Spaces:
Running
on
Zero
Running
on
Zero
File size: 8,439 Bytes
42f4126 aea4015 42f4126 aea4015 04fa68c aea4015 89817e2 42f4126 aea4015 42f4126 aea4015 42f4126 aea4015 46ef1e4 42f4126 46ef1e4 42f4126 aea4015 42f4126 070daeb 42f4126 46ef1e4 42f4126 46ef1e4 42f4126 aea4015 42f4126 aea4015 42f4126 aea4015 42f4126 aea4015 42f4126 aea4015 42f4126 aea4015 42f4126 aea4015 42f4126 46ef1e4 42f4126 46ef1e4 42f4126 46ef1e4 42f4126 46ef1e4 42f4126 aea4015 46ef1e4 42f4126 46ef1e4 42f4126 46ef1e4 42f4126 46ef1e4 42f4126 46ef1e4 42f4126 aea4015 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 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 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 |
import re
import threading
import gradio as gr
import spaces
import transformers
from transformers import pipeline
# λͺ¨λΈκ³Ό ν ν¬λμ΄μ λ‘λ©
model_name = "CohereForAI/c4ai-command-r7b-arabic-02-2025"
if gr.NO_RELOAD:
pipe = pipeline(
"text-generation",
model=model_name,
device_map="auto",
torch_dtype="auto",
)
# μ΅μ’
λ΅λ³μ κ°μ§νκΈ° μν λ§μ»€
ANSWER_MARKER = "**λ΅λ³**"
# λ¨κ³λ³ μΆλ‘ μ μμνλ λ¬Έμ₯λ€
rethink_prepends = [
"μ, μ΄μ λ€μμ νμ
ν΄μΌ ν©λλ€ ",
"μ μκ°μλ ",
"μ μλ§μ, μ μκ°μλ ",
"λ€μ μ¬νμ΄ λ§λμ§ νμΈν΄ λ³΄κ² μ΅λλ€ ",
"λν κΈ°μ΅ν΄μΌ ν κ²μ ",
"λ λ€λ₯Έ μ£Όλͺ©ν μ μ ",
"κ·Έλ¦¬κ³ μ λ λ€μκ³Ό κ°μ μ¬μ€λ κΈ°μ΅ν©λλ€ ",
"μ΄μ μΆ©λΆν μ΄ν΄νλ€κ³ μκ°ν©λλ€ ",
"μ§κΈκΉμ§μ μ 보λ₯Ό λ°νμΌλ‘, μλ μ§λ¬Έμ μ¬μ©λ μΈμ΄λ‘ λ΅λ³νκ² μ΅λλ€:"
"\n{question}\n"
f"\n{ANSWER_MARKER}\n",
]
# μμ νμ λ¬Έμ ν΄κ²°μ μν μ€μ
latex_delimiters = [
{"left": "$$", "right": "$$", "display": True},
{"left": "$", "right": "$", "display": False},
]
def reformat_math(text):
"""Gradio ꡬ문(Katex)μ μ¬μ©νλλ‘ MathJax κ΅¬λΆ κΈ°νΈ μμ .
μ΄κ²μ Gradioμμ μν 곡μμ νμνκΈ° μν μμ ν΄κ²°μ±
μ
λλ€. νμ¬λ‘μλ
λ€λ₯Έ latex_delimitersλ₯Ό μ¬μ©νμ¬ μμλλ‘ μλνκ² νλ λ°©λ²μ μ°Ύμ§ λͺ»νμ΅λλ€...
