Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -1,118 +1,161 @@
|
|
1 |
import os
|
|
|
|
|
2 |
import gradio as gr
|
3 |
from gradio_client import Client, handle_file
|
4 |
from openai import OpenAI
|
5 |
|
6 |
-
# ---
|
7 |
NV_API_KEY = os.environ.get("NV_API_KEY")
|
8 |
if not NV_API_KEY:
|
9 |
-
raise
|
10 |
|
11 |
-
#
|
12 |
florence = Client("gokaygokay/Florence-2")
|
13 |
|
14 |
-
#
|
15 |
-
llm = OpenAI(
|
16 |
-
|
17 |
-
api_key=NV_API_KEY
|
18 |
-
)
|
19 |
|
20 |
-
|
21 |
-
|
22 |
-
"""Делаем подробную подпись через Florence-2."""
|
23 |
try:
|
|
|
24 |
result = florence.predict(
|
25 |
image=handle_file(image_path),
|
26 |
task_prompt="More Detailed Caption",
|
27 |
text_input=None,
|
28 |
model_id="microsoft/Florence-2-large",
|
29 |
-
api_name="/process_image"
|
30 |
)
|
|
|
31 |
return result if isinstance(result, str) else str(result)
|
32 |
except Exception as e:
|
33 |
return f"[Ошибка при генерации подписи: {e}]"
|
34 |
|
35 |
-
|
36 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
if not image_path:
|
38 |
-
|
|
|
|
|
39 |
|
|
|
40 |
caption = get_caption(image_path)
|
41 |
|
|
|
42 |
system_prompt = (
|
43 |
-
"
|
44 |
-
|
45 |
-
"
|
|
|
|
|
46 |
)
|
47 |
|
48 |
-
|
49 |
history.append([user_message, ""])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
50 |
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
#
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
["https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/flowers.png"],
|
76 |
]
|
77 |
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
)
|
83 |
|
|
|
|
|
84 |
with gr.Row():
|
85 |
with gr.Column(scale=4):
|
86 |
-
image_input = gr.Image(
|
87 |
-
|
88 |
-
|
89 |
-
label="Примеры",
|
90 |
-
columns=4,
|
91 |
-
height="auto",
|
92 |
-
preview=True
|
93 |
-
)
|
94 |
-
user_input = gr.Textbox(label="Ваш вопрос", placeholder="Например: Что изображено на фото?")
|
95 |
send_btn = gr.Button("Отправить")
|
|
|
|
|
|
|
96 |
|
97 |
with gr.Column(scale=6):
|
98 |
-
chatbot = gr.Chatbot(label="Чат", height=
|
99 |
-
clear_btn = gr.Button("Очистить чат")
|
100 |
|
101 |
-
#
|
102 |
-
def
|
103 |
-
return
|
104 |
|
105 |
-
gallery.select(
|
106 |
|
107 |
-
|
108 |
-
|
109 |
-
inputs=[image_input, user_input, chatbot],
|
110 |
-
outputs=[chatbot]
|
111 |
-
)
|
112 |
|
113 |
-
clear_btn.click(lambda:
|
114 |
|
115 |
-
# Запуск
|
116 |
if __name__ == "__main__":
|
117 |
-
demo.launch()
|
118 |
|
|
|
1 |
import os
|
2 |
+
from typing import Generator, List, Tuple
|
3 |
+
|
4 |
import gradio as gr
|
5 |
from gradio_client import Client, handle_file
|
6 |
from openai import OpenAI
|
7 |
|
8 |
+
# --- Конфигурация (в HF Spaces добавь NV_API_KEY в Secrets) ---
|
9 |
NV_API_KEY = os.environ.get("NV_API_KEY")
|
10 |
if not NV_API_KEY:
|
11 |
+
raise RuntimeError("Добавьте NV_API_KEY в Secrets Hugging Face Space")
|
12 |
|
13 |
+
# Florence-2 (публичный wrapper)
|
14 |
florence = Client("gokaygokay/Florence-2")
|
15 |
|
16 |
+
# OpenAI-compatible client (NVIDIA integrate)
|
17 |
+
llm = OpenAI(base_url="https://integrate.api.nvidia.com/v1", api_key=NV_API_KEY)
|
18 |
+
|
|
|
|
|
19 |
|
20 |
+
def get_caption(image_path: str) -> str:
|
21 |
+
"""Запрос 'More Detailed Caption' к Florence-2. image_path может быть URL или локальный путь."""
