Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,35 +1,32 @@
|
|
| 1 |
import gradio as gr
|
| 2 |
import torch
|
| 3 |
-
import os
|
| 4 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 5 |
from huggingface_hub import hf_hub_download
|
| 6 |
from functools import lru_cache
|
| 7 |
|
| 8 |
-
# ---
|
| 9 |
-
#
|
| 10 |
MODEL_NAME = "Kenan023214/PyroNet-mini"
|
| 11 |
-
DEVICE = "cpu" #
|
| 12 |
MAX_NEW_TOKENS = 256
|
| 13 |
MAX_CONTEXT_TOKENS = 2048
|
| 14 |
|
| 15 |
-
# Загрузка модели и токенизатора
|
| 16 |
@lru_cache(maxsize=1)
|
| 17 |
def load_model():
|
| 18 |
-
"""
|
| 19 |
print("Loading model and tokenizer...")
|
| 20 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
| 21 |
model = AutoModelForCausalLM.from_pretrained(
|
| 22 |
MODEL_NAME,
|
| 23 |
device_map=DEVICE,
|
| 24 |
-
torch_dtype=torch.float32 #
|
| 25 |
)
|
| 26 |
print("Model loaded.")
|
| 27 |
return tokenizer, model
|
| 28 |
|
| 29 |
-
# Загрузка файлов шаблонов из репозитория
|
| 30 |
@lru_cache(maxsize=1)
|
| 31 |
def download_templates():
|
| 32 |
-
"""
|
| 33 |
print("Downloading chat templates...")
|
| 34 |
for lang in ["ru", "en", "uk"]:
|
| 35 |
hf_hub_download(
|
|
@@ -43,13 +40,13 @@ def download_templates():
|
|
| 43 |
tokenizer, model = load_model()
|
| 44 |
download_templates()
|
| 45 |
|
| 46 |
-
# ---
|
| 47 |
def num_tokens_of_text(text: str) -> int:
|
| 48 |
-
"""
|
| 49 |
return len(tokenizer.encode(text, add_special_tokens=False))
|
| 50 |
|
| 51 |
def trim_history_to_max_tokens(messages, max_tokens):
|
| 52 |
-
"""
|
| 53 |
rev = list(reversed(messages))
|
| 54 |
total = 0
|
| 55 |
kept = []
|
|
@@ -62,7 +59,7 @@ def trim_history_to_max_tokens(messages, max_tokens):
|
|
| 62 |
return list(reversed(kept))
|
| 63 |
|
| 64 |
def build_messages_for_template(history_messages, reasoning: bool, language: str):
|
| 65 |
-
"""
|
| 66 |
if language == 'ru':
|
| 67 |
system_message = "Ты — дружелюбный ассистент, который говорит на русском. Отвечай кратко, но по делу."
|
| 68 |
reasoning_instruction = ("[REASONING MODE]\n"
|
|
@@ -87,7 +84,7 @@ def build_messages_for_template(history_messages, reasoning: bool, language: str
|
|
| 87 |
return messages
|
| 88 |
|
| 89 |
def extract_assistant_reply(raw_generated_text: str) -> str:
|
| 90 |
-
"""
|
| 91 |
text = raw_generated_text
|
| 92 |
if "<|assistant|>" in text:
|
| 93 |
text = text.split("<|assistant|>")[-1]
|
|
@@ -95,23 +92,19 @@ def extract_assistant_reply(raw_generated_text: str) -> str:
|
|
| 95 |
text = text.replace(tag, "")
|
| 96 |
return text.strip()
|
| 97 |
|
| 98 |
-
# ---
|
| 99 |
def generate_response(user_text: str, history, reasoning: bool, language: str):
|
| 100 |
-
"""
|
| 101 |
|
| 102 |
-
# Добавляем user-сообщение в историю
|
| 103 |
history.append({"role": "user", "content": user_text})
|
| 104 |
|
| 105 |
-
# Подрезаем историю, чтобы вход не стал слишком большим
|
| 106 |
trimmed_history = trim_history_to_max_tokens(history, MAX_CONTEXT_TOKENS)
|
| 107 |
|
| 108 |
-
# Собираем messages с возможной инструкцией reasoning
|
| 109 |
messages_for_template = build_messages_for_template(trimmed_history, reasoning, language)
|
| 110 |
|
| 111 |
-
#
|
| 112 |
template_file = f"chat_template_{language}.jinja"
|
| 113 |
|
| 114 |
-
# Применяем шаблон и токенизируем
|
| 115 |
text = tokenizer.apply_chat_template(
|
| 116 |
messages_for_template,
|
| 117 |
template_path=template_file,
|
|
@@ -121,7 +114,6 @@ def generate_response(user_text: str, history, reasoning: bool, language: str):
