seawolf2357's picture
Update app.py
ced8ba1 verified
raw
history blame
21.5 kB
#!/usr/bin/env python
import os
import re
import tempfile
from collections.abc import Iterator
from threading import Thread
import cv2
import gradio as gr
import spaces
import torch
from loguru import logger
from PIL import Image
from transformers import AutoProcessor, Gemma3ForConditionalGeneration, TextIteratorStreamer
# CSV/TXT ๋ถ„์„
import pandas as pd
# PDF ํ…์ŠคํŠธ ์ถ”์ถœ
import PyPDF2
MAX_CONTENT_CHARS = 4000 # ๋„ˆ๋ฌด ํฐ ํŒŒ์ผ์„ ๋ง‰๊ธฐ ์œ„ํ•ด ์ตœ๋Œ€ ํ‘œ์‹œ 4000์ž
model_id = os.getenv("MODEL_ID", "google/gemma-3-27b-it")
processor = AutoProcessor.from_pretrained(model_id, padding_side="left")
model = Gemma3ForConditionalGeneration.from_pretrained(
model_id,
device_map="auto",
torch_dtype=torch.bfloat16,
attn_implementation="eager"
)
MAX_NUM_IMAGES = int(os.getenv("MAX_NUM_IMAGES", "5"))
##################################################
# CSV, TXT, PDF ๋ถ„์„ ํ•จ์ˆ˜
##################################################
def analyze_csv_file(path: str) -> str:
"""
CSV ํŒŒ์ผ์„ ์ „์ฒด ๋ฌธ์ž์—ด๋กœ ๋ณ€ํ™˜. ๋„ˆ๋ฌด ๊ธธ ๊ฒฝ์šฐ ์ผ๋ถ€๋งŒ ํ‘œ์‹œ.
"""
try:
df = pd.read_csv(path)
# ๋ฐ์ดํ„ฐ ํ”„๋ ˆ์ž„ ํฌ๊ธฐ ์ œํ•œ (ํ–‰/์—ด ์ˆ˜๊ฐ€ ๋งŽ์€ ๊ฒฝ์šฐ)
if df.shape[0] > 50 or df.shape[1] > 10:
df = df.iloc[:50, :10]
df_str = df.to_string()
if len(df_str) > MAX_CONTENT_CHARS:
df_str = df_str[:MAX_CONTENT_CHARS] + "\n...(truncated)..."
return f"**[CSV File: {os.path.basename(path)}]**\n\n{df_str}"
except Exception as e:
return f"Failed to read CSV ({os.path.basename(path)}): {str(e)}"
def analyze_txt_file(path: str) -> str:
"""
TXT ํŒŒ์ผ ์ „๋ฌธ ์ฝ๊ธฐ. ๋„ˆ๋ฌด ๊ธธ๋ฉด ์ผ๋ถ€๋งŒ ํ‘œ์‹œ.
"""
try:
with open(path, "r", encoding="utf-8") as f:
text = f.read()
if len(text) > MAX_CONTENT_CHARS:
text = text[:MAX_CONTENT_CHARS] + "\n...(truncated)..."
return f"**[TXT File: {os.path.basename(path)}]**\n\n{text}"
except Exception as e:
return f"Failed to read TXT ({os.path.basename(path)}): {str(e)}"
def pdf_to_markdown(pdf_path: str) -> str:
"""
PDF โ†’ Markdown. ํŽ˜์ด์ง€๋ณ„๋กœ ๊ฐ„๋‹จํžˆ ํ…์ŠคํŠธ ์ถ”์ถœ.
