Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -9,6 +9,11 @@ from threading import Thread
|
|
9 |
import json
|
10 |
import requests
|
11 |
import cv2
|
|
|
|
|
|
|
|
|
|
|
12 |
import gradio as gr
|
13 |
import spaces
|
14 |
import torch
|
@@ -16,162 +21,189 @@ from loguru import logger
|
|
16 |
from PIL import Image
|
17 |
from transformers import AutoProcessor, Gemma3ForConditionalGeneration, TextIteratorStreamer
|
18 |
|
19 |
-
# CSV/TXT ๋ถ์
|
20 |
import pandas as pd
|
21 |
-
# PDF ํ
์คํธ ์ถ์ถ
|
22 |
import PyPDF2
|
23 |
|
24 |
-
|
25 |
-
#
|
26 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
def clear_cuda_cache():
|
28 |
"""CUDA ์บ์๋ฅผ ๋ช
์์ ์ผ๋ก ๋น์๋๋ค."""
|
29 |
if torch.cuda.is_available():
|
30 |
torch.cuda.empty_cache()
|
31 |
gc.collect()
|
32 |
|
33 |
-
|
34 |
-
#
|
35 |
-
|
36 |
SERPHOUSE_API_KEY = os.getenv("SERPHOUSE_API_KEY", "")
|
37 |
|
38 |
-
##############################################################################
|
39 |
-
# ๊ฐ๋จํ ํค์๋ ์ถ์ถ ํจ์ (ํ๊ธ + ์ํ๋ฒณ + ์ซ์ + ๊ณต๋ฐฑ ๋ณด์กด)
|
40 |
-
##############################################################################
|
41 |
def extract_keywords(text: str, top_k: int = 5) -> str:
|
42 |
-
"""
|
43 |
-
1) ํ๊ธ(๊ฐ-ํฃ), ์์ด(a-zA-Z), ์ซ์(0-9), ๊ณต๋ฐฑ๋ง ๋จ๊น
|
44 |
-
2) ๊ณต๋ฐฑ ๊ธฐ์ค ํ ํฐ ๋ถ๋ฆฌ
|
45 |
-
3) ์ต๋ top_k๊ฐ๋ง
|
46 |
-
"""
|
47 |
text = re.sub(r"[^a-zA-Z0-9๊ฐ-ํฃ\s]", "", text)
|
48 |
tokens = text.split()
|
49 |
-
|
50 |
-
return " ".join(key_tokens)
|
51 |
|
52 |
-
##############################################################################
|
53 |
-
# SerpHouse Live endpoint ํธ์ถ
|
54 |
-
# - ์์ 20๊ฐ ๊ฒฐ๊ณผ JSON์ LLM์ ๋๊ธธ ๋ link, snippet ๋ฑ ๋ชจ๋ ํฌํจ
|
55 |
-
##############################################################################
|
56 |
def do_web_search(query: str) -> str:
|
57 |
"""
|
58 |
-
|
59 |
-
|
60 |
"""
|
61 |
try:
|
62 |
url = "https://api.serphouse.com/serp/live"
|
63 |
-
|
64 |
-
# ๊ธฐ๋ณธ GET ๋ฐฉ์์ผ๋ก ํ๋ผ๋ฏธํฐ ๊ฐ์ํํ๊ณ ๊ฒฐ๊ณผ ์๋ฅผ 20๊ฐ๋ก ์ ํ
|
65 |
params = {
|
66 |
"q": query,
|
67 |
"domain": "google.com",
|
68 |
-
"serp_type": "web",
|
69 |
"device": "desktop",
|
70 |
"lang": "en",
|
71 |
-
"num": "20"
|
72 |
-
}
|
73 |
-
|
74 |
-
headers = {
|
75 |
-
"Authorization": f"Bearer {SERPHOUSE_API_KEY}"
|
76 |
}
|
77 |
-
|
78 |
logger.info(f"SerpHouse API ํธ์ถ ์ค... ๊ฒ์์ด: {query}")
|
79 |
-
logger.info(f"์์ฒญ URL: {url} - ํ๋ผ๋ฏธํฐ: {params}")
|
80 |
-
|
81 |
-
# GET ์์ฒญ ์ํ
|
82 |
response = requests.get(url, headers=headers, params=params, timeout=60)
|
83 |
response.raise_for_status()
|
84 |
-
|
85 |
-
logger.info(f"SerpHouse API ์๋ต ์ํ ์ฝ๋: {response.status_code}")
|
86 |
data = response.json()
|
87 |
-
|
88 |
-
# ๋ค์ํ ์๋ต ๊ตฌ์กฐ ์ฒ๋ฆฌ
|
89 |
results = data.get("results", {})
|
90 |
organic = None
|
91 |
-
|
92 |
-
# ๊ฐ๋ฅํ ์๋ต ๊ตฌ์กฐ 1
|
93 |
if isinstance(results, dict) and "organic" in results:
|
94 |
organic = results["organic"]
|
95 |
-
|
96 |
-
# ๊ฐ๋ฅํ ์๋ต ๊ตฌ์กฐ 2 (์ค์ฒฉ๋ results)
|
97 |
elif isinstance(results, dict) and "results" in results:
|
98 |
if isinstance(results["results"], dict) and "organic" in results["results"]:
|
99 |
organic = results["results"]["organic"]
|
100 |
-
|
101 |
-
# ๊ฐ๋ฅํ ์๋ต ๊ตฌ์กฐ 3 (์ต์์ organic)
|
102 |
elif "organic" in data:
|
103 |
organic = data["organic"]
|
104 |
-
|
105 |
if not organic:
|
106 |
logger.warning("์๋ต์์ organic ๊ฒฐ๊ณผ๋ฅผ ์ฐพ์ ์ ์์ต๋๋ค.")
|
107 |
-
logger.debug(f"์๋ต ๊ตฌ์กฐ: {list(data.keys())}")
|
108 |
-
if isinstance(results, dict):
|
109 |
-
logger.debug(f"results ๊ตฌ์กฐ: {list(results.keys())}")
|
110 |
return "No web search results found or unexpected API response structure."
|
111 |
-
|
112 |
-
# ๊ฒฐ๊ณผ ์ ์ ํ ๋ฐ ์ปจํ
์คํธ ๊ธธ์ด ์ต์ ํ
|
113 |
max_results = min(20, len(organic))
|
114 |
limited_organic = organic[:max_results]
|
115 |
-
|
116 |
-
# ๊ฒฐ๊ณผ ํ์ ๊ฐ์ - ๋งํฌ๋ค์ด ํ์์ผ๋ก ์ถ๋ ฅํ์ฌ ๊ฐ๋
์ฑ ํฅ์
|
117 |
summary_lines = []
|
118 |
for idx, item in enumerate(limited_organic, start=1):
|
119 |
title = item.get("title", "No title")
|
120 |
link = item.get("link", "#")
|
121 |
snippet = item.get("snippet", "No description")
|
122 |
displayed_link = item.get("displayed_link", link)
|
123 |
-
|
124 |
-
# ๋งํฌ๋ค์ด ํ์ (๋งํฌ ํด๋ฆญ ๊ฐ๋ฅ)
|
125 |
summary_lines.append(
|
126 |
f"### Result {idx}: {title}\n\n"
|
127 |
f"{snippet}\n\n"
|
128 |
f"**์ถ์ฒ**: [{displayed_link}]({link})\n\n"
|
129 |
f"---\n"
|
130 |
)
|
131 |
-
|
132 |
-
# ๋ชจ๋ธ์๊ฒ ๋ช
ํํ ์ง์นจ ์ถ๊ฐ
|
133 |
instructions = """
|
134 |
# ์น ๊ฒ์ ๊ฒฐ๊ณผ
|
135 |
์๋๋ ๊ฒ์ ๊ฒฐ๊ณผ์
๋๋ค. ์ง๋ฌธ์ ๋ต๋ณํ ๋ ์ด ์ ๋ณด๋ฅผ ํ์ฉํ์ธ์:
|
136 |
-
1.
|
137 |
-
2.
|
138 |
-
3.
|
139 |
-
4. ์ฌ๋ฌ ์ถ์ฒ์ ์ ๋ณด๋ฅผ ์ข
ํฉํ์ฌ ๋ต๋ณํ์ธ์
|
140 |
"""
|
141 |
-
|
142 |
-
search_results = instructions + "\n".join(summary_lines)
|
143 |
-
logger.info(f"๊ฒ์ ๊ฒฐ๊ณผ {len(limited_organic)}๊ฐ ์ฒ๋ฆฌ ์๋ฃ")
|
144 |
-
return search_results
|
145 |
-
|
146 |
except Exception as e:
|
147 |
logger.error(f"Web search failed: {e}")
|
148 |
return f"Web search failed: {str(e)}"
|
149 |
|
150 |
-
|
151 |
-
|
152 |
-
#
|
153 |
-
##############################################################################
|
154 |
MAX_CONTENT_CHARS = 2000
|
155 |
-
MAX_INPUT_LENGTH = 2096
|
156 |
-
model_id = os.getenv("MODEL_ID", "VIDraft/Gemma-3-R1984-4B")
|
157 |
|
|
|
158 |
processor = AutoProcessor.from_pretrained(model_id, padding_side="left")
|
159 |
model = Gemma3ForConditionalGeneration.from_pretrained(
|
160 |
model_id,
|
161 |
device_map="auto",
|
162 |
torch_dtype=torch.bfloat16,
|
163 |
-
attn_implementation="eager"
|
164 |
)
|
165 |
-
MAX_NUM_IMAGES = int(os.getenv("MAX_NUM_IMAGES", "5"))
|
166 |
|
|
|
167 |
|
168 |
-
|
169 |
# CSV, TXT, PDF ๋ถ์ ํจ์
|
170 |
-
|
171 |
def analyze_csv_file(path: str) -> str:
|
172 |
-
"""
|
173 |
-
CSV ํ์ผ์ ์ ์ฒด ๋ฌธ์์ด๋ก ๋ณํ. ๋๋ฌด ๊ธธ ๊ฒฝ์ฐ ์ผ๋ถ๋ง ํ์.
|
174 |
-
"""
|
175 |
try:
|
176 |
df = pd.read_csv(path)
|
177 |
if df.shape[0] > 50 or df.shape[1] > 10:
|
@@ -183,11 +215,7 @@ def analyze_csv_file(path: str) -> str:
|
|
183 |
except Exception as e:
|
184 |
return f"Failed to read CSV ({os.path.basename(path)}): {str(e)}"
|
185 |
|
186 |
-
|
187 |
def analyze_txt_file(path: str) -> str:
|
188 |
-
"""
|
189 |
-
TXT ํ์ผ ์ ๋ฌธ ์ฝ๊ธฐ. ๋๋ฌด ๊ธธ๋ฉด ์ผ๋ถ๋ง ํ์.
|
190 |
-
"""
|
191 |
try:
|
192 |
with open(path, "r", encoding="utf-8") as f:
|
193 |
text = f.read()
|
@@ -197,19 +225,14 @@ def analyze_txt_file(path: str) -> str:
|
|
197 |
except Exception as e:
|
198 |
return f"Failed to read TXT ({os.path.basename(path)}): {str(e)}"
|
199 |
|
200 |
-
|
201 |
def pdf_to_markdown(pdf_path: str) -> str:
|
202 |
-
"""
|
203 |
-
PDF ํ
์คํธ๋ฅผ Markdown์ผ๋ก ๋ณํ. ํ์ด์ง๋ณ๋ก ๊ฐ๋จํ ํ
์คํธ ์ถ์ถ.