|
|
| 121 |
|
| 122 |
inputs = tokenizer(text, return_tensors="pt").to(DEVICE)
|
| 123 |
|
| 124 |
-
# Генерация
|
| 125 |
with torch.no_grad():
|
| 126 |
outputs = model.generate(
|
| 127 |
**inputs,
|
|
@@ -132,48 +124,40 @@ def generate_response(user_text: str, history, reasoning: bool, language: str):
|
|
| 132 |
pad_token_id=tokenizer.eos_token_id
|
| 133 |
)
|
| 134 |
|
| 135 |
-
# Декодируем и очищаем ответ
|
| 136 |
raw = tokenizer.decode(outputs[0], skip_special_tokens=False)
|
| 137 |
reply = extract_assistant_reply(raw)
|
| 138 |
|
| 139 |
-
# Добавляем ассистента в историю
|
| 140 |
history.append({"role": "assistant", "content": reply})
|
| 141 |
|
| 142 |
-
# Gradio ожидает возвращение списка [пользователь, ассистент]
|
| 143 |
-
# Мы возвращаем всю историю для корректного отображения
|
| 144 |
return "", history
|
| 145 |
|
| 146 |
-
# ---
|
| 147 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
| 148 |
gr.Markdown("# PyroNet-mini Chat")
|
| 149 |
-
gr.Markdown("
|
| 150 |
|
| 151 |
chatbot = gr.Chatbot(height=500)
|
| 152 |
|
| 153 |
with gr.Row():
|
| 154 |
with gr.Column(scale=4):
|
| 155 |
msg = gr.Textbox(
|
| 156 |
-
label="
|
| 157 |
-
placeholder="
|
| 158 |
container=False
|
| 159 |
)
|
| 160 |
with gr.Column(scale=1, min_width=100):
|
| 161 |
language_dropdown = gr.Dropdown(
|
| 162 |
choices=["ru", "en", "uk"],
|
| 163 |
-
value="
|
| 164 |
-
label="
|
| 165 |
container=False
|
| 166 |
)
|
| 167 |
reasoning_checkbox = gr.Checkbox(
|
| 168 |
-
label="
|
| 169 |
)
|
| 170 |
|
| 171 |
-
btn_send = gr.Button("
|
| 172 |
-
btn_clear = gr.Button("
|
| 173 |
-
|
| 174 |
-
# Обработчики событий
|
| 175 |
-
def reset_history():
|
| 176 |
-
return [], None
|
| 177 |
|
| 178 |
btn_send.click(
|
| 179 |
fn=generate_response,
|
|
@@ -193,4 +177,3 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
|
| 193 |
|
| 194 |
if __name__ == "__main__":
|
| 195 |
demo.launch()
|
| 196 |
-
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
import torch
|
|
|
|
| 3 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 4 |
from huggingface_hub import hf_hub_download
|
| 5 |
from functools import lru_cache
|
| 6 |
|
| 7 |
+
# --- Hugging Face Space Configuration ---
|
| 8 |
+
# Load the model and tokenizer only once when the app starts
|
| 9 |
MODEL_NAME = "Kenan023214/PyroNet-mini"
|
| 10 |
+
DEVICE = "cpu" # Use CPU for basic Space
|
| 11 |
MAX_NEW_TOKENS = 256
|
| 12 |
MAX_CONTEXT_TOKENS = 2048
|
| 13 |
|
|
|
|
| 14 |
@lru_cache(maxsize=1)
|
| 15 |
def load_model():
|
| 16 |
+
"""Loads the model and tokenizer, caching them for performance."""
|
| 17 |
print("Loading model and tokenizer...")
|
| 18 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
| 19 |
model = AutoModelForCausalLM.from_pretrained(
|
| 20 |
MODEL_NAME,
|
| 21 |
device_map=DEVICE,
|
| 22 |
+
torch_dtype=torch.float32 # Use float32 for CPU compatibility
|
| 23 |
)
|
| 24 |
print("Model loaded.")
|
| 25 |
return tokenizer, model
|
| 26 |
|
|
|
|
| 27 |
@lru_cache(maxsize=1)
|
| 28 |
def download_templates():
|
| 29 |
+
"""Downloads template files from the model repository."""
|
| 30 |
print("Downloading chat templates...")
|
| 31 |
for lang in ["ru", "en", "uk"]:
|
| 32 |
hf_hub_download(
|
|
|
|
| 40 |
tokenizer, model = load_model()
|
| 41 |
download_templates()
|
| 42 |
|
| 43 |
+
# --- Utilities ---
|
| 44 |
def num_tokens_of_text(text: str) -> int:
|
| 45 |
+
"""Approximate number of tokens for a given text."""
