 |
|
| 202 |
+
# --- Inference Function ---
|
|
|
|
| 203 |
def predict(message, history):
|
| 204 |
"""
|
| 205 |
+
Generates response using HyperCLOVAX.
|
| 206 |
+
Assumes 'history' is in the Gradio 'messages' format: List[Dict].
|
| 207 |
"""
|
| 208 |
if model is None or tokenizer is None:
|
| 209 |
return "오류: 모델이 로드되지 않았습니다."
|
| 210 |
|
| 211 |
+
system_prompt = get_system_prompt()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 212 |
|
| 213 |
+
# Start with system prompt
|
| 214 |
+
chat_history_formatted = [
|
| 215 |
+
{"role": "tool_list", "content": ""}, # As required by model card
|
| 216 |
+
{"role": "system", "content": system_prompt}
|
| 217 |
+
]
|
| 218 |
+
|
| 219 |
+
# Append history (List of {'role': 'user'/'assistant', 'content': '...'})
|
| 220 |
+
if isinstance(history, list): # Check if history is a list
|
| 221 |
+
for turn in history:
|
| 222 |
+
# Validate turn format
|
| 223 |
+
if isinstance(turn, dict) and "role" in turn and "content" in turn:
|
| 224 |
+
chat_history_formatted.append(turn)
|
| 225 |
+
# Handle potential older tuple format
|
| 226 |
+
elif isinstance(turn, (list, tuple)) and len(turn) == 2:
|
| 227 |
+
print(f"Warning: Received history item in tuple format: {turn}. Converting to messages format.")
|
| 228 |
+
chat_history_formatted.append({"role": "user", "content": turn[0]})
|
| 229 |
+
if turn[1]: # Ensure assistant message exists
|
| 230 |
+
chat_history_formatted.append({"role": "assistant", "content": turn[1]})
|
| 231 |
+
else:
|
| 232 |
+
print(f"Warning: Skipping unexpected history format item: {turn}")
|
| 233 |
+
|
| 234 |
+
# Append the latest user message
|
| 235 |
+
chat_history_formatted.append({"role": "user", "content": message})
|
| 236 |
|
| 237 |
inputs = None
|
| 238 |
output_ids = None
|
| 239 |
|
| 240 |
try:
|
| 241 |
+
inputs = tokenizer.apply_chat_template(
|
| 242 |
+
chat_history_formatted,
|
| 243 |
+
add_generation_prompt=True,
|
| 244 |
+
return_dict=True,
|
| 245 |
+
return_tensors="pt"
|
| 246 |
+
).to(device)
|
| 247 |
+
input_length = inputs['input_ids'].shape[1]
|
| 248 |
+
print(f"\nInput tokens: {input_length}")
|
| 249 |
|
| 250 |
except Exception as e:
|
| 251 |
+
print(f"!!! Error applying chat template: {e}")
|
| 252 |
return f"오류: 입력 형식을 처리하는 중 문제가 발생했습니다. ({e})"
|
| 253 |
|
| 254 |
try:
|
| 255 |
+
print("Generating response...")
|
| 256 |
generation_start_time = time.time()
|
| 257 |
|
| 258 |
+
# Prepare generation arguments, handling empty stop_token_ids_list
|
| 259 |
gen_kwargs = {
|
| 260 |
"max_new_tokens": MAX_NEW_TOKENS,
|
| 261 |
"pad_token_id": tokenizer.eos_token_id if tokenizer.eos_token_id is not None else tokenizer.pad_token_id,
|
| 262 |
"do_sample": True,
|
| 263 |
"temperature": 0.7,
|
| 264 |
"top_p": 0.9,
|
|
|
|
| 265 |
}
|
|
|
|
| 266 |
if stop_token_ids_list:
|
| 267 |
gen_kwargs["eos_token_id"] = stop_token_ids_list
|
| 268 |
else:
|
| 269 |
+
print("Generation Warning: No stop tokens defined.")
|
| 270 |
|
| 271 |
with torch.no_grad():
|
| 272 |
output_ids = model.generate(**inputs, **gen_kwargs)
|
| 273 |
|
| 274 |
generation_time = time.time() - generation_start_time
|
| 275 |
+
print(f"Generation complete in {generation_time:.2f} seconds.")
|
| 276 |
|
| 277 |
except Exception as e:
|
| 278 |
+
print(f"!!! Error during model generation: {e}")
|
| 279 |
if inputs is not None: del inputs
|
| 280 |
if output_ids is not None: del output_ids
|
| 281 |
gc.collect()
|
| 282 |
return f"오류: 응답을 생성하는 중 문제가 발생했습니다. ({e})"
|
| 283 |
|
| 284 |
+
# Decode the response
|
| 285 |
response = "오류: 응답 생성에 실패했습니다."