|
|
|
22 |
try:
|
23 |
+
# handle_file поддерживает URL и локальные файлы
|
24 |
result = florence.predict(
|
25 |
image=handle_file(image_path),
|
26 |
task_prompt="More Detailed Caption",
|
27 |
text_input=None,
|
28 |
model_id="microsoft/Florence-2-large",
|
29 |
+
api_name="/process_image",
|
30 |
)
|
31 |
+
# result может быть строкой или структурой — нормализуем
|
32 |
return result if isinstance(result, str) else str(result)
|
33 |
except Exception as e:
|
34 |
return f"[Ошибка при генерации подписи: {e}]"
|
35 |
|
36 |
+
|
37 |
+
def _extract_text_from_chunk(chunk) -> str:
|
38 |
+
"""Универсальная попытка извлечь текстовый фрагмент из stream-chunk."""
|
39 |
+
try:
|
40 |
+
# объект-атрибутный стиль
|
41 |
+
if hasattr(chunk, "choices"):
|
42 |
+
choice = chunk.choices[0]
|
43 |
+
delta = getattr(choice, "delta", None)
|
44 |
+
if delta is not None:
|
45 |
+
txt = getattr(delta, "content", None) or getattr(delta, "reasoning_content", None)
|
46 |
+
return txt or ""
|
47 |
+
# dict-стиль
|
48 |
+
if isinstance(chunk, dict):
|
49 |
+
choices = chunk.get("choices", [])
|
50 |
+
if choices:
|
51 |
+
delta = choices[0].get("delta", {})
|
52 |
+
return delta.get("content") or delta.get("reasoning_content") or ""
|
53 |
+
except Exception:
|
54 |
+
return ""
|
55 |
+
return ""
|
56 |
+
|
57 |
+
|
58 |
+
def chat_stream(image_path: str, user_message: str, history: List[Tuple[str, str]]):
|
59 |
+
"""
|
60 |
+
Generator для Gradio: сначала возвращает caption, затем по мере прихода токенов
|
61 |
+
обновляет последний ответ ассистента.
|
62 |
+
Возвращаемые значения — кортежи (history, caption) соответствующие outputs.
|
63 |
+
"""
|
64 |
+
history = history or []
|
65 |
+
|
66 |
if not image_path:
|
67 |
+
history.append([user_message, "Пожалуйста, загрузите изображение."])
|
68 |
+
yield history, ""
|
69 |
+
return
|
70 |
|
71 |
+
# Получаем подробную подпись
|
72 |
caption = get_caption(image_path)
|
73 |
|
74 |
+
# Сборка системного промпта
|
75 |
system_prompt = (
|
76 |
+
"You are 'multimodal gpt-oss 120b'. Use the provided 'More Detailed Caption' as authoritative visual context.\n\n"
|
77 |
+
"Image Caption START >>>\n"
|
78 |
+
f"{caption}\n"
|
79 |
+
"<<< Image Caption END.\n"
|
80 |
+
"When answering, mention visible details and be explicit when uncertain."