"""
text_chunks = []
try:
with open(pdf_path, "rb") as f:
reader = PyPDF2.PdfReader(f)
# ์ตœ๋Œ€ 5ํŽ˜์ด์ง€๋งŒ ์ฒ˜๋ฆฌ
max_pages = min(5, len(reader.pages))
for page_num in range(max_pages):
page = reader.pages[page_num]
page_text = page.extract_text() or ""
page_text = page_text.strip()
if page_text:
# ํŽ˜์ด์ง€๋ณ„ ํ…์ŠคํŠธ๋„ ์ œํ•œ
if len(page_text) > MAX_CONTENT_CHARS // max_pages:
page_text = page_text[:MAX_CONTENT_CHARS // max_pages] + "...(truncated)"
text_chunks.append(f"## Page {page_num+1}\n\n{page_text}\n")
if len(reader.pages) > max_pages:
text_chunks.append(f"\n...(Showing {max_pages} of {len(reader.pages)} pages)...")
except Exception as e:
return f"Failed to read PDF ({os.path.basename(pdf_path)}): {str(e)}"
full_text = "\n".join(text_chunks)
if len(full_text) > MAX_CONTENT_CHARS:
full_text = full_text[:MAX_CONTENT_CHARS] + "\n...(truncated)..."
return f"**[PDF File: {os.path.basename(pdf_path)}]**\n\n{full_text}"
##################################################
# ์ด๋ฏธ์ง€/๋น„๋””์˜ค ์—…๋กœ๋“œ ์ œํ•œ ๊ฒ€์‚ฌ
##################################################
def count_files_in_new_message(paths: list[str]) -> tuple[int, int]:
image_count = 0
video_count = 0
for path in paths:
if path.endswith(".mp4"):
video_count += 1
elif re.search(r"\.(png|jpg|jpeg|gif|webp)$", path, re.IGNORECASE):
image_count += 1
return image_count, video_count
def count_files_in_history(history: list[dict]) -> tuple[int, int]:
image_count = 0
video_count = 0
for item in history:
if item["role"] != "user" or isinstance(item["content"], str):
continue
if isinstance(item["content"], list) and len(item["content"]) > 0:
file_path = item["content"][0]
if isinstance(file_path, str):
if file_path.endswith(".mp4"):
video_count += 1
elif re.search(r"\.(png|jpg|jpeg|gif|webp)$", file_path, re.IGNORECASE):
image_count += 1
return image_count, video_count
def validate_media_constraints(message: dict, history: list[dict]) -> bool:
"""
- ๋น„๋””์˜ค 1๊ฐœ ์ดˆ๊ณผ ๋ถˆ๊ฐ€
- ๋น„๋””์˜ค์™€ ์ด๋ฏธ์ง€ ํ˜ผํ•ฉ ๋ถˆ๊ฐ€
- ์ด๋ฏธ์ง€ ๊ฐœ์ˆ˜ MAX_NUM_IMAGES ์ดˆ๊ณผ ๋ถˆ๊ฐ€
- <image> ํƒœ๊ทธ๊ฐ€ ์žˆ์œผ๋ฉด ํƒœ๊ทธ ์ˆ˜์™€ ์‹ค์ œ ์ด๋ฏธ์ง€ ์ˆ˜ ์ผ์น˜
- CSV, TXT, PDF ๋“ฑ์€ ์—ฌ๊ธฐ์„œ ์ œํ•œํ•˜์ง€ ์•Š์Œ
"""
# ์ด๋ฏธ์ง€์™€ ๋น„๋””์˜ค ํŒŒ์ผ๋งŒ ํ•„ํ„ฐ๋ง
media_files = []
for f in message["files"]:
if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE) or f.endswith(".mp4"):
media_files.append(f)
new_image_count, new_video_count = count_files_in_new_message(media_files)
history_image_count, history_video_count = count_files_in_history(history)
image_count = history_image_count + new_image_count
video_count = history_video_count + new_video_count
if video_count > 1:
gr.Warning("Only one video is supported.")
return False
if video_count == 1:
if image_count > 0:
gr.Warning("Mixing images and videos is not allowed.")
return False
if "<image>" in message["text"]:
gr.Warning("Using <image> tags with video files is not supported.")
return False
if video_count == 0 and image_count > MAX_NUM_IMAGES:
gr.Warning(f"You can upload up to {MAX_NUM_IMAGES} images.")