|
| 46 |
return len(tokenizer.encode(text, add_special_tokens=False))
|
| 47 |
|
| 48 |
def trim_history_to_max_tokens(messages, max_tokens):
|
| 49 |
+
"""Trims the message history to fit within a token limit."""
|
| 50 |
rev = list(reversed(messages))
|
| 51 |
total = 0
|
| 52 |
kept = []
|
|
|
|
| 59 |
return list(reversed(kept))
|
| 60 |
|
| 61 |
def build_messages_for_template(history_messages, reasoning: bool, language: str):
|
| 62 |
+
"""Prepares messages for the chat template."""
|
| 63 |
if language == 'ru':
|
| 64 |
system_message = "Ты — дружелюбный ассистент, который говорит на русском. Отвечай кратко, но по делу."
|
| 65 |
reasoning_instruction = ("[REASONING MODE]\n"
|
|
|
|
| 84 |
return messages
|
| 85 |
|
| 86 |
def extract_assistant_reply(raw_generated_text: str) -> str:
|
| 87 |
+
"""Removes extra tokens and returns only the assistant's reply."""
|
| 88 |
text = raw_generated_text
|
| 89 |
if "<|assistant|>" in text:
|
| 90 |
text = text.split("<|assistant|>")[-1]
|
|
|
|
| 92 |
text = text.replace(tag, "")
|
| 93 |
return text.strip()
|
| 94 |
|
| 95 |
+
# --- Main function for Gradio ---
|
| 96 |
def generate_response(user_text: str, history, reasoning: bool, language: str):
|
| 97 |
+
"""Processes user input and generates a response."""
|
| 98 |
|
|
|
|
| 99 |
history.append({"role": "user", "content": user_text})
|
| 100 |
|
|
|
|
| 101 |
trimmed_history = trim_history_to_max_tokens(history, MAX_CONTEXT_TOKENS)
|
| 102 |
|
|
|
|
| 103 |
messages_for_template = build_messages_for_template(trimmed_history, reasoning, language)
|
| 104 |
|
| 105 |
+
# Select the template file from the local files
|
| 106 |
template_file = f"chat_template_{language}.jinja"
|
| 107 |
|
|
|
|
| 108 |
text = tokenizer.apply_chat_template(
|
| 109 |
messages_for_template,
|
| 110 |
template_path=template_file,
|
|
|
|
| 114 |
|
| 115 |
inputs = tokenizer(text, return_tensors="pt").to(DEVICE)
|
| 116 |
|
|
|
|
| 117 |
with torch.no_grad():
|
| 118 |
outputs = model.generate(
|
| 119 |
**inputs,
|
|
|
|
| 124 |
pad_token_id=tokenizer.eos_token_id
|
| 125 |
)
|
| 126 |
|
|
|
|
| 127 |
raw = tokenizer.decode(outputs[0], skip_special_tokens=False)
|
| 128 |
reply = extract_assistant_reply(raw)
|
| 129 |
|
|
|
|
| 130 |
history.append({"role": "assistant", "content": reply})
|
| 131 |
|
|
|
|
|
|
|
| 132 |
return "", history
|
| 133 |
|
| 134 |
+
# --- Gradio Interface ---
|
| 135 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
| 136 |
gr.Markdown("# PyroNet-mini Chat")
|
| 137 |
+
gr.Markdown("A demonstration of PyroNet-mini (based on a custom model) with multilingual templates and a reasoning mode.")
|
| 138 |
|
| 139 |
chatbot = gr.Chatbot(height=500)
|
| 140 |
|
| 141 |
with gr.Row():
|
| 142 |
with gr.Column(scale=4):
|
| 143 |
msg = gr.Textbox(
|
| 144 |
+
label="Your Prompt",
|
| 145 |
+
placeholder="Write your message here...",
|
| 146 |
container=False
|
| 147 |
)
|
| 148 |
with gr.Column(scale=1, min_width=100):
|
| 149 |
language_dropdown = gr.Dropdown(
|
| 150 |
choices=["ru", "en", "uk"],
|
| 151 |
+
value="en",
|
| 152 |
+
label="Language",
|
| 153 |
container=False
|
| 154 |
)
|
| 155 |
reasoning_checkbox = gr.Checkbox(
|
| 156 |
+
label="Enable Reasoning Mode"
|
| 157 |
)
|
| 158 |
|
| 159 |
+
btn_send = gr.Button("Send")
|
| 160 |
+
btn_clear = gr.Button("Clear")
|
|
|
|
|
|
|
|
|
|
|
|
|
| 161 |
|
| 162 |
btn_send.click(
|
| 163 |
fn=generate_response,
|
|
|
|
| 177 |
|
| 178 |
if __name__ == "__main__":
|
| 179 |
demo.launch()
|
|
|