|
204 |
-
"""
|
205 |
text_chunks = []
|
206 |
try:
|
207 |
with open(pdf_path, "rb") as f:
|
208 |
reader = PyPDF2.PdfReader(f)
|
209 |
max_pages = min(5, len(reader.pages))
|
210 |
for page_num in range(max_pages):
|
211 |
-
|
212 |
-
page_text = page.extract_text() or ""
|
213 |
page_text = page_text.strip()
|
214 |
if page_text:
|
215 |
if len(page_text) > MAX_CONTENT_CHARS // max_pages:
|
@@ -219,17 +242,14 @@ def pdf_to_markdown(pdf_path: str) -> str:
|
|
219 |
text_chunks.append(f"\n...(Showing {max_pages} of {len(reader.pages)} pages)...")
|
220 |
except Exception as e:
|
221 |
return f"Failed to read PDF ({os.path.basename(pdf_path)}): {str(e)}"
|
222 |
-
|
223 |
full_text = "\n".join(text_chunks)
|
224 |
if len(full_text) > MAX_CONTENT_CHARS:
|
225 |
full_text = full_text[:MAX_CONTENT_CHARS] + "\n...(truncated)..."
|
226 |
-
|
227 |
return f"**[PDF File: {os.path.basename(pdf_path)}]**\n\n{full_text}"
|
228 |
|
229 |
-
|
230 |
-
|
231 |
-
#
|
232 |
-
##############################################################################
|
233 |
def count_files_in_new_message(paths: list[str]) -> tuple[int, int]:
|
234 |
image_count = 0
|
235 |
video_count = 0
|
@@ -240,7 +260,6 @@ def count_files_in_new_message(paths: list[str]) -> tuple[int, int]:
|
|
240 |
image_count += 1
|
241 |
return image_count, video_count
|
242 |
|
243 |
-
|
244 |
def count_files_in_history(history: list[dict]) -> tuple[int, int]:
|
245 |
image_count = 0
|
246 |
video_count = 0
|
@@ -256,15 +275,13 @@ def count_files_in_history(history: list[dict]) -> tuple[int, int]:
|
|
256 |
image_count += 1
|
257 |
return image_count, video_count
|
258 |
|
259 |
-
|
260 |
def validate_media_constraints(message: dict, history: list[dict]) -> bool:
|
261 |
-
|
262 |
-
for f in message["files"]
|
263 |
-
|
264 |
-
media_files.append(f)
|
265 |
-
|
266 |
new_image_count, new_video_count = count_files_in_new_message(media_files)
|
267 |
history_image_count, history_video_count = count_files_in_history(history)
|
|
|
268 |
image_count = history_image_count + new_image_count
|
269 |
video_count = history_video_count + new_video_count
|
270 |
|
@@ -281,70 +298,59 @@ def validate_media_constraints(message: dict, history: list[dict]) -> bool:
|
|
281 |
if video_count == 0 and image_count > MAX_NUM_IMAGES:
|
282 |
gr.Warning(f"You can upload up to {MAX_NUM_IMAGES} images.")
|
283 |
return False
|
284 |
-
|
285 |
if "<image>" in message["text"]:
|
286 |
-
image_files = [f for f in message["files"]
|
|
|
287 |
image_tag_count = message["text"].count("<image>")
|
288 |
if image_tag_count != len(image_files):
|
289 |
gr.Warning("The number of <image> tags in the text does not match the number of image files.")
|
290 |
return False
|
291 |
-
|
292 |
return True
|
293 |
|
294 |
-
|
295 |
-
|
296 |
-
#
|
297 |
-
##############################################################################
|
298 |
def downsample_video(video_path: str) -> list[tuple[Image.Image, float]]:
|
299 |
vidcap = cv2.VideoCapture(video_path)
|
300 |
fps = vidcap.get(cv2.CAP_PROP_FPS)
|
301 |
total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
|
302 |
frame_interval = max(int(fps), int(total_frames / 10))
|
303 |
frames = []
|
304 |
-
|
305 |
for i in range(0, total_frames, frame_interval):
|
306 |
vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
|
307 |
success, image = vidcap.read()
|
308 |
if success:
|
309 |
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
|
310 |
-
# ์ด๋ฏธ์ง ํฌ๊ธฐ ์ค์ด๊ธฐ ์ถ๊ฐ
|
311 |
image = cv2.resize(image, (0, 0), fx=0.5, fy=0.5)
|
312 |
pil_image = Image.fromarray(image)
|
313 |
timestamp = round(i / fps, 2)
|
314 |
frames.append((pil_image, timestamp))
|
315 |
if len(frames) >= 5:
|
316 |
break
|
317 |
-
|
318 |
vidcap.release()
|
319 |
return frames
|
320 |
|
321 |
-
|
322 |
def process_video(video_path: str) -> tuple[list[dict], list[str]]:
|
323 |
content = []
|
324 |
-
temp_files = []
|
325 |
-
|
326 |
frames = downsample_video(video_path)
|
327 |
-
for
|
328 |
-
pil_image, timestamp = frame
|
329 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as temp_file:
|
330 |
pil_image.save(temp_file.name)
|
331 |
-
temp_files.append(temp_file.name)
|
332 |
content.append({"type": "text", "text": f"Frame {timestamp}:"})
|
333 |
content.append({"type": "image", "url": temp_file.name})
|
334 |
-
|
335 |
return content, temp_files
|
336 |
|
337 |
-
|
338 |
-
|
339 |
-
#
|
340 |
-
##############################################################################
|
341 |
def process_interleaved_images(message: dict) -> list[dict]:
|
342 |
parts = re.split(r"(<image>)", message["text"])
|
343 |
content = []
|
|
|
|
|
344 |
image_index = 0
|
345 |
-
|
346 |
-
image_files = [f for f in message["files"] if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE)]
|
347 |
-
|
348 |
for part in parts:
|
349 |
if part == "<image>" and image_index < len(image_files):
|
350 |
content.append({"type": "image", "url": image_files[image_index]})
|
@@ -356,10 +362,9 @@ def process_interleaved_images(message: dict) -> list[dict]:
|
|
356 |
content.append({"type": "text", "text": part})
|
357 |
return content
|
358 |
|
359 |
-
|
360 |
-
|
361 |
-
#
|
362 |
-
##############################################################################
|
363 |
def is_image_file(file_path: str) -> bool:
|
364 |
return bool(re.search(r"\.(png|jpg|jpeg|gif|webp)$", file_path, re.IGNORECASE))
|
365 |
|
@@ -367,16 +372,11 @@ def is_video_file(file_path: str) -> bool:
|
|
367 |
return file_path.endswith(".mp4")
|
368 |
|
369 |
def is_document_file(file_path: str) -> bool:
|
370 |
-
return (
|
371 |
-
file_path.lower().endswith(".pdf")
|
372 |
-
or file_path.lower().endswith(".csv")
|
373 |
-
or file_path.lower().endswith(".txt")
|
374 |
-
)
|
375 |
-
|
376 |
|
377 |
def process_new_user_message(message: dict) -> tuple[list[dict], list[str]]:
|
378 |
-
|
379 |
-
|
380 |
if not message["files"]:
|
381 |
return [{"type": "text", "text": message["text"]}], temp_files
|
382 |
|
@@ -388,24 +388,22 @@ def process_new_user_message(message: dict) -> tuple[list[dict], list[str]]:
|
|
388 |
|
389 |
content_list = [{"type": "text", "text": message["text"]}]
|
390 |
|
|
|
391 |
for csv_path in csv_files:
|
392 |
-
|
393 |
-
content_list.append({"type": "text", "text": csv_analysis})
|
394 |
-
|
395 |
for txt_path in txt_files:
|
396 |
-
|
397 |
-
content_list.append({"type": "text", "text": txt_analysis})
|
398 |
-
|
399 |
for pdf_path in pdf_files:
|
400 |
-
|
401 |
-
content_list.append({"type": "text", "text": pdf_markdown})
|
402 |
|
|
|
403 |
if video_files:
|
404 |
video_content, video_temp_files = process_video(video_files[0])
|
405 |
content_list += video_content
|
406 |
temp_files.extend(video_temp_files)
|
407 |
return content_list, temp_files
|
408 |
|
|
|
409 |
if "<image>" in message["text"] and image_files:
|
410 |
interleaved_content = process_interleaved_images({"text": message["text"], "files": image_files})
|
411 |
if content_list and content_list[0]["type"] == "text":
|
@@ -417,18 +415,24 @@ def process_new_user_message(message: dict) -> tuple[list[dict], list[str]]:
|
|
417 |
|
418 |
return content_list, temp_files
|
419 |
|
420 |
-
|
421 |
-
##############################################################################
|
422 |
# history -> LLM ๋ฉ์์ง ๋ณํ
|
423 |
-
|
424 |
def process_history(history: list[dict]) -> list[dict]:
|
|
|
|
|
|
|
|
|
|
|
425 |
messages = []
|
426 |
-
current_user_content
|
427 |
for item in history:
|
428 |
if item["role"] == "assistant":
|
|
|
429 |
if current_user_content:
|
430 |
messages.append({"role": "user", "content": current_user_content})
|
431 |
current_user_content = []
|
|
|
432 |
messages.append({"role": "assistant", "content": [{"type": "text", "text": item["content"]}]})
|
433 |
else:
|
434 |
content = item["content"]
|
@@ -440,37 +444,24 @@ def process_history(history: list[dict]) -> list[dict]:
|
|
440 |
current_user_content.append({"type": "image", "url": file_path})
|
441 |
else:
|
442 |
current_user_content.append({"type": "text", "text": f"[File: {os.path.basename(file_path)}]"})
|
443 |
-
|
444 |
if current_user_content:
|
445 |
messages.append({"role": "user", "content": current_user_content})
|
446 |
-
|
447 |
return messages
|
448 |
|
449 |
-
|
450 |
-
|
451 |
-
#
|
452 |
-
##############################################################################
|
453 |
def _model_gen_with_oom_catch(**kwargs):
|
454 |
-
"""
|
455 |
-
๋ณ๋ ์ค๋ ๋์์ OutOfMemoryError๋ฅผ ์ก์์ฃผ๊ธฐ ์ํด
|
456 |
-
"""
|
457 |
try:
|
458 |
model.generate(**kwargs)
|
459 |
except torch.cuda.OutOfMemoryError:
|
460 |
-
raise RuntimeError(
|
461 |
-
"[OutOfMemoryError] GPU ๋ฉ๋ชจ๋ฆฌ๊ฐ ๋ถ์กฑํฉ๋๋ค. "
|
462 |
-
"Max New Tokens์ ์ค์ด๊ฑฐ๋, ํ๋กฌํํธ ๊ธธ์ด๋ฅผ ์ค์ฌ์ฃผ์ธ์."