return False
# ์ด๋ฏธ์ง€ ํƒœ๊ทธ ๊ฒ€์ฆ (์‹ค์ œ ์ด๋ฏธ์ง€ ํŒŒ์ผ๋งŒ ๊ณ„์‚ฐ)
if "<image>" in message["text"]:
# ์ด๋ฏธ์ง€ ํŒŒ์ผ๋งŒ ํ•„ํ„ฐ๋ง
image_files = [f for f in message["files"] if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE)]
image_tag_count = message["text"].count("<image>")
if image_tag_count != len(image_files):
gr.Warning("The number of <image> tags in the text does not match the number of image files.")
return False
return True
##################################################
# ๋น„๋””์˜ค ์ฒ˜๋ฆฌ
##################################################
def downsample_video(video_path: str) -> list[tuple[Image.Image, float]]:
vidcap = cv2.VideoCapture(video_path)
fps = vidcap.get(cv2.CAP_PROP_FPS)
total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
# ๋” ์ ์€ ํ”„๋ ˆ์ž„์„ ์ถ”์ถœํ•˜๋„๋ก ์กฐ์ •
frame_interval = max(int(fps), int(total_frames / 10)) # ์ดˆ๋‹น 1ํ”„๋ ˆ์ž„ ๋˜๋Š” ์ตœ๋Œ€ 10ํ”„๋ ˆ์ž„
frames = []
for i in range(0, total_frames, frame_interval):
vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
success, image = vidcap.read()
if success:
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
pil_image = Image.fromarray(image)
timestamp = round(i / fps, 2)
frames.append((pil_image, timestamp))
# ์ตœ๋Œ€ 5ํ”„๋ ˆ์ž„๋งŒ ์‚ฌ์šฉ
if len(frames) >= 5:
break
vidcap.release()
return frames
def process_video(video_path: str) -> list[dict]:
content = []
frames = downsample_video(video_path)
for frame in frames:
pil_image, timestamp = frame
with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as temp_file:
pil_image.save(temp_file.name)
content.append({"type": "text", "text": f"Frame {timestamp}:"})
content.append({"type": "image", "url": temp_file.name})
logger.debug(f"{content=}")
return content
##################################################
# interleaved <image> ์ฒ˜๋ฆฌ
##################################################
def process_interleaved_images(message: dict) -> list[dict]:
parts = re.split(r"(<image>)", message["text"])
content = []
image_index = 0
# ์ด๋ฏธ์ง€ ํŒŒ์ผ๋งŒ ํ•„ํ„ฐ๋ง
image_files = [f for f in message["files"] if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE)]
for part in parts:
if part == "<image>" and image_index < len(image_files):
content.append({"type": "image", "url": image_files[image_index]})
image_index += 1
elif part.strip():
content.append({"type": "text", "text": part.strip()})
else:
# ๊ณต๋ฐฑ์ด๊ฑฐ๋‚˜ \n ๊ฐ™์€ ๊ฒฝ์šฐ
if isinstance(part, str) and part != "<image>":