|
| 286 |
if output_ids is not None:
|
| 287 |
try:
|
| 288 |
new_tokens = output_ids[0, input_length:]
|
| 289 |
response = tokenizer.decode(new_tokens, skip_special_tokens=True)
|
| 290 |
+
print(f"Output tokens: {len(new_tokens)}")
|
| 291 |
del new_tokens
|
| 292 |
except Exception as e:
|
| 293 |
+
print(f"!!! Error decoding response: {e}")
|
| 294 |
response = "오류: 응답을 디코딩하는 중 문제가 발생했습니다."
|
| 295 |
|
| 296 |
+
# Clean up memory
|
| 297 |
if inputs is not None: del inputs
|
| 298 |
if output_ids is not None: del output_ids
|
| 299 |
gc.collect()
|
| 300 |
+
print("Memory cleaned.")
|
| 301 |
+
|
| 302 |
+
return response
|
| 303 |
+
|
| 304 |
+
# --- Additional UI components ---
|
| 305 |
+
def create_welcome_markdown():
|
| 306 |
+
return """
|
| 307 |
+
# 🇰🇷 네이버 HyperCLOVA X SEED
|
| 308 |
+
|
| 309 |
+
한국의 기술력으로 개발된 네이버의 초거대 AI 언어모델 'HyperCLOVA X'를 경험해보세요.
|
| 310 |
+
이 데모는 0.5B 파라미터 경량 모델을 사용하여 한국어 자연어 처리 능력을 보여줍니다.
|
| 311 |
+
|
| 312 |
+
**사용 방법**:
|
| 313 |
+
- 아래 채팅창에 질문이나 요청을 입력하세요
|
| 314 |
+
- 한국어로 다양한 주제에 대한 대화를 나눠보세요
|
| 315 |
+
- 예시 질문을 클릭하여 빠르게 시작할 수도 있습니다
|
| 316 |
+
"""
|
| 317 |
+
|
| 318 |
+
# --- Gradio Interface Setup ---
|
| 319 |
+
print("--- Setting up Gradio Interface ---")
|
| 320 |
+
|
| 321 |
+
with gr.Blocks(css=custom_css) as demo:
|
| 322 |
+
gr.Markdown(create_welcome_markdown(), elem_id="intro-message")
|
| 323 |
+
|
| 324 |
+
chatbot = gr.ChatInterface(
|
| 325 |
+
fn=predict,
|
| 326 |
+
title="",
|
| 327 |
+
description="",
|
| 328 |
+
examples=[
|
| 329 |
+
["네이버 클로바X는 무엇인가요?"],
|
| 330 |
+
["슈뢰딩거 방정식과 양자역학의 관계를 설명해주세요."],
|
| 331 |
+
["딥러닝 모델 학습 과정을 단계별로 알려줘."],
|
| 332 |
+
["제주도 여행 계획을 세우고 있는데, 3박 4일 추천 코스 좀 짜줄래?"],
|
| 333 |
+
["한국 역사에서 가장 중요한 사건 5가지는 무엇인가요?"],
|
| 334 |
+
["인공지능 윤리에 대해 설명해주세요."],
|
| 335 |
+
],
|
| 336 |
+
cache_examples=False,
|
| 337 |
+
submit_btn="보내기",
|
| 338 |
+
retry_btn="다시 시도",
|
| 339 |
+
undo_btn="취소",
|
| 340 |
+
clear_btn="새로운 대화",
|
| 341 |
+
)
|
| 342 |
+
|
| 343 |
+
with gr.Accordion("모델 정보", open=False):
|
| 344 |
+
gr.Markdown(f"""
|
| 345 |
+
- **모델**: {MODEL_ID}
|
| 346 |
+
- **환경**: ZeroGPU 공유 환경에서 실행 중
|
| 347 |
+
- **토큰 제한**: 최대 생성 토큰 수는 {MAX_NEW_TOKENS}개로 제한됩니다.
|
| 348 |
+
- **하드웨어**: {"GPU" if device.type == "cuda" else "CPU"} 환경에서 실행 중
|
| 349 |
+
""")
|
| 350 |
+
|
| 351 |
+
gr.Markdown(
|
| 352 |
+
"© 2025 네이버 HyperCLOVA X 데모 | Powered by Hugging Face & ZeroGPU",
|
| 353 |
+
elem_classes="footer"
|
| 354 |
+
)
|
| 355 |
+
|
| 356 |
+
# --- Application Launch ---
|
| 357 |
if __name__ == "__main__":
|
| 358 |
if load_successful:
|
| 359 |
warmup_model()
|
| 360 |
else:
|
| 361 |
+
print("Skipping warm-up because model loading failed.")
|
| 362 |
|
| 363 |
+
print("--- Launching Gradio App ---")
|
| 364 |
demo.queue().launch(
|
| 365 |
+
# share=True # Uncomment for public link
|
| 366 |
+
# server_name="0.0.0.0" # Uncomment for local network access
|
| 367 |
)
|