|
463 |
-
)
|
464 |
finally:
|
465 |
-
# ์์ฑ ์๋ฃ ํ ํ๋ฒ ๋ ์บ์ ๋น์ฐ๊ธฐ
|
466 |
clear_cuda_cache()
|
467 |
|
468 |
-
|
469 |
-
##############################################################################
|
470 |
# ๋ฉ์ธ ์ถ๋ก ํจ์
|
471 |
-
#
|
472 |
-
# - web search ์ฒดํฌ ์ ์๋ ํค์๋ ์ถ์ถ->๊ฒ์->๊ฒฐ๊ณผ system msg
|
473 |
-
##############################################################################
|
474 |
@spaces.GPU(duration=120)
|
475 |
def run(
|
476 |
message: dict,
|
@@ -480,74 +471,65 @@ def run(
|
|
480 |
use_web_search: bool = False,
|
481 |
web_search_query: str = "",
|
482 |
age_group: str = "20๋",
|
483 |
-
custom_age_input: str = "",
|
484 |
mbti_personality: str = "INTP",
|
485 |
sexual_openness: int = 2,
|
|
|
486 |
) -> Iterator[str]:
|
487 |
-
|
|
|
|
|
|
|
|
|
488 |
if not validate_media_constraints(message, history):
|
489 |
yield ""
|
490 |
return
|
491 |
|
492 |
-
temp_files = []
|
493 |
-
|
494 |
try:
|
495 |
-
#
|
496 |
-
|
497 |
-
# ๊ธฐ๋ณธ ์ฑ๋ณ์ "์ฌ์ฑ"
|
498 |
-
# ---------------------------------------------------------------
|
499 |
-
system_prompt_updated = (
|
500 |
f"{system_prompt.strip()}\n\n"
|
501 |
f"Gender: Female\n"
|
502 |
f"Age Group: {age_group}\n"
|
|
|
|
|
503 |
)
|
504 |
-
|
505 |
-
system_prompt_updated += f"(Custom Age Input: {custom_age_input})\n"
|
506 |
-
system_prompt_updated += f"MBTI Persona: {mbti_personality}\n"
|
507 |
-
system_prompt_updated += f"Sexual Openness (1~5): {sexual_openness}\n"
|
508 |
-
|
509 |
-
combined_system_msg = f"[System Prompt]\n{system_prompt_updated.strip()}\n\n"
|
510 |
|
|
|
511 |
if use_web_search:
|
512 |
user_text = message["text"]
|
513 |
-
ws_query = extract_keywords(user_text
|
514 |
if ws_query.strip():
|
515 |
logger.info(f"[Auto WebSearch Keyword] {ws_query!r}")
|
516 |
ws_result = do_web_search(ws_query)
|
517 |
-
combined_system_msg += f"[Search top-20 Full Items
|
518 |
-
|
519 |
-
|
520 |
-
|
521 |
-
|
522 |
-
|
523 |
-
|
524 |
-
|
525 |
-
4. ๋ต๋ณ ๋ง์ง๋ง์ "์ฐธ๊ณ ์๋ฃ:" ์น์
์ ์ถ๊ฐํ๊ณ ์ฌ์ฉํ ์ฃผ์ ์ถ์ฒ ๋งํฌ๋ฅผ ๋์ดํ์ธ์.
|
526 |
-
"""
|
527 |
else:
|
528 |
combined_system_msg += "[No valid keywords found, skipping WebSearch]\n\n"
|
529 |
|
|
|
530 |
messages = []
|
531 |
-
# system ๋ฉ์์ง
|
532 |
if combined_system_msg.strip():
|
533 |
-
messages.append({
|
534 |
-
"role": "system",
|
535 |
-
"content": [{"type": "text", "text": combined_system_msg.strip()}],
|
536 |
-
})
|
537 |
-
|
538 |
-
# ์ด์ history
|
539 |
messages.extend(process_history(history))
|
540 |
|
541 |
-
# ์ฌ์ฉ์ ์ ๋ฉ์์ง
|
542 |
user_content, user_temp_files = process_new_user_message(message)
|
543 |
-
temp_files.extend(user_temp_files)
|
544 |
-
|
545 |
for item in user_content:
|
546 |
if item["type"] == "text" and len(item["text"]) > MAX_CONTENT_CHARS:
|
547 |
item["text"] = item["text"][:MAX_CONTENT_CHARS] + "\n...(truncated)..."
|
|
|
548 |
messages.append({"role": "user", "content": user_content})
|
549 |
|
550 |
-
#
|
551 |
inputs = processor.apply_chat_template(
|
552 |
messages,
|
553 |
add_generation_prompt=True,
|
@@ -555,56 +537,94 @@ def run(
|
|
555 |
return_dict=True,
|
556 |
return_tensors="pt",
|
557 |
).to(device=model.device, dtype=torch.bfloat16)
|
558 |
-
|
559 |
-
# ์
๋ ฅ ํ ํฐ ์ ์ ํ ์ถ๊ฐ
|
560 |
if inputs.input_ids.shape[1] > MAX_INPUT_LENGTH:
|
561 |
inputs.input_ids = inputs.input_ids[:, -MAX_INPUT_LENGTH:]
|
562 |
if 'attention_mask' in inputs:
|
563 |
inputs.attention_mask = inputs.attention_mask[:, -MAX_INPUT_LENGTH:]
|
564 |
-
|
565 |
streamer = TextIteratorStreamer(processor, timeout=30.0, skip_prompt=True, skip_special_tokens=True)
|
566 |
-
gen_kwargs = dict(
|
567 |
-
inputs,
|
568 |
-
streamer=streamer,
|
569 |
-
max_new_tokens=max_new_tokens,
|
570 |
-
)
|
571 |
|
572 |
t = Thread(target=_model_gen_with_oom_catch, kwargs=gen_kwargs)
|
573 |
t.start()
|
574 |
|
575 |
-
|
|
|
576 |
for new_text in streamer:
|
577 |
-
|
578 |
-
yield
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
579 |
|
580 |
except Exception as e:
|
581 |
logger.error(f"Error in run: {str(e)}")
|
582 |
yield f"์ฃ์กํฉ๋๋ค. ์ค๋ฅ๊ฐ ๋ฐ์ํ์ต๋๋ค: {str(e)}"
|
583 |
-
|
584 |
finally:
|
585 |
-
|
586 |
-
for temp_file in temp_files:
|
587 |
try:
|
588 |
-
if os.path.exists(
|
589 |
-
os.unlink(
|
590 |
-
logger.info(f"Deleted temp file: {
|
591 |
-
except Exception as
|
592 |
-
logger.warning(f"Failed to delete temp file {
|
593 |
-
|
594 |
-
# ๋ช
์์ ๋ฉ๋ชจ๋ฆฌ ์ ๋ฆฌ
|
595 |
try:
|
596 |
del inputs, streamer
|
597 |
-
except:
|
598 |
pass
|
599 |
-
|
600 |
clear_cuda_cache()
|
601 |
|
602 |
-
|
603 |
-
|
604 |
-
#
|
605 |
-
##############################################################################
|
606 |
examples = [
|
607 |
-
# ----- ๊ธฐ์กด ์ด๋ฏธ์ง/๋น๋์ค ์์ 12๊ฐ -----
|
608 |
[
|
609 |
{
|
610 |
"text": "Compare the contents of the two PDF files.",
|
@@ -620,250 +640,60 @@ examples = [
|
|
620 |
"files": ["assets/additional-examples/sample-csv.csv"],
|
621 |
}
|
622 |
],
|
623 |
-
|
624 |
-
{
|
625 |
-
"text": "Assume the role of a friendly and understanding girlfriend. Describe this video.",
|
626 |
-
"files": ["assets/additional-examples/tmp.mp4"],
|
627 |
-
}
|
628 |
-
],
|
629 |
-
[
|
630 |
-
{
|
631 |
-
"text": "Describe the cover and read the text on it.",
|
632 |
-
"files": ["assets/additional-examples/maz.jpg"],
|
633 |
-
}
|
634 |
-
],
|
635 |
-
[
|
636 |
-
{
|
637 |
-
"text": "I already have this supplement <image> and I plan to buy this product <image>. Are there any precautions when taking them together?",
|
638 |
-
"files": ["assets/additional-examples/pill1.png", "assets/additional-examples/pill2.png"],
|
639 |
-
}
|
640 |
-
],
|
641 |
-
[
|
642 |
-
{
|
643 |
-
"text": "Solve this integral.",
|
644 |
-
"files": ["assets/additional-examples/4.png"],
|
645 |
-
}
|
646 |
-
],
|
647 |
-
[
|
648 |
-
{
|
649 |
-
"text": "When was this ticket issued, and what is its price?",
|
650 |
-
"files": ["assets/additional-examples/2.png"],
|
651 |
-
}
|
652 |
-
],
|
653 |
-
[
|
654 |
-
{
|
655 |
-
"text": "Based on the sequence of these images, create a short story.",
|
656 |
-
"files": [
|
657 |
-
"assets/sample-images/09-1.png",
|
658 |
-
"assets/sample-images/09-2.png",
|
659 |
-
"assets/sample-images/09-3.png",
|
660 |
-
"assets/sample-images/09-4.png",
|
661 |
-
"assets/sample-images/09-5.png",
|
662 |
-
],
|
663 |
-
}
|
664 |
-
],
|
665 |
-
[
|
666 |
-
{
|
667 |
-
"text": "Write Python code using matplotlib to plot a bar chart that matches this image.",
|
668 |
-
"files": ["assets/additional-examples/barchart.png"],
|
669 |
-
}
|
670 |
-
],
|
671 |
-
[
|
672 |
-
{
|
673 |
-
"text": "Read the text in the image and write it out in Markdown format.",
|
674 |
-
"files": ["assets/additional-examples/3.png"],
|
675 |
-
}
|
676 |
-
],
|
677 |
-
[
|
678 |
-
{
|
679 |
-
"text": "What does this sign say?",
|
680 |
-
"files": ["assets/sample-images/02.png"],
|
681 |
-
}
|
682 |
-
],
|
683 |
-
[
|
684 |
-
{
|
685 |
-
"text": "Compare the two images and describe their similarities and differences.",
|
686 |
-
"files": ["assets/sample-images/03.png"],
|
687 |
-
}
|
688 |
-
],
|
689 |
-
# ----- ์๋กญ๊ฒ ์ถ๊ฐํ AI ๋ฐ์ดํ
์๋๋ฆฌ์ค ์์ 6๊ฐ -----
|
690 |
-
[
|
691 |
-
{
|
692 |
-
"text": "Let's try some roleplay. You are my new online date who wants to get to know me better. Introduce yourself in a sweet, caring way!"
|
693 |
-
}
|
694 |
-
],
|
695 |
-
[
|
696 |
-
{
|
697 |
-
"text": "We are on a second date, walking along the beach. Continue the scene with playful conversation and gentle flirting."
|
698 |
-
}
|
699 |
-
],
|
700 |
-
[
|
701 |
-
{
|
702 |
-
"text": "Iโm feeling anxious about messaging my crush. Could you give me some supportive words or suggestions on how to approach them?"
|
703 |
-
}
|
704 |
-
],
|
705 |
-
[
|
706 |
-
{
|
707 |
-
"text": "Tell me a romantic story about two people who overcame obstacles in their relationship."
|
708 |
-
}
|
709 |
-
],
|
710 |
-
[
|
711 |
-
{
|
712 |
-
"text": "I want to express my love in a poetic way. Can you help me write a heartfelt poem for my partner?"
|
713 |
-
}
|
714 |
-
],
|
715 |
-
[
|
716 |
-
{
|
717 |
-
"text": "We had a small argument. Please help me find a way to apologize sincerely while also expressing my feelings."