content.append({"type": "text", "text": part})
return content
##################################################
# PDF + CSV + TXT + ์ด๋ฏธ์ง€/๋น„๋””์˜ค
##################################################
def is_image_file(file_path: str) -> bool:
"""์ด๋ฏธ์ง€ ํŒŒ์ผ์ธ์ง€ ํ™•์ธ"""
return bool(re.search(r"\.(png|jpg|jpeg|gif|webp)$", file_path, re.IGNORECASE))
def is_video_file(file_path: str) -> bool:
"""๋น„๋””์˜ค ํŒŒ์ผ์ธ์ง€ ํ™•์ธ"""
return file_path.endswith(".mp4")
def is_document_file(file_path: str) -> bool:
"""๋ฌธ์„œ ํŒŒ์ผ์ธ์ง€ ํ™•์ธ (PDF, CSV, TXT)"""
return (file_path.lower().endswith(".pdf") or
file_path.lower().endswith(".csv") or
file_path.lower().endswith(".txt"))
def process_new_user_message(message: dict) -> list[dict]:
if not message["files"]:
return [{"type": "text", "text": message["text"]}]
# 1) ํŒŒ์ผ ๋ถ„๋ฅ˜
video_files = [f for f in message["files"] if is_video_file(f)]
image_files = [f for f in message["files"] if is_image_file(f)]
csv_files = [f for f in message["files"] if f.lower().endswith(".csv")]
txt_files = [f for f in message["files"] if f.lower().endswith(".txt")]
pdf_files = [f for f in message["files"] if f.lower().endswith(".pdf")]
# 2) ์‚ฌ์šฉ์ž ์›๋ณธ text ์ถ”๊ฐ€
content_list = [{"type": "text", "text": message["text"]}]
# 3) CSV
for csv_path in csv_files:
csv_analysis = analyze_csv_file(csv_path)
content_list.append({"type": "text", "text": csv_analysis})
# 4) TXT
for txt_path in txt_files:
txt_analysis = analyze_txt_file(txt_path)
content_list.append({"type": "text", "text": txt_analysis})
# 5) PDF
for pdf_path in pdf_files:
pdf_markdown = pdf_to_markdown(pdf_path)
content_list.append({"type": "text", "text": pdf_markdown})
# 6) ๋น„๋””์˜ค (ํ•œ ๊ฐœ๋งŒ ํ—ˆ์šฉ)
if video_files:
content_list += process_video(video_files[0])
return content_list
# 7) ์ด๋ฏธ์ง€ ์ฒ˜๋ฆฌ
if "<image>" in message["text"] and image_files:
# interleaved
interleaved_content = process_interleaved_images({"text": message["text"], "files": image_files})
# ์›๋ณธ content_list ์•ž๋ถ€๋ถ„(ํ…์ŠคํŠธ)์„ ์ œ๊ฑฐํ•˜๊ณ  interleaved๋กœ ๋Œ€์ฒด
if content_list[0]["type"] == "text":
content_list = content_list[1:] # ์›๋ณธ ํ…์ŠคํŠธ ์ œ๊ฑฐ
return interleaved_content + content_list # interleaved + ๋‚˜๋จธ์ง€ ๋ฌธ์„œ ๋ถ„์„ ๋‚ด์šฉ
else:
# ์ผ๋ฐ˜ ์—ฌ๋Ÿฌ ์žฅ
for img_path in image_files:
content_list.append({"type": "image", "url": img_path})
return content_list
##################################################