|
81 |
)
|
82 |
|
83 |
+
# Добавляем сообщение пользователя
|
84 |
history.append([user_message, ""])
|
85 |
+
# Первый yield — чтобы UI сразу показал пользовательское сообщение и подпись
|
86 |
+
yield history, caption
|
87 |
+
|
88 |
+
assistant_text = ""
|
89 |
+
try:
|
90 |
+
stream = llm.chat.completions.create(
|
91 |
+
model="openai/gpt-oss-120b",
|
92 |
+
messages=[
|
93 |
+
{"role": "system", "content": system_prompt},
|
94 |
+
{"role": "user", "content": user_message},
|
95 |
+
],
|
96 |
+
temperature=0.8,
|
97 |
+
top_p=1.0,
|
98 |
+
max_tokens=1024,
|
99 |
+
stream=True,
|
100 |
+
)
|
101 |
|
102 |
+
for chunk in stream:
|
103 |
+
piece = _extract_text_from_chunk(chunk)
|
104 |
+
if not piece:
|
105 |
+
continue
|
106 |
+
assistant_text += piece
|
107 |
+
history[-1][1] = assistant_text
|
108 |
+
yield history, caption
|
109 |
+
|
110 |
+
except Exception as e:
|
111 |
+
# В случае ошибки — покажем её в чате
|
112 |
+
history[-1][1] = f"[Ошибка стриминга LLM: {e}]"
|
113 |
+
yield history, caption
|
114 |
+
|
115 |
+
# Финальный yield (гарантируем состояние завершения)
|
116 |
+
yield history, caption
|
117 |
+
|
118 |
+
|
119 |
+
# --- UI (для HF Spaces) ---
|
120 |
+
EXAMPLE_IMAGES = [
|
121 |
+
# список простых строк (URL или локальные пути). НИКАКИХ вложенных списков!
|
122 |
+
"https://raw.githubusercontent.com/gradio-app/gradio/main/test/test_files/bus.png",
|
123 |
+
"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/cats.png",
|
124 |
+
"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/cheetah.jpg",
|
125 |
+
"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/flowers.png",
|
|
|
126 |
]
|
127 |
|
128 |
+
css = """
|
129 |
+
#title {text-align:center; margin-bottom: -18px;}
|
130 |
+
.gradio-container { max-width: 1100px; margin: auto; }
|
131 |
+
"""
|
|
|
132 |
|
133 |
+
with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
|
134 |
+
gr.Markdown("<h2 id='title'>🖼️ multimodal gpt-oss 120b — визуальный чат</h2>")
|
135 |
with gr.Row():
|
136 |
with gr.Column(scale=4):
|
137 |
+
image_input = gr.Image(label="Загрузите картинку или выберите из галереи", type="filepath", tool="editor")
|
138 |
+
raw_caption = gr.Textbox(label="More Detailed Caption (Florence-2)", interactive=False)
|
139 |
+
user_input = gr.Textbox(label="Вопрос по изображению", placeholder="Например: 'Что происходит на фото?'")
|
|
|
|
|
|
|
|
|
|
|
|
|
140 |
send_btn = gr.Button("Отправить")
|
141 |
+
clear_btn = gr.Button("Очистить чат")
|
142 |
+
gr.Markdown("**Галерея примеров (клик — подставить в загрузчик)**")
|
143 |
+
gallery = gr.Gallery(value=EXAMPLE_IMAGES, columns=4, label="Примеры", show_label=False).style(grid=[4], height="auto")
|
144 |
|
145 |
with gr.Column(scale=6):
|
146 |
+
chatbot = gr.Chatbot(label="Чат с моделью", height=600)
|
|
|
147 |
|
148 |
+
# Клик по картинке в галерее -> вставляем URL/путь в image_input
|
149 |
+
def pick_example(img_url: str):
|
150 |
+
return img_url
|
151 |
|
152 |
+
gallery.select(fn=pick_example, inputs=[gallery], outputs=[image_input])
|
153 |
|
154 |
+
# Кнопка отправки: п��ивязываем генератор, который возвращает (chat_history, caption)
|
155 |
+
send_btn.click(fn=chat_stream, inputs=[image_input, user_input, chatbot], outputs=[chatbot, raw_caption])
|
|
|
|
|
|
|
156 |
|
157 |
+
clear_btn.click(lambda: [], None, chatbot)
|
158 |
|
|
|
159 |
if __name__ == "__main__":
|
160 |
+
demo.launch(server_name="0.0.0.0", server_port=int(os.environ.get("PORT", 7860)))
|
161 |
|