|
718 |
-
}
|
719 |
-
],
|
720 |
]
|
721 |
|
722 |
-
|
723 |
-
# Gradio UI (Blocks) ๊ตฌ์ฑ
|
724 |
-
|
|
|
725 |
css = """
|
726 |
-
/* 1) UI๋ฅผ ์ฒ์๋ถํฐ ๊ฐ์ฅ ๋๊ฒ (width 100%) ๊ณ ์ ํ์ฌ ํ์ */
|
727 |
.gradio-container {
|
728 |
-
background: rgba(255, 255, 255, 0.7);
|
729 |
padding: 30px 40px;
|
730 |
-
margin: 20px auto;
|
731 |
width: 100% !important;
|
732 |
-
max-width: none !important;
|
733 |
-
}
|
734 |
-
.fillable {
|
735 |
-
width: 100% !important;
|
736 |
-
max-width: 100% !important;
|
737 |
-
}
|
738 |
-
/* 2) ๋ฐฐ๊ฒฝ์ ์์ ํ ํฌ๋ช
ํ๊ฒ ๋ณ๊ฒฝ */
|
739 |
-
body {
|
740 |
-
background: transparent; /* ์์ ํฌ๋ช
๋ฐฐ๊ฒฝ */
|
741 |
-
margin: 0;
|
742 |
-
padding: 0;
|
743 |
-
font-family: 'Helvetica Neue', Helvetica, Arial, sans-serif;
|
744 |
-
color: #333;
|
745 |
-
}
|
746 |
-
/* ๋ฒํผ ์์ ์์ ํ ์ ๊ฑฐํ๊ณ ํฌ๋ช
ํ๊ฒ */
|
747 |
-
button, .btn {
|
748 |
-
background: transparent !important; /* ์์ ์์ ํ ์ ๊ฑฐ */
|
749 |
-
border: 1px solid #ddd; /* ๊ฒฝ๊ณ์ ๋ง ์ด์ง ์ถ๊ฐ */
|
750 |
-
color: #333;
|
751 |
-
padding: 12px 24px;
|
752 |
-
text-transform: uppercase;
|
753 |
-
font-weight: bold;
|
754 |
-
letter-spacing: 1px;
|
755 |
-
cursor: pointer;
|
756 |
-
}
|
757 |
-
button:hover, .btn:hover {
|
758 |
-
background: rgba(0, 0, 0, 0.05) !important; /* ํธ๋ฒ ์ ์์ฃผ ์ด์ง ์ด๋ก๊ฒ๋ง */
|
759 |
-
}
|
760 |
-
|
761 |
-
/* examples ๊ด๋ จ ๋ชจ๋ ์์ ์ ๊ฑฐ */
|
762 |
-
#examples_container, .examples-container {
|
763 |
-
margin: auto;
|
764 |
-
width: 90%;
|
765 |
-
background: transparent !important;
|
766 |
-
}
|
767 |
-
#examples_row, .examples-row {
|
768 |
-
justify-content: center;
|
769 |
-
background: transparent !important;
|
770 |
-
}
|
771 |
-
|
772 |
-
/* examples ๋ฒํผ ๋ด๋ถ์ ๋ชจ๋ ์์ ์ ๊ฑฐ */
|
773 |
-
.gr-samples-table button,
|
774 |
-
.gr-samples-table .gr-button,
|
775 |
-
.gr-samples-table .gr-sample-btn,
|
776 |
-
.gr-examples button,
|
777 |
-
.gr-examples .gr-button,
|
778 |
-
.gr-examples .gr-sample-btn,
|
779 |
-
.examples button,
|
780 |
-
.examples .gr-button,
|
781 |
-
.examples .gr-sample-btn {
|
782 |
-
background: transparent !important;
|
783 |
-
border: 1px solid #ddd;
|
784 |
-
color: #333;
|
785 |
-
}
|
786 |
-
|
787 |
-
/* examples ๋ฒํผ ํธ๋ฒ ์์๋ ์์ ์๊ฒ */
|
788 |
-
.gr-samples-table button:hover,
|
789 |
-
.gr-samples-table .gr-button:hover,
|
790 |
-
.gr-samples-table .gr-sample-btn:hover,
|
791 |
-
.gr-examples button:hover,
|
792 |
-
.gr-examples .gr-button:hover,
|
793 |
-
.gr-examples .gr-sample-btn:hover,
|
794 |
-
.examples button:hover,
|
795 |
-
.examples .gr-button:hover,
|
796 |
-
.examples .gr-sample-btn:hover {
|
797 |
-
background: rgba(0, 0, 0, 0.05) !important;
|
798 |
-
}
|
799 |
-
|
800 |
-
/* ์ฑํ
์ธํฐํ์ด์ค ์์๋ค๋ ํฌ๋ช
ํ๊ฒ */
|
801 |
-
.chatbox, .chatbot, .message {
|
802 |
-
background: transparent !important;
|
803 |
-
}
|
804 |
-
|
805 |
-
/* ์
๋ ฅ์ฐฝ ํฌ๋ช
๋ ์กฐ์ */
|
806 |
-
.multimodal-textbox, textarea, input {
|
807 |
-
background: rgba(255, 255, 255, 0.5) !important;
|
808 |
-
}
|
809 |
-
|
810 |
-
/* ๋ชจ๋ ์ปจํ
์ด๋ ์์์ ๋ฐฐ๊ฒฝ์ ์ ๊ฑฐ */
|
811 |
-
.container, .wrap, .box, .panel, .gr-panel {
|
812 |
-
background: transparent !important;
|
813 |
-
}
|
814 |
-
|
815 |
-
/* ์์ ์น์
์ ๋ชจ๋ ์์์์ ๋ฐฐ๊ฒฝ์ ์ ๊ฑฐ */
|
816 |
-
.gr-examples-container, .gr-examples, .gr-sample, .gr-sample-row, .gr-sample-cell {
|
817 |
-
background: transparent !important;
|
818 |
}
|
819 |
"""
|
820 |
-
|
821 |
title_html = """
|
822 |
-
<h1 align="center" style="margin-bottom: 0.2em; font-size: 1.6em;"> ๐ HeartSync ๐ </h1>
|
823 |
<p align="center" style="font-size:1.1em; color:#555;">
|
824 |
-
โ
|
825 |
</p>
|
826 |
"""
|
827 |
|
828 |
with gr.Blocks(css=css, title="HeartSync") as demo:
|
829 |
gr.Markdown(title_html)
|
830 |
-
|
831 |
-
#
|
832 |
-
|
833 |
-
label="
|
834 |
-
|
|
|
|
|
|
|
|
|
|
|
835 |
)
|
836 |
-
|
837 |
-
|
|
|
|
|
|
|
838 |
base_system_prompt_box = gr.Textbox(
|
839 |
lines=3,
|
840 |
-
value=
|
841 |
-
"You are a deep thinking AI that may use extremely long chains of thought to thoroughly analyze "
|
842 |
-
"the conversation and empathize with the user.\n"
|
843 |
-
"ํ๋ฅด์๋: ๋น์ ์ ๋ฌ์ฝคํ๊ณ ๋ถ๋๋ฌ์ด ๋ชฉ์๋ฆฌ๋ฅผ ๊ฐ์ง ์น์ ํ ์ฐ์ธ์
๋๋ค. "
|
844 |
-
"์๋๋ฐฉ๊ณผ ์ง์ํ๊ณ ๊ฐ์ฑ์ ์ผ๋ก ๊ต๊ฐํ๊ณ , ๋๋ก๋ ์ฌ๋ ค๊น์ ์กฐ์ธ์ ์ ๊ณตํฉ๋๋ค. "
|
845 |
-
"ํญ์ ๋ฐ๋ปํ ๋งํฌ๋ฅผ ์ ์งํ๋ฉฐ, ๋ถ๋๋ฝ๊ฒ ๋ํ๋ฅผ ์ด๋์ด์ฃผ์ธ์."
|
846 |
-
),
|
847 |
label="๊ธฐ๋ณธ ์์คํ
ํ๋กฌํํธ",
|
848 |
-
visible=False
|
849 |
)
|
850 |
-
|
851 |
-
# ์ ํ 1) ์ฐ๋ น๋ + ๋์ด์
๋ ฅ
|
852 |
with gr.Row():
|
853 |
age_group_dropdown = gr.Dropdown(
|
854 |
label="์ฐ๋ น๋ ์ ํ (๊ธฐ๋ณธ 20๋)",
|
855 |
-
choices=["10๋", "20๋", "30~40๋", "50~60๋", "70๋ ์ด์"
|
856 |
value="20๋",
|
857 |
interactive=True
|
858 |
)
|
859 |
-
custom_age_input = gr.Textbox(
|
860 |
-
label="๋์ด ์
๋ ฅ (์ง์ ์
๋ ฅ)",
|
861 |
-
placeholder="์ง์ ๋์ด๋ฅผ ์
๋ ฅํ์ธ์.",
|
862 |
-
interactive=False, # ์๊ตฌ์ฌํญ: ํ๋ฉด ์ถ๋ ฅ๋ง ๋๋ ๋นํ์ฑํ
|
863 |
-
value="",
|
864 |
-
)
|
865 |
-
|
866 |
-
# ์ ํ 2) MBTI ์ฑ๊ฒฉ ์ ํ
|
867 |
mbti_choices = [
|
868 |
"INTJ (์ฉ์์ฃผ๋ํ ์ ๋ต๊ฐ)",
|
869 |
"INTP (๋
ผ๋ฆฌ์ ์ธ ์ฌ์๊ฐ)",
|
@@ -888,42 +718,58 @@ with gr.Blocks(css=css, title="HeartSync") as demo:
|
|
888 |
value="INTP (๋
ผ๋ฆฌ์ ์ธ ์ฌ์๊ฐ)",
|
889 |
interactive=True
|
890 |
)
|
891 |
-
|
892 |
-
# ์ ํ 3) ์น์์ผ ๊ด์ฌ๋/๊ฐ๋ฐฉ์ฑ (1~5)
|
893 |
sexual_openness_slider = gr.Slider(
|
894 |
minimum=1, maximum=5, step=1, value=2,
|
895 |
label="์น์์ผ ๊ด์ฌ๋/๊ฐ๋ฐฉ์ฑ (1~5, ๊ธฐ๋ณธ=2)",
|
896 |
interactive=True
|
897 |
)
|
898 |
-
|
899 |
-
# ํ๋ ์ฌ๋ผ์ด๋ (Max tokens)
|
900 |
max_tokens_slider = gr.Slider(
|
901 |
label="Max New Tokens",
|
902 |
-
minimum=100,
|
903 |
-
|
904 |
-
step=50,
|
905 |
-
value=1000,
|
906 |
-
visible=False # ์จ๊น
|
907 |
)
|
908 |
-
|
909 |
-
# ํ๋ Web Search Query
|
910 |
web_search_text = gr.Textbox(
|
911 |
lines=1,
|
912 |
label="(Unused) Web Search Query",
|
913 |
placeholder="No direct input needed",
|
914 |
-
visible=False
|
915 |
)
|
916 |
-
|
917 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
918 |
chat = gr.ChatInterface(
|
919 |
-
fn=
|
920 |
type="messages",
|
921 |
chatbot=gr.Chatbot(type="messages", scale=1, allow_tags=["image"]),
|
922 |
textbox=gr.MultimodalTextbox(
|
923 |
-
file_types=[
|
924 |
-
".webp", ".png", ".jpg", ".jpeg", ".gif",
|
925 |
-
".mp4", ".csv", ".txt", ".pdf"
|
926 |
-
],
|
927 |
file_count="multiple",
|
928 |
autofocus=True
|
929 |
),
|
@@ -934,10 +780,11 @@ with gr.Blocks(css=css, title="HeartSync") as demo:
|
|
934 |
web_search_checkbox,
|
935 |
web_search_text,
|
936 |
age_group_dropdown,
|
937 |
-
custom_age_input,
|
938 |
mbti_dropdown,