# history -> LLM ๋ฉ”์‹œ์ง€ ๋ณ€ํ™˜
##################################################
def process_history(history: list[dict]) -> list[dict]:
messages = []
current_user_content: list[dict] = []
for item in history:
if item["role"] == "assistant":
# user_content๊ฐ€ ์Œ“์—ฌ์žˆ๋‹ค๋ฉด user ๋ฉ”์‹œ์ง€๋กœ ์ €์žฅ
if current_user_content:
messages.append({"role": "user", "content": current_user_content})
current_user_content = []
# ๊ทธ ๋’ค item์€ assistant
messages.append({"role": "assistant", "content": [{"type": "text", "text": item["content"]}]})
else:
# user
content = item["content"]
if isinstance(content, str):
current_user_content.append({"type": "text", "text": content})
elif isinstance(content, list) and len(content) > 0:
file_path = content[0]
if is_image_file(file_path):
current_user_content.append({"type": "image", "url": file_path})
else:
# ๋น„์ด๋ฏธ์ง€ ํŒŒ์ผ์€ ํ…์ŠคํŠธ๋กœ ์ฒ˜๋ฆฌ
current_user_content.append({"type": "text", "text": f"[File: {os.path.basename(file_path)}]"})
# ๋งˆ์ง€๋ง‰ ์‚ฌ์šฉ์ž ๋ฉ”์‹œ์ง€๊ฐ€ ์ฒ˜๋ฆฌ๋˜์ง€ ์•Š์€ ๊ฒฝ์šฐ ์ถ”๊ฐ€
if current_user_content:
messages.append({"role": "user", "content": current_user_content})
return messages
##################################################
# ๋ฉ”์ธ ์ถ”๋ก  ํ•จ์ˆ˜
##################################################
@spaces.GPU(duration=120)
def run(message: dict, history: list[dict], system_prompt: str = "", max_new_tokens: int = 512) -> Iterator[str]:
if not validate_media_constraints(message, history):
yield ""
return
try:
messages = []
if system_prompt:
messages.append({"role": "system", "content": [{"type": "text", "text": system_prompt}]})
messages.extend(process_history(history))
# ์‚ฌ์šฉ์ž ๋ฉ”์‹œ์ง€ ์ฒ˜๋ฆฌ
user_content = process_new_user_message(message)
# ํ† ํฐ ์ˆ˜๋ฅผ ์ค„์ด๊ธฐ ์œ„ํ•ด ๋„ˆ๋ฌด ๊ธด ํ…์ŠคํŠธ๋Š” ์ž˜๋ผ๋‚ด๊ธฐ
for item in user_content:
if item["type"] == "text" and len(item["text"]) > MAX_CONTENT_CHARS:
item["text"] = item["text"][:MAX_CONTENT_CHARS] + "\n...(truncated)..."
messages.append({"role": "user", "content": user_content})
# ๋ชจ๋ธ ์ž…๋ ฅ ์ƒ์„ฑ ์ „ ์ตœ์ข… ํ™•์ธ
# ์ด๋ฏธ์ง€๋‚˜ ๋น„๋””์˜ค๊ฐ€ ์•„๋‹Œ ํŒŒ์ผ๋“ค์€ ๋ชจ๋ธ์˜ "image" ํŒŒ์ดํ”„๋ผ์ธ์œผ๋กœ ์ „๋‹ฌ๋˜์ง€ ์•Š๋„๋ก ํ•„ํ„ฐ๋ง
for msg in messages:
if msg["role"] != "user":
continue
filtered_content = []
for item in msg["content"]:
if item["type"] == "image":
if is_image_file(item["url"]):
filtered_content.append(item)
else:
# ์ด๋ฏธ์ง€ ํŒŒ์ผ์ด ์•„๋‹Œ ๊ฒฝ์šฐ ํ…์ŠคํŠธ๋กœ ๋ณ€ํ™˜
filtered_content.append({
"type": "text",
"text": f"[Non-image file: {os.path.basename(item['url'])}]"
})
else:
filtered_content.append(item)
msg["content"] = filtered_content
# ๋ชจ๋ธ ์ž…๋ ฅ ์ƒ์„ฑ
inputs = processor.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt",
).to(device=model.device, dtype=torch.bfloat16)
# ํ…์ŠคํŠธ ์ƒ์„ฑ ์ŠคํŠธ๋ฆฌ๋จธ ์„ค์ •
streamer = TextIteratorStreamer(processor, timeout=30.0, skip_prompt=True, skip_special_tokens=True)
gen_kwargs = dict(
inputs,
streamer=streamer,
max_new_tokens=max_new_tokens,
)
# ๋ณ„๋„ ์Šค๋ ˆ๋“œ์—์„œ ํ…์ŠคํŠธ ์ƒ์„ฑ
t = Thread(target=model.generate, kwargs=gen_kwargs)
t.start()
# ๊ฒฐ๊ณผ ์ŠคํŠธ๋ฆฌ๋ฐ
output = ""
for new_text in streamer:
output += new_text
yield output
except Exception as e:
logger.error(f"Error in run: {str(e)}")
yield f"์ฃ„์†กํ•ฉ๋‹ˆ๋‹ค. ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ–ˆ์Šต๋‹ˆ๋‹ค: {str(e)}"
##################################################
# ์˜ˆ์‹œ๋“ค (ํ•œ๊ธ€ํ™” ๋ฒ„์ „)
##################################################
examples = [
[
{
"text": "PDF ํŒŒ์ผ ๋‚ด์šฉ์„ ์š”์•ฝ, ๋ถ„์„ํ•˜๋ผ.",
"files": ["assets/additional-examples/pdf.pdf"],
}
],
[
{
"text": "CSV ํŒŒ์ผ ๋‚ด์šฉ์„ ์š”์•ฝ, ๋ถ„์„ํ•˜๋ผ",
"files": ["assets/additional-examples/sample-csv.csv"],
}
],
[
{
"text": "๋™์ผํ•œ ๋ง‰๋Œ€ ๊ทธ๋ž˜ํ”„๋ฅผ ๊ทธ๋ฆฌ๋Š” matplotlib ์ฝ”๋“œ๋ฅผ ์ž‘์„ฑํ•ด์ฃผ์„ธ์š”.",
"files": ["assets/additional-examples/barchart.png"],
}
],
[
{
"text": "์ด ์˜์ƒ์—์„œ ์ด์ƒํ•œ ์ ์ด ๋ฌด์—‡์ธ๊ฐ€์š”?",
"files": ["assets/additional-examples/tmp.mp4"],
}
],
[
{
"text": "์ด๋ฏธ ์ด ์˜์–‘์ œ๋ฅผ <image> ๊ฐ€์ง€๊ณ  ์žˆ๊ณ , ์ด ์ œํ’ˆ <image>์„ ์ƒˆ๋กœ ์‚ฌ๋ ค ํ•ฉ๋‹ˆ๋‹ค. ํ•จ๊ป˜ ์„ญ์ทจํ•  ๋•Œ ์ฃผ์˜ํ•ด์•ผ ํ•  ์ ์ด ์žˆ์„๊นŒ์š”?",
"files": ["assets/additional-examples/pill1.png", "assets/additional-examples/pill2.png"],
}
],
[
{
"text": "์ด๋ฏธ์ง€์˜ ์‹œ๊ฐ์  ์š”์†Œ์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ์‹œ๋ฅผ ์ž‘์„ฑํ•ด์ฃผ์„ธ์š”.",
"files": ["assets/sample-images/06-1.png", "assets/sample-images/06-2.png"],
}
],
[
{
"text": "์ด๋ฏธ์ง€์˜ ์‹œ๊ฐ์  ์š”์†Œ๋ฅผ ํ† ๋Œ€๋กœ ์งง์€ ์•…๊ณก์„ ์ž‘๊ณกํ•ด์ฃผ์„ธ์š”.",
"files": [
"assets/sample-images/07-1.png",
"assets/sample-images/07-2.png",
"assets/sample-images/07-3.png",
"assets/sample-images/07-4.png",
],
}
],
[
{
"text": "์ด ์ง‘์—์„œ ๋ฌด์Šจ ์ผ์ด ์žˆ์—ˆ์„์ง€ ์งง์€ ์ด์•ผ๊ธฐ๋ฅผ ์ง€์–ด๋ณด์„ธ์š”.",
"files": ["assets/sample-images/08.png"],
}
],
[
{
"text": "์ด๋ฏธ์ง€๋“ค์˜ ์ˆœ์„œ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์งง์€ ์ด์•ผ๊ธฐ๋ฅผ ๋งŒ๋“ค์–ด ์ฃผ์„ธ์š”.",