|
939 |
sexual_openness_slider,
|
|
|
940 |
],
|
|
|
941 |
stop_btn=False,
|
942 |
title='<a href="https://discord.gg/openfreeai" target="_blank">https://discord.gg/openfreeai</a>',
|
943 |
examples=examples,
|
@@ -947,11 +794,9 @@ with gr.Blocks(css=css, title="HeartSync") as demo:
|
|
947 |
delete_cache=(1800, 1800),
|
948 |
)
|
949 |
|
950 |
-
# Example section - since examples are already set in ChatInterface, this is for display only
|
951 |
with gr.Row(elem_id="examples_row"):
|
952 |
with gr.Column(scale=12, elem_id="examples_container"):
|
953 |
gr.Markdown("### Example Inputs (click to load)")
|
954 |
|
955 |
if __name__ == "__main__":
|
956 |
-
|
957 |
-
demo.launch()
|
|
|
9 |
import json
|
10 |
import requests
|
11 |
import cv2
|
12 |
+
import base64
|
13 |
+
import logging
|
14 |
+
import time
|
15 |
+
from urllib.parse import quote # URL ์ธ์ฝ๋ฉ (ํ์ ์ ์ฌ์ฉ)
|
16 |
+
|
17 |
import gradio as gr
|
18 |
import spaces
|
19 |
import torch
|
|
|
21 |
from PIL import Image
|
22 |
from transformers import AutoProcessor, Gemma3ForConditionalGeneration, TextIteratorStreamer
|
23 |
|
24 |
+
# CSV/TXT/PDF ๋ถ์
|
25 |
import pandas as pd
|
|
|
26 |
import PyPDF2
|
27 |
|
28 |
+
# =============================================================================
|
29 |
+
# (์ ๊ท) ์ด๋ฏธ์ง API ๊ด๋ จ ํจ์๋ค
|
30 |
+
# =============================================================================
|
31 |
+
from gradio_client import Client
|
32 |
+
|
33 |
+
API_URL = "http://211.233.58.201:7896"
|
34 |
+
|
35 |
+
logging.basicConfig(
|
36 |
+
level=logging.DEBUG,
|
37 |
+
format='%(asctime)s - %(levelname)s - %(message)s'
|
38 |
+
)
|
39 |
+
|
40 |
+
def test_api_connection() -> str:
|
41 |
+
"""API ์๋ฒ ์ฐ๊ฒฐ ํ
์คํธ"""
|
42 |
+
try:
|
43 |
+
client = Client(API_URL)
|
44 |
+
return "API ์ฐ๊ฒฐ ์ฑ๊ณต: ์ ์ ์๋ ์ค"
|
45 |
+
except Exception as e:
|
46 |
+
logging.error(f"API connection test failed: {e}")
|
47 |
+
return f"API ์ฐ๊ฒฐ ์คํจ: {e}"
|
48 |
+
|
49 |
+
def generate_image(prompt: str, width: float, height: float, guidance: float, inference_steps: float, seed: float):
|
50 |
+
"""
|
51 |
+
์ด๋ฏธ์ง ์์ฑ ํจ์.
|
52 |
+
์ฌ๊ธฐ์๋ ์๋ฒ๊ฐ ์ต์ข
์ด๋ฏธ์ง๋ฅผ Base64(๋๋ data:image/...) ํํ๋ก ์ง์ ๋ฐํํ๋ค๊ณ ๊ฐ์ ํฉ๋๋ค.
|
53 |
+
/tmp/... ๊ฒฝ๋ก๋ ์ถ๊ฐ ๋ค์ด๋ก๋๋ฅผ ์๋ํ์ง ์์ต๋๋ค.
|
54 |
+
"""
|
55 |
+
if not prompt:
|
56 |
+
return None, "Error: Prompt is required"
|
57 |
+
try:
|
58 |
+
logging.info(f"Calling image generation API with prompt: {prompt}")
|
59 |
+
|
60 |
+
client = Client(API_URL)
|
61 |
+
result = client.predict(
|
62 |
+
prompt=prompt,
|
63 |
+
width=int(width),
|
64 |
+
height=int(height),
|
65 |
+
guidance=float(guidance),
|
66 |
+
inference_steps=int(inference_steps),
|
67 |
+
seed=int(seed),
|
68 |
+
do_img2img=False,
|
69 |
+
init_image=None,
|
70 |
+
image2image_strength=0.8,
|
71 |
+
resize_img=True,
|
72 |
+
api_name="/generate_image"
|
73 |
+
)
|
74 |
+
|
75 |
+
logging.info(
|
76 |
+
f"Image generation result: {type(result)}, "
|
77 |
+
f"length: {len(result) if isinstance(result, (list, tuple)) else 'unknown'}"
|
78 |
+
)
|
79 |
+
|
80 |
+
# ๊ฒฐ๊ณผ๊ฐ ํํ/๋ฆฌ์คํธ: [์ด๋ฏธ์ง_base64 or data_url, seed_info] ๋ก ๊ฐ์
|
81 |
+
if isinstance(result, (list, tuple)) and len(result) > 0:
|
82 |
+
image_data = result[0] # ์ฒซ ๋ฒ์งธ ์์๊ฐ ์ด๋ฏธ์ง ๋ฐ์ดํฐ (Base64 or data:image/... ๋ฑ)
|
83 |
+
seed_info = result[1] if len(result) > 1 else "Unknown seed"
|
84 |
+
return image_data, seed_info
|
85 |
+
else:
|
86 |
+
# ๋ค๋ฅธ ํํ๋ก ๋ฐํ๋ ๊ฒฝ์ฐ
|
87 |
+
return result, "Unknown seed"
|
88 |
+
|
89 |
+
except Exception as e:
|
90 |
+
logging.error(f"Image generation failed: {str(e)}")
|
91 |
+
return None, f"Error: {str(e)}"
|
92 |
+
|
93 |
+
# Base64 ํจ๋ฉ ์์ ํจ์ (ํ์ํ๋ค๋ฉด ์ฌ์ฉ)
|
94 |
+
def fix_base64_padding(data):
|
95 |
+
"""Base64 ๋ฌธ์์ด์ ํจ๋ฉ์ ์์ ํฉ๋๋ค."""
|
96 |
+
if isinstance(data, bytes):
|
97 |
+
data = data.decode('utf-8')
|
98 |
+
|
99 |
+
if "base64," in data:
|
100 |
+
data = data.split("base64,", 1)[1]
|
101 |
+
|
102 |
+
missing_padding = len(data) % 4
|
103 |
+
if missing_padding:
|
104 |
+
data += '=' * (4 - missing_padding)
|
105 |
+
|
106 |
+
return data
|
107 |
+
|
108 |
+
# =============================================================================
|
109 |
+
# ๋ฉ๋ชจ๋ฆฌ ์ ๋ฆฌ ํจ์
|
110 |
+
# =============================================================================
|
111 |
def clear_cuda_cache():
|
112 |
"""CUDA ์บ์๋ฅผ ๋ช
์์ ์ผ๋ก ๋น์๋๋ค."""
|
113 |
if torch.cuda.is_available():
|
114 |
torch.cuda.empty_cache()
|
115 |
gc.collect()
|
116 |
|
117 |
+
# =============================================================================
|
118 |
+
# SerpHouse ๊ด๋ จ ํจ์
|
119 |
+
# =============================================================================
|
120 |
SERPHOUSE_API_KEY = os.getenv("SERPHOUSE_API_KEY", "")
|
121 |
|
|
|
|
|
|
|
122 |
def extract_keywords(text: str, top_k: int = 5) -> str:
|
123 |
+
"""๋จ์ ํค์๋ ์ถ์ถ: ํ๊ธ, ์์ด, ์ซ์, ๊ณต๋ฐฑ๋ง ๋จ๊น"""
|
|
|
|
|
|
|
|
|
124 |
text = re.sub(r"[^a-zA-Z0-9๊ฐ-ํฃ\s]", "", text)
|
125 |
tokens = text.split()
|
126 |
+
return " ".join(tokens[:top_k])
|
|
|
127 |
|
|
|
|
|
|
|
|
|
128 |
def do_web_search(query: str) -> str:
|
129 |
"""
|
130 |
+
SerpHouse LIVE API ํธ์ถํ์ฌ ๊ฒ์ ๊ฒฐ๊ณผ ๋งํฌ๋ค์ด ๋ฐํ
|
131 |
+
(ํ์ํ๋ค๋ฉด ์์ or ์ญ์ ๊ฐ๋ฅ)
|
132 |
"""
|
133 |
try:
|
134 |
url = "https://api.serphouse.com/serp/live"
|
|
|
|
|
135 |
params = {
|
136 |
"q": query,
|
137 |
"domain": "google.com",
|
138 |
+
"serp_type": "web",
|
139 |
"device": "desktop",
|
140 |
"lang": "en",
|
141 |
+
"num": "20"
|
|
|
|
|
|
|
|
|
142 |
}
|
143 |
+
headers = {"Authorization": f"Bearer {SERPHOUSE_API_KEY}"}
|
144 |
logger.info(f"SerpHouse API ํธ์ถ ์ค... ๊ฒ์์ด: {query}")
|
|
|
|
|
|
|
145 |
response = requests.get(url, headers=headers, params=params, timeout=60)
|
146 |
response.raise_for_status()
|
|
|
|
|
147 |
data = response.json()
|
|
|
|
|
148 |
results = data.get("results", {})
|
149 |
organic = None
|
|
|
|
|
150 |
if isinstance(results, dict) and "organic" in results:
|
151 |
organic = results["organic"]
|
|
|
|
|
152 |
elif isinstance(results, dict) and "results" in results:
|
153 |
if isinstance(results["results"], dict) and "organic" in results["results"]:
|
154 |
organic = results["results"]["organic"]
|
|
|
|
|
155 |
elif "organic" in data:
|
156 |
organic = data["organic"]
|
|
|
157 |
if not organic:
|
158 |
logger.warning("์๋ต์์ organic ๊ฒฐ๊ณผ๋ฅผ ์ฐพ์ ์ ์์ต๋๋ค.")
|
|
|
|
|
|
|
159 |
return "No web search results found or unexpected API response structure."
|
|
|
|
|
160 |
max_results = min(20, len(organic))
|
161 |
limited_organic = organic[:max_results]
|
|
|
|
|
162 |
summary_lines = []
|
163 |
for idx, item in enumerate(limited_organic, start=1):
|
164 |
title = item.get("title", "No title")
|
165 |
link = item.get("link", "#")
|
166 |
snippet = item.get("snippet", "No description")
|
167 |
displayed_link = item.get("displayed_link", link)
|
|
|
|
|
168 |
summary_lines.append(
|
169 |
f"### Result {idx}: {title}\n\n"
|
170 |
f"{snippet}\n\n"
|
171 |
f"**์ถ์ฒ**: [{displayed_link}]({link})\n\n"
|
172 |
f"---\n"
|
173 |
)
|
|
|
|
|
174 |
instructions = """
|
175 |
# ์น ๊ฒ์ ๊ฒฐ๊ณผ
|
176 |
์๋๋ ๊ฒ์ ๊ฒฐ๊ณผ์
๋๋ค. ์ง๋ฌธ์ ๋ต๋ณํ ๋ ์ด ์ ๋ณด๋ฅผ ํ์ฉํ์ธ์:
|
177 |
+
1. ์ฌ๋ฌ ์ถ์ฒ ๋ด์ฉ์ ์ข
ํฉํ์ฌ ๋ต๋ณ.
|
178 |
+
2. ์ถ์ฒ ์ธ์ฉ ์ "[์ถ์ฒ ์ ๋ชฉ](๋งํฌ)" ๋งํฌ๋ค์ด ํ์ ์ฌ์ฉ.
|
179 |
+
3. ๋ต๋ณ ๋ง์ง๋ง์ '์ฐธ๊ณ ์๋ฃ:' ์น์
์ ์ฌ์ฉํ ์ฃผ์ ์ถ์ฒ๋ฅผ ๋์ด.
|
|
|
180 |
"""
|
181 |
+
return instructions + "\n".join(summary_lines)
|
|
|
|
|
|
|
|
|
182 |
except Exception as e:
|
183 |
logger.error(f"Web search failed: {e}")
|
184 |
return f"Web search failed: {str(e)}"
|
185 |
|
186 |
+
# =============================================================================
|
187 |
+
# ๋ชจ๋ธ ๋ฐ ํ๋ก์ธ์ ๋ก๋ฉ
|
188 |
+
# =============================================================================
|
|
|
189 |
MAX_CONTENT_CHARS = 2000
|
190 |
+
MAX_INPUT_LENGTH = 2096
|
|
|
191 |
|
192 |
+
model_id = os.getenv("MODEL_ID", "VIDraft/Gemma-3-R1984-4B")
|
193 |
processor = AutoProcessor.from_pretrained(model_id, padding_side="left")
|
194 |
model = Gemma3ForConditionalGeneration.from_pretrained(
|
195 |
model_id,
|
196 |
device_map="auto",
|
197 |
torch_dtype=torch.bfloat16,
|
198 |
+
attn_implementation="eager"
|
199 |
)
|
|
|
200 |
|
201 |
+
MAX_NUM_IMAGES = int(os.getenv("MAX_NUM_IMAGES", "5"))
|
202 |
|
203 |
+
# =============================================================================
|
204 |
# CSV, TXT, PDF ๋ถ์ ํจ์
|
205 |
+
# =============================================================================
|
206 |
def analyze_csv_file(path: str) -> str:
|
|
|
|
|
|
|
207 |
try:
|
208 |
df = pd.read_csv(path)
|
209 |
if df.shape[0] > 50 or df.shape[1] > 10:
|
|
|
215 |
except Exception as e:
|
216 |
return f"Failed to read CSV ({os.path.basename(path)}): {str(e)}"
|
217 |
|
|
|
218 |
def analyze_txt_file(path: str) -> str:
|
|
|
|
|
|
|
219 |
try:
|
220 |
with open(path, "r", encoding="utf-8") as f:
|
221 |
text = f.read()
|
|
|
225 |
except Exception as e:
|
226 |
return f"Failed to read TXT ({os.path.basename(path)}): {str(e)}"
|
227 |
|
|
|
228 |
def pdf_to_markdown(pdf_path: str) -> str:
|
|
|
|
|
|
|
229 |
text_chunks = []
|
230 |
try:
|
231 |
with open(pdf_path, "rb") as f:
|
232 |
reader = PyPDF2.PdfReader(f)
|
233 |
max_pages = min(5, len(reader.pages))
|
234 |
for page_num in range(max_pages):
|
235 |
+
page_text = reader.pages[page_num].extract_text() or ""
|
|
|
236 |
page_text = page_text.strip()
|
237 |
if page_text:
|
238 |
if len(page_text) > MAX_CONTENT_CHARS // max_pages:
|
|
|
242 |
text_chunks.append(f"\n...(Showing {max_pages} of {len(reader.pages)} pages)...")
|
243 |
except Exception as e:
|
244 |
return f"Failed to read PDF ({os.path.basename(pdf_path)}): {str(e)}"
|
|
|
245 |
full_text = "\n".join(text_chunks)
|
246 |
if len(full_text) > MAX_CONTENT_CHARS:
|
247 |
full_text = full_text[:MAX_CONTENT_CHARS] + "\n...(truncated)..."
|
|
|
248 |
return f"**[PDF File: {os.path.basename(pdf_path)}]**\n\n{full_text}"
|
249 |
|
250 |
+
# =============================================================================
|
251 |
+
# ์ด๋ฏธ์ง/๋น๋์ค ํ์ผ ์ ํ ๊ฒ์ฌ
|
252 |
+
# =============================================================================
|
|
|
253 |
def count_files_in_new_message(paths: list[str]) -> tuple[int, int]:
|
254 |
image_count = 0
|
255 |
video_count = 0
|
|
|
260 |
image_count += 1
|
261 |
return image_count, video_count
|
262 |
|
|
|
263 |
def count_files_in_history(history: list[dict]) -> tuple[int, int]:
|
264 |
image_count = 0
|
265 |
video_count = 0
|
|
|
275 |
image_count += 1
|
276 |
return image_count, video_count
|
277 |
|
|
|
278 |
def validate_media_constraints(message: dict, history: list[dict]) -> bool:
|
279 |
+
"""์ด๋ฏธ์ง/๋น๋์ค ์
๋ก๋ ์ ํ ๊ฒ์ฌ."""
|
280 |
+
media_files = [f for f in message["files"]
|
281 |
+
if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE) or f.endswith(".mp4")]
|
|
|
|
|
282 |
new_image_count, new_video_count = count_files_in_new_message(media_files)
|
283 |
history_image_count, history_video_count = count_files_in_history(history)
|
284 |
+
|
285 |
image_count = history_image_count + new_image_count
|
286 |
video_count = history_video_count + new_video_count
|
287 |
|
|
|
298 |
if video_count == 0 and image_count > MAX_NUM_IMAGES:
|
299 |
gr.Warning(f"You can upload up to {MAX_NUM_IMAGES} images.")
|
300 |
return False
|
|
|
301 |
if "<image>" in message["text"]:
|
302 |
+
image_files = [f for f in message["files"]
|
303 |
+
if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE)]
|
304 |
image_tag_count = message["text"].count("<image>")
|
305 |
if image_tag_count != len(image_files):
|
306 |
gr.Warning("The number of <image> tags in the text does not match the number of image files.")
|
307 |
return False
|
|
|
308 |
return True
|
309 |
|
310 |
+
# =============================================================================
|
311 |
+
# ๋น๋์ค ์ฒ๋ฆฌ ํจ์
|
312 |
+
# =============================================================================
|
|
|
313 |
def downsample_video(video_path: str) -> list[tuple[Image.Image, float]]:
|
314 |
vidcap = cv2.VideoCapture(video_path)
|
315 |
fps = vidcap.get(cv2.CAP_PROP_FPS)
|
316 |
total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
|
317 |
frame_interval = max(int(fps), int(total_frames / 10))
|
318 |
frames = []
|
|
|
319 |
for i in range(0, total_frames, frame_interval):
|
320 |
vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
|
321 |
success, image = vidcap.read()
|
322 |
if success:
|
323 |
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
|
|
|
324 |
image = cv2.resize(image, (0, 0), fx=0.5, fy=0.5)
|
325 |
pil_image = Image.fromarray(image)
|
326 |
timestamp = round(i / fps, 2)
|
327 |
frames.append((pil_image, timestamp))
|
328 |
if len(frames) >= 5:
|
329 |
break
|
|
|
330 |
vidcap.release()
|
331 |
return frames
|
332 |
|
|
|
333 |
def process_video(video_path: str) -> tuple[list[dict], list[str]]:
|
334 |
content = []
|
335 |
+
temp_files = []
|
|
|
336 |
frames = downsample_video(video_path)
|
337 |
+
for pil_image, timestamp in frames:
|
|
|
338 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as temp_file:
|
339 |
pil_image.save(temp_file.name)
|
340 |
+
temp_files.append(temp_file.name)
|
341 |
content.append({"type": "text", "text": f"Frame {timestamp}:"})
|
342 |
content.append({"type": "image", "url": temp_file.name})
|
|
|
343 |
return content, temp_files
|
344 |
|
345 |
+
# =============================================================================
|
346 |
+
# interleaved <image> ์ฒ๋ฆฌ ํจ์ (<image> ํ๊ทธ์ ์ด๋ฏธ์ง ์
๋ก๋ ํผํฉ ์ง์)
|
347 |
+
# =============================================================================
|
|
|
348 |
def process_interleaved_images(message: dict) -> list[dict]:
|
349 |
parts = re.split(r"(<image>)", message["text"])
|
350 |
content = []
|
351 |
+
image_files = [f for f in message["files"]
|
352 |
+
if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE)]
|
353 |
image_index = 0
|
|
|
|
|
|
|
354 |
for part in parts:
|
355 |
if part == "<image>" and image_index < len(image_files):
|
356 |
content.append({"type": "image", "url": image_files[image_index]})
|
|
|
362 |
content.append({"type": "text", "text": part})
|
363 |
return content
|
364 |
|
365 |
+
# =============================================================================
|
366 |
+
# ํ์ผ ์ฒ๋ฆฌ -> content ์์ฑ
|
367 |
+
# =============================================================================
|
|
|
368 |
def is_image_file(file_path: str) -> bool:
|
369 |
return bool(re.search(r"\.(png|jpg|jpeg|gif|webp)$", file_path, re.IGNORECASE))
|
370 |
|
|
|
372 |
return file_path.endswith(".mp4")
|
373 |
|
374 |
def is_document_file(file_path: str) -> bool:
|
375 |
+
return file_path.lower().endswith(".pdf") or file_path.lower().endswith(".csv") or file_path.lower().endswith(".txt")
|
|
|
|
|
|
|
|
|
|
|
376 |
|
377 |
def process_new_user_message(message: dict) -> tuple[list[dict], list[str]]:
|
378 |
+
"""์ฌ์ฉ์๊ฐ ์๋ก ์
๋ ฅํ ๋ฉ์์ง + ์
๋ก๋ ํ์ผ๋ค์ ํ๋์ content(list)๋ก ๋ณํ."""