"files": [
"assets/sample-images/09-1.png",
"assets/sample-images/09-2.png",
"assets/sample-images/09-3.png",
"assets/sample-images/09-4.png",
"assets/sample-images/09-5.png",
],
}
],
[
{
"text": "์ด ์„ธ๊ณ„์—์„œ ์‚ด๊ณ  ์žˆ์„ ์ƒ๋ฌผ๋“ค์„ ์ƒ์ƒํ•ด์„œ ๋ฌ˜์‚ฌํ•ด์ฃผ์„ธ์š”.",
"files": ["assets/sample-images/10.png"],
}
],
[
{
"text": "์ด๋ฏธ์ง€์— ์ ํžŒ ํ…์ŠคํŠธ๋ฅผ ์ฝ์–ด์ฃผ์„ธ์š”.",
"files": ["assets/additional-examples/1.png"],
}
],
[
{
"text": "์ด ํ‹ฐ์ผ“์€ ์–ธ์ œ ๋ฐœ๊ธ‰๋œ ๊ฒƒ์ด๊ณ , ๊ฐ€๊ฒฉ์€ ์–ผ๋งˆ์ธ๊ฐ€์š”?",
"files": ["assets/additional-examples/2.png"],
}
],
[
{
"text": "์ด๋ฏธ์ง€์— ์žˆ๋Š” ํ…์ŠคํŠธ๋ฅผ ๊ทธ๋Œ€๋กœ ์ฝ์–ด์„œ ๋งˆํฌ๋‹ค์šด ํ˜•ํƒœ๋กœ ์ ์–ด์ฃผ์„ธ์š”.",
"files": ["assets/additional-examples/3.png"],
}
],
[
{
"text": "์ด ์ ๋ถ„์„ ํ’€์–ด์ฃผ์„ธ์š”.",
"files": ["assets/additional-examples/4.png"],
}
],
[
{
"text": "์ด ์ด๋ฏธ์ง€๋ฅผ ๊ฐ„๋‹จํžˆ ์บก์…˜์œผ๋กœ ์„ค๋ช…ํ•ด์ฃผ์„ธ์š”.",
"files": ["assets/sample-images/01.png"],
}
],
[
{
"text": "์ด ํ‘œ์ง€ํŒ์—๋Š” ๋ฌด์Šจ ๋ฌธ๊ตฌ๊ฐ€ ์ ํ˜€ ์žˆ๋‚˜์š”?",
"files": ["assets/sample-images/02.png"],
}
],
[
{
"text": "๋‘ ์ด๋ฏธ์ง€๋ฅผ ๋น„๊ตํ•ด์„œ ๊ณตํ†ต์ ๊ณผ ์ฐจ์ด์ ์„ ๋งํ•ด์ฃผ์„ธ์š”.",
"files": ["assets/sample-images/03.png"],
}
],
[
{
"text": "์ด๋ฏธ์ง€์— ๋ณด์ด๋Š” ๋ชจ๋“  ์‚ฌ๋ฌผ๊ณผ ๊ทธ ์ƒ‰์ƒ์„ ๋‚˜์—ดํ•ด์ฃผ์„ธ์š”.",
"files": ["assets/sample-images/04.png"],
}
],
[
{
"text": "์žฅ๋ฉด์˜ ๋ถ„์œ„๊ธฐ๋ฅผ ๋ฌ˜์‚ฌํ•ด์ฃผ์„ธ์š”.",
"files": ["assets/sample-images/05.png"],
}
],
]
##################################################
# Gradio ์ธํ„ฐํŽ˜์ด์Šค ์„ค์ •
##################################################
demo = gr.ChatInterface(
fn=run,
type="messages",
chatbot=gr.Chatbot(type="messages", scale=1, allow_tags=["image"]),
# .webp, .png, .jpg, .jpeg, .gif, .mp4, .csv, .txt, .pdf ๋ชจ๋‘ ํ—ˆ์šฉ
textbox=gr.MultimodalTextbox(
file_types=[
".webp", ".png", ".jpg", ".jpeg", ".gif",
".mp4", ".csv", ".txt", ".pdf"
],
file_count="multiple",
autofocus=True
),
multimodal=True,
additional_inputs=[
gr.Textbox(
label="System Prompt",
value=(
"You are a deeply thoughtful AI. Consider problems thoroughly and derive "
"correct solutions through systematic reasoning. Please answer in korean."
)
),
gr.Slider(label="Max New Tokens", minimum=100, maximum=8000, step=50, value=2000),
],
stop_btn=False,
title="Vidraft-Gemma-3-27B",
examples=examples,
run_examples_on_click=False,
cache_examples=False,
css_paths="style.css",
delete_cache=(1800, 1800),
)
if __name__ == "__main__":
demo.launch()