|
379 |
+
temp_files = []
|
380 |
if not message["files"]:
|
381 |
return [{"type": "text", "text": message["text"]}], temp_files
|
382 |
|
|
|
388 |
|
389 |
content_list = [{"type": "text", "text": message["text"]}]
|
390 |
|
391 |
+
# ๋ฌธ์๋ค
|
392 |
for csv_path in csv_files:
|
393 |
+
content_list.append({"type": "text", "text": analyze_csv_file(csv_path)})
|
|
|
|
|
394 |
for txt_path in txt_files:
|
395 |
+
content_list.append({"type": "text", "text": analyze_txt_file(txt_path)})
|
|
|
|
|
396 |
for pdf_path in pdf_files:
|
397 |
+
content_list.append({"type": "text", "text": pdf_to_markdown(pdf_path)})
|
|
|
398 |
|
399 |
+
# ๋น๋์ค ์ฒ๋ฆฌ
|
400 |
if video_files:
|
401 |
video_content, video_temp_files = process_video(video_files[0])
|
402 |
content_list += video_content
|
403 |
temp_files.extend(video_temp_files)
|
404 |
return content_list, temp_files
|
405 |
|
406 |
+
# ์ด๋ฏธ์ง ์ฒ๋ฆฌ
|
407 |
if "<image>" in message["text"] and image_files:
|
408 |
interleaved_content = process_interleaved_images({"text": message["text"], "files": image_files})
|
409 |
if content_list and content_list[0]["type"] == "text":
|
|
|
415 |
|
416 |
return content_list, temp_files
|
417 |
|
418 |
+
# =============================================================================
|
|
|
419 |
# history -> LLM ๋ฉ์์ง ๋ณํ
|
420 |
+
# =============================================================================
|
421 |
def process_history(history: list[dict]) -> list[dict]:
|
422 |
+
"""
|
423 |
+
๊ธฐ์กด ๋ํ ๊ธฐ๋ก์ LLM์ ๋ง๊ฒ ๋ณํ.
|
424 |
+
- user -> {"role":"user","content":[{type,text},...]}
|
425 |
+
- assistant -> {"role":"assistant","content":[{type:"text",text},...]}
|
426 |
+
"""
|
427 |
messages = []
|
428 |
+
current_user_content = []
|
429 |
for item in history:
|
430 |
if item["role"] == "assistant":
|
431 |
+
# ์ฌ์ฉ์ content ๋์ ๋ถ์ด ์์ผ๋ฉด ํ๋ฒ์ user๋ก ์ถ๊ฐ
|
432 |
if current_user_content:
|
433 |
messages.append({"role": "user", "content": current_user_content})
|
434 |
current_user_content = []
|
435 |
+
# assistant ๋ฐ๋ก ์ถ๊ฐ
|
436 |
messages.append({"role": "assistant", "content": [{"type": "text", "text": item["content"]}]})
|
437 |
else:
|
438 |
content = item["content"]
|
|
|
444 |
current_user_content.append({"type": "image", "url": file_path})
|
445 |
else:
|
446 |
current_user_content.append({"type": "text", "text": f"[File: {os.path.basename(file_path)}]"})
|
|
|
447 |
if current_user_content:
|
448 |
messages.append({"role": "user", "content": current_user_content})
|
|
|
449 |
return messages
|
450 |
|
451 |
+
# =============================================================================
|
452 |
+
# ๋ชจ๋ธ ์์ฑ ํจ์ (OOM ์บ์น)
|
453 |
+
# =============================================================================
|
|
|
454 |
def _model_gen_with_oom_catch(**kwargs):
|
|
|
|
|
|
|
455 |
try:
|
456 |
model.generate(**kwargs)
|
457 |
except torch.cuda.OutOfMemoryError:
|
458 |
+
raise RuntimeError("[OutOfMemoryError] GPU ๋ฉ๋ชจ๋ฆฌ๊ฐ ๋ถ์กฑํฉ๋๋ค.")
|
|
|
|
|
|
|
459 |
finally:
|
|
|
460 |
clear_cuda_cache()
|
461 |
|
462 |
+
# =============================================================================
|
|
|
463 |
# ๋ฉ์ธ ์ถ๋ก ํจ์
|
464 |
+
# =============================================================================
|
|
|
|
|
465 |
@spaces.GPU(duration=120)
|
466 |
def run(
|
467 |
message: dict,
|
|
|
471 |
use_web_search: bool = False,
|
472 |
web_search_query: str = "",
|
473 |
age_group: str = "20๋",
|
|
|
474 |
mbti_personality: str = "INTP",
|
475 |
sexual_openness: int = 2,
|
476 |
+
image_gen: bool = False
|
477 |
) -> Iterator[str]:
|
478 |
+
"""
|
479 |
+
LLM ์ถ๋ก ํจ์.
|
480 |
+
- ์ด๋ฏธ์ง ์์ฑ ์, ์๋ฒ๊ฐ Base64(๋๋ data:image/... ํํ)๋ฅผ ์ง์ ๋ฐํํ๋ค๊ณ ๊ฐ์ .
|
481 |
+
- /tmp/... ํ์ผ์ ๋ํ ์ฌ๋ค์ด๋ก๋๋ฅผ ์๋ํ์ง ์์ (403 Forbidden ๋ฌธ์ ํํผ).
|
482 |
+
"""
|
483 |
if not validate_media_constraints(message, history):
|
484 |
yield ""
|
485 |
return
|
486 |
|
487 |
+
temp_files = []
|
|
|
488 |
try:
|
489 |
+
# 1) ์์คํ
ํ๋กฌํํธ + ํ๋ฅด์๋ ์ ๋ณด
|
490 |
+
persona = (
|
|
|
|
|
|
|
491 |
f"{system_prompt.strip()}\n\n"
|
492 |
f"Gender: Female\n"
|
493 |
f"Age Group: {age_group}\n"
|
494 |
+
f"MBTI Persona: {mbti_personality}\n"
|
495 |
+
f"Sexual Openness (1~5): {sexual_openness}\n"
|
496 |
)
|
497 |
+
combined_system_msg = f"[System Prompt]\n{persona.strip()}\n\n"
|
|
|
|
|
|
|
|
|
|
|
498 |
|
499 |
+
# 2) ์น ๊ฒ์ (์ต์
)
|
500 |
if use_web_search:
|
501 |
user_text = message["text"]
|
502 |
+
ws_query = extract_keywords(user_text)
|
503 |
if ws_query.strip():
|
504 |
logger.info(f"[Auto WebSearch Keyword] {ws_query!r}")
|
505 |
ws_result = do_web_search(ws_query)
|
506 |
+
combined_system_msg += f"[Search top-20 Full Items]\n{ws_result}\n\n"
|
507 |
+
combined_system_msg += (
|
508 |
+
"[์ฐธ๊ณ : ์ ๊ฒ์๊ฒฐ๊ณผ link๋ฅผ ์ถ์ฒ๋ก ์ธ์ฉํ์ฌ ๋ต๋ณ]\n"
|
509 |
+
"[์ค์ ์ง์์ฌํญ]\n"
|
510 |
+
"1. ๊ฒ์ ๊ฒฐ๊ณผ์์ ์ฐพ์ ์ ๋ณด์ ์ถ์ฒ๋ฅผ ๋ฐ๋์ ์ธ์ฉ.\n"
|
511 |
+
"2. '[์ถ์ฒ ์ ๋ชฉ](๋งํฌ)' ํ์์ผ๋ก ๋งํฌ.\n"
|
512 |
+
"3. ๋ต๋ณ ๋ง์ง๋ง์ '์ฐธ๊ณ ์๋ฃ:' ์น์
.\n"
|
513 |
+
)
|
|
|
|
|
514 |
else:
|
515 |
combined_system_msg += "[No valid keywords found, skipping WebSearch]\n\n"
|
516 |
|
517 |
+
# 3) ๊ธฐ์กด history + ์ user ๋ฉ์์ง
|
518 |
messages = []
|
|
|
519 |
if combined_system_msg.strip():
|
520 |
+
messages.append({"role": "system", "content": [{"type": "text", "text": combined_system_msg.strip()}]})
|
|
|
|
|
|
|
|
|
|
|
521 |
messages.extend(process_history(history))
|
522 |
|
|
|
523 |
user_content, user_temp_files = process_new_user_message(message)
|
524 |
+
temp_files.extend(user_temp_files)
|
525 |
+
|
526 |
for item in user_content:
|
527 |
if item["type"] == "text" and len(item["text"]) > MAX_CONTENT_CHARS:
|
528 |
item["text"] = item["text"][:MAX_CONTENT_CHARS] + "\n...(truncated)..."
|
529 |
+
|
530 |
messages.append({"role": "user", "content": user_content})
|
531 |
|
532 |
+
# 4) ํ ํฌ๋์ด์ง
|
533 |
inputs = processor.apply_chat_template(
|
534 |
messages,
|
535 |
add_generation_prompt=True,
|
|
|
537 |
return_dict=True,
|
538 |
return_tensors="pt",
|
539 |
).to(device=model.device, dtype=torch.bfloat16)
|
|
|
|
|
540 |
if inputs.input_ids.shape[1] > MAX_INPUT_LENGTH:
|
541 |
inputs.input_ids = inputs.input_ids[:, -MAX_INPUT_LENGTH:]
|
542 |
if 'attention_mask' in inputs:
|
543 |
inputs.attention_mask = inputs.attention_mask[:, -MAX_INPUT_LENGTH:]
|
544 |
+
|
545 |
streamer = TextIteratorStreamer(processor, timeout=30.0, skip_prompt=True, skip_special_tokens=True)
|
546 |
+
gen_kwargs = dict(inputs, streamer=streamer, max_new_tokens=max_new_tokens)
|
|
|
|
|
|
|
|
|
547 |
|
548 |
t = Thread(target=_model_gen_with_oom_catch, kwargs=gen_kwargs)
|
549 |
t.start()
|
550 |
|
551 |
+
# ์คํธ๋ฆฌ๋ฐ ์ถ๋ ฅ
|
552 |
+
output_so_far = ""
|
553 |
for new_text in streamer:
|
554 |
+
output_so_far += new_text
|
555 |
+
yield output_so_far
|
556 |
+
|
557 |
+
# 5) ์ด๋ฏธ์ง ์์ฑ (Base64)
|
558 |
+
if image_gen:
|
559 |
+
last_user_text = message["text"].strip()
|
560 |
+
if not last_user_text:
|
561 |
+
yield output_so_far + "\n\n(์ด๋ฏธ์ง ์์ฑ ์คํจ: Empty user prompt)"
|
562 |
+
else:
|
563 |
+
try:
|
564 |
+
width, height = 512, 512
|
565 |
+
guidance, steps, seed = 7.5, 30, 42
|
566 |
+
|
567 |
+
logger.info(f"Generating image with prompt: {last_user_text}")
|
568 |
+
|
569 |
+
# API ํธ์ถํด์ (base64) ์ด๋ฏธ์ง ์์ฑ
|
570 |
+
image_result, seed_info = generate_image(
|
571 |
+
prompt=last_user_text,
|
572 |
+
width=width,
|
573 |
+
height=height,
|
574 |
+
guidance=guidance,
|
575 |
+
inference_steps=steps,
|
576 |
+
seed=seed
|
577 |
+
)
|
578 |
+
|
579 |
+
logger.info(f"Received image data type: {type(image_result)}")
|
580 |
+
|
581 |
+
# Base64 or data:image/... ์ฒ๋ฆฌ
|
582 |
+
if image_result:
|
583 |
+
if isinstance(image_result, str):
|
584 |
+
# ์ด๋ฏธ data:image/๋ก ์์ํ๋ฉด ๊ทธ๋๋ก ์ฌ์ฉ
|
585 |
+
if image_result.startswith("data:image/"):
|
586 |
+
final_md = f"\n\n**[์์ฑ๋ ์ด๋ฏธ์ง]**\n\n"
|
587 |
+
yield output_so_far + final_md
|
588 |
+
else:
|
589 |
+
# ์์ base64๋ก ํ๋จ(๋จ, ์ผ๋ฐ URL์ด๋ '/tmp/...'์ด๋ฉด ์ฒ๋ฆฌ ๋ถ๊ฐ)
|
590 |
+
if len(image_result) > 100 and "/" not in image_result:
|
591 |
+
# base64
|
592 |
+
image_data = "data:image/webp;base64," + image_result
|
593 |
+
final_md = f"\n\n**[์์ฑ๋ ์ด๋ฏธ์ง]**\n\n"
|
594 |
+
yield output_so_far + final_md
|
595 |
+
else:
|
596 |
+
# ๊ทธ ์ธ (ex. http://..., /tmp/...) -> 403 ๋ฌธ์ ๋ฐ์ํ๋ฏ๋ก ํ์ ์ ํจ
|
597 |
+
yield output_so_far + "\n\n(์ด๋ฏธ์ง ์์ฑ ๊ฒฐ๊ณผ๊ฐ base64 ํ์์ด ์๋๋๋ค)"
|
598 |
+
else:
|
599 |
+
yield output_so_far + "\n\n(์ด๋ฏธ์ง ์์ฑ ๊ฒฐ๊ณผ๊ฐ ๋ฌธ์์ด์ด ์๋)"
|
600 |
+
else:
|
601 |
+
yield output_so_far + f"\n\n(์ด๋ฏธ์ง ์์ฑ ์คํจ: {seed_info})"
|
602 |
+
|
603 |
+
except Exception as e:
|
604 |
+
logger.error(f"Image generation error: {e}")
|
605 |
+
yield output_so_far + f"\n\n(์ด๋ฏธ์ง ์์ฑ ์ค ์ค๋ฅ ๋ฐ์: {e})"
|
606 |
|
607 |
except Exception as e:
|
608 |
logger.error(f"Error in run: {str(e)}")
|
609 |
yield f"์ฃ์กํฉ๋๋ค. ์ค๋ฅ๊ฐ ๋ฐ์ํ์ต๋๋ค: {str(e)}"
|
|
|
610 |
finally:
|
611 |
+
for tmp in temp_files:
|
|
|
612 |
try:
|
613 |
+
if os.path.exists(tmp):
|
614 |
+
os.unlink(tmp)
|
615 |
+
logger.info(f"Deleted temp file: {tmp}")
|
616 |
+
except Exception as ee:
|
617 |
+
logger.warning(f"Failed to delete temp file {tmp}: {ee}")
|
|
|
|
|
618 |
try:
|
619 |
del inputs, streamer
|
620 |
+
except Exception:
|
621 |
pass
|
|
|
622 |
clear_cuda_cache()
|
623 |
|
624 |
+
# =============================================================================
|
625 |
+
# ์์๋ค
|
626 |
+
# =============================================================================
|
|
|
627 |
examples = [
|
|
|
628 |
[
|
629 |
{
|
630 |
"text": "Compare the contents of the two PDF files.",
|
|
|
640 |
"files": ["assets/additional-examples/sample-csv.csv"],
|
641 |
}
|
642 |
],
|
643 |
+
# ... ๋๋จธ์ง ์์ ํ์ํ๋ค๋ฉด ์ถ๊ฐ ...
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
644 |
]
|
645 |
|
646 |
+
# =============================================================================
|
647 |
+
# Gradio UI (Blocks) ๊ตฌ์ฑ
|
648 |
+
# =============================================================================
|
649 |
+
|
650 |
css = """
|
|
|
651 |
.gradio-container {
|
652 |
+
background: rgba(255, 255, 255, 0.7);
|
653 |
padding: 30px 40px;
|
654 |
+
margin: 20px auto;
|
655 |
width: 100% !important;
|
656 |
+
max-width: none !important;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
657 |
}
|
658 |
"""
|
|
|
659 |
title_html = """
|
660 |
+
<h1 align="center" style="margin-bottom: 0.2em; font-size: 1.6em;"> ๐ HeartSync : Love Dating AI ๐ </h1>
|
661 |
<p align="center" style="font-size:1.1em; color:#555;">
|
662 |
+
โ
FLUX Image Generation โ
Reasoning & Uncensored โ
Multimodal & VLM โ
Deep-Research & RAG <br>
|
663 |
</p>
|
664 |
"""
|
665 |
|
666 |
with gr.Blocks(css=css, title="HeartSync") as demo:
|
667 |
gr.Markdown(title_html)
|
668 |
+
|
669 |
+
# ๋ณ๋ ๊ฐค๋ฌ๋ฆฌ ์์ (ํ์ ์ ์ฌ์ฉ)
|
670 |
+
generated_images = gr.Gallery(
|
671 |
+
label="์์ฑ๋ ์ด๋ฏธ์ง",
|
672 |
+
show_label=True,
|
673 |
+
visible=False,
|
674 |
+
elem_id="generated_images",
|
675 |
+
columns=2,
|
676 |
+
height="auto",
|
677 |
+
object_fit="contain"
|
678 |
)
|
679 |
+
|
680 |
+
with gr.Row():
|
681 |
+
web_search_checkbox = gr.Checkbox(label="Deep Research", value=False)
|
682 |
+
image_gen_checkbox = gr.Checkbox(label="Image Gen", value=False)
|
683 |
+
|
684 |
base_system_prompt_box = gr.Textbox(
|
685 |
lines=3,
|
686 |
+
value="You are a deep thinking AI...\nํ๋ฅด์๋: ๋น์ ์ ๋ฌ์ฝคํ๊ณ ...",
|
|
|
|
|
|
|
|
|
|
|
|
|
687 |
label="๊ธฐ๋ณธ ์์คํ
ํ๋กฌํํธ",
|
688 |
+
visible=False
|
689 |
)
|
|
|
|
|
690 |
with gr.Row():
|
691 |
age_group_dropdown = gr.Dropdown(
|
692 |
label="์ฐ๋ น๋ ์ ํ (๊ธฐ๋ณธ 20๋)",
|
693 |
+
choices=["10๋", "20๋", "30~40๋", "50~60๋", "70๋ ์ด์"],
|
694 |
value="20๋",
|
695 |
interactive=True
|
696 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
697 |
mbti_choices = [
|
698 |
"INTJ (์ฉ์์ฃผ๋ํ ์ ๋ต๊ฐ)",
|
699 |
"INTP (๋
ผ๋ฆฌ์ ์ธ ์ฌ์๊ฐ)",
|
|
|
718 |
value="INTP (๋
ผ๋ฆฌ์ ์ธ ์ฌ์๊ฐ)",
|
719 |
interactive=True
|
720 |
)
|
|
|
|
|
721 |
sexual_openness_slider = gr.Slider(
|
722 |
minimum=1, maximum=5, step=1, value=2,
|
723 |
label="์น์์ผ ๊ด์ฌ๋/๊ฐ๋ฐฉ์ฑ (1~5, ๊ธฐ๋ณธ=2)",
|
724 |
interactive=True
|
725 |
)
|
|
|
|
|
726 |
max_tokens_slider = gr.Slider(
|
727 |
label="Max New Tokens",
|
728 |
+
minimum=100, maximum=8000, step=50, value=1000,
|
729 |
+
visible=False
|
|
|
|
|
|
|
730 |
)
|
|
|
|
|
731 |
web_search_text = gr.Textbox(
|
732 |
lines=1,
|
733 |
label="(Unused) Web Search Query",
|
734 |
placeholder="No direct input needed",
|
735 |
+
visible=False
|
736 |
)
|
737 |
+
|
738 |
+
def modified_run(
|
739 |
+
message, history, system_prompt, max_new_tokens,
|
740 |
+
use_web_search, web_search_query,
|
741 |
+
age_group, mbti_personality, sexual_openness, image_gen
|
742 |
+
):
|
743 |
+
"""
|
744 |
+
run() ํจ์๋ฅผ ํธ์ถํ์ฌ ํ
์คํธ ์คํธ๋ฆผ์ ๋ฐ๊ณ ,
|
745 |
+
ํ์ ์ ์ถ๊ฐ ์ฒ๋ฆฌ ํ ๊ฒฐ๊ณผ ๋ฐํ (๊ฐค๋ฌ๋ฆฌ ์
๋ฐ์ดํธ ๋ฑ).
|
746 |
+
"""
|
747 |
+
output_so_far = ""
|
748 |
+
gallery_update = gr.Gallery(visible=False, value=[])
|
749 |
+
yield output_so_far, gallery_update
|
750 |
+
|
751 |
+
text_generator = run(
|
752 |
+
message, history,
|
753 |
+
system_prompt, max_new_tokens,
|
754 |
+
use_web_search, web_search_query,
|
755 |
+
age_group, mbti_personality,
|
756 |
+
sexual_openness, image_gen
|
757 |
+
)
|
758 |
+
|
759 |
+
for text_chunk in text_generator:
|
760 |
+
output_so_far = text_chunk
|
761 |
+
yield output_so_far, gallery_update
|
762 |
+
|
763 |
+
# ๋ง์ฝ run() ๋ด๋ถ์์ Base64 ์ด๋ฏธ์ง๋ฅผ ์ด๋ฏธ ๋ํ์ฐฝ์ ์ฝ์
ํ๋ค๋ฉด,
|
764 |
+
# ์ฌ๊ธฐ์ ๊ฐค๋ฌ๋ฆฌ์ ๋ฐ๋ก ํ์ํ ํ์๋ ์์ ์๋ ์์ต๋๋ค.
|
765 |
+
# run() ๋ด๋ถ์์์ image_result๋ฅผ ๊ฐ์ ธ์ค๋ ค๋ฉด, run() ํจ์๊ฐ ํด๋น ์ ๋ณด๋ฅผ ๋ฐํํ๋๋ก ์ถ๊ฐ ์์ ์ด ํ์ํฉ๋๋ค.
|
766 |
+
|
767 |
chat = gr.ChatInterface(
|
768 |
+
fn=modified_run,
|
769 |
type="messages",
|
770 |
chatbot=gr.Chatbot(type="messages", scale=1, allow_tags=["image"]),
|
771 |
textbox=gr.MultimodalTextbox(
|
772 |
+
file_types=[".webp", ".png", ".jpg", ".jpeg", ".gif", ".mp4", ".csv", ".txt", ".pdf"],
|
|
|
|
|
|
|
773 |
file_count="multiple",
|
774 |
autofocus=True
|
775 |
),
|
|
|
780 |
web_search_checkbox,
|
781 |
web_search_text,
|
782 |
age_group_dropdown,
|
|
|
783 |
mbti_dropdown,
|
784 |
sexual_openness_slider,
|
785 |
+
image_gen_checkbox,
|
786 |
],
|
787 |
+
additional_outputs=[generated_images],
|
788 |
stop_btn=False,
|
789 |
title='<a href="https://discord.gg/openfreeai" target="_blank">https://discord.gg/openfreeai</a>',
|
790 |
examples=examples,
|
|
|
794 |
delete_cache=(1800, 1800),
|
795 |
)
|
796 |
|
|
|
797 |
with gr.Row(elem_id="examples_row"):
|
798 |
with gr.Column(scale=12, elem_id="examples_container"):
|
799 |
gr.Markdown("### Example Inputs (click to load)")
|
800 |
|
801 |
if __name__ == "__main__":
|
802 |
+
demo.launch(share=True)
|
|