"""
text = re.sub(r"\\\[\s*(.*?)\s*\\\]", r"$$\1$$", text, flags=re.DOTALL)
text = re.sub(r"\\\(\s*(.*?)\s*\\\)", r"$\1$", text, flags=re.DOTALL)
return text
def user_input(message, history_original, history_thinking):
"""μ¬μ©μ μ
λ ₯μ νμ€ν 리μ μΆκ°νκ³ μ
λ ₯ ν
μ€νΈ μμ λΉμ°κΈ°"""
return "", history_original + [
gr.ChatMessage(role="user", content=message.replace(ANSWER_MARKER, ""))
], history_thinking + [
gr.ChatMessage(role="user", content=message.replace(ANSWER_MARKER, ""))
]
def rebuild_messages(history: list):
"""μ€κ° μκ° κ³Όμ μμ΄ λͺ¨λΈμ΄ μ¬μ©ν νμ€ν 리μμ λ©μμ§ μ¬κ΅¬μ±"""
messages = []
for h in history:
if isinstance(h, dict) and not h.get("metadata", {}).get("title", False):
messages.append(h)
elif (
isinstance(h, gr.ChatMessage)
and h.metadata.get("title")
and isinstance(h.content, str)
):
messages.append({"role": h.role, "content": h.content})
return messages
@spaces.GPU
def bot_original(
history: list,
max_num_tokens: int,
do_sample: bool,
temperature: float,
):
"""μλ³Έ λͺ¨λΈμ΄ μ§λ¬Έμ λ΅λ³νλλ‘ νκΈ° (μΆλ‘ κ³Όμ μμ΄)"""
# λμ€μ μ€λ λμμ ν ν°μ μ€νΈλ¦ΌμΌλ‘ κ°μ Έμ€κΈ° μν¨
streamer = transformers.TextIteratorStreamer(
pipe.tokenizer, # pyright: ignore
skip_special_tokens=True,
skip_prompt=True,
)
# 보쑰μ λ©μμ§ μ€λΉ
history.append(
gr.ChatMessage(
role="assistant",
content=str(""),
)
)
# νμ¬ μ±ν
μ νμλ λ©μμ§
messages = rebuild_messages(history[:-1]) # λ§μ§λ§ λΉ λ©μμ§ μ μΈ
# μλ³Έ λͺ¨λΈμ μΆλ‘ μμ΄ λ°λ‘ λ΅λ³
t = threading.Thread(
target=pipe,
args=(messages,),
kwargs=dict(
max_new_tokens=max_num_tokens,
streamer=streamer,
do_sample=do_sample,
temperature=temperature,
),
)
t.start()
for token in streamer:
history[-1].content += token
history[-1].content = reformat_math(history[-1].content)
yield history
t.join()
yield history
@spaces.GPU
def bot_thinking(
history: list,
max_num_tokens: int,
final_num_tokens: int,
do_sample: bool,
temperature: float,
):
"""μΆλ‘ κ³Όμ μ ν¬ν¨νμ¬ λͺ¨λΈμ΄ μ§λ¬Έμ λ΅λ³νλλ‘ νκΈ°"""
# λμ€μ μ€λ λμμ ν ν°μ μ€νΈλ¦ΌμΌλ‘ κ°μ Έμ€κΈ° μν¨
streamer = transformers.TextIteratorStreamer(
pipe.tokenizer, # pyright: ignore
skip_special_tokens=True,
skip_prompt=True,
)
# νμν κ²½μ° μΆλ‘ μ μ§λ¬Έμ λ€μ μ½μ
νκΈ° μν¨
question = history[-1]["content"]
# 보쑰μ λ©μμ§ μ€λΉ
history.append(
gr.ChatMessage(
role="assistant",
content=str(""),
metadata={"title": "π§ μκ° μ€...", "status": "pending"},
)
)
# νμ¬ μ±ν
μ νμλ μΆλ‘ κ³Όμ
messages = rebuild_messages(history)
for i, prepend in enumerate(rethink_prepends):
if i > 0:
messages[-1]["content"] += "\n\n"
messages[-1]["content"] += prepend.format(question=question)
num_tokens = int(
max_num_tokens if ANSWER_MARKER not in prepend else final_num_tokens
)
t = threading.Thread(
target=pipe,
args=(messages,),
kwargs=dict(
max_new_tokens=num_tokens,
streamer=streamer,
do_sample=do_sample,
temperature=temperature,
),
)
t.start()
# μ λ΄μ©μΌλ‘ νμ€ν 리 μ¬κ΅¬μ±
history[-1].content += prepend.format(question=question)
if ANSWER_MARKER in prepend:
history[-1].metadata = {"title": "π μ¬κ³ κ³Όμ ", "status": "done"}
# μκ° μ’
λ£, μ΄μ λ΅λ³μ
λλ€ (μ€κ° λ¨κ³μ λν λ©νλ°μ΄ν° μμ)
history.append(gr.ChatMessage(role="assistant", content=""))
for token in streamer:
history[-1].content += token
history[-1].content = reformat_math(history[-1].content)
yield history
t.join()
yield history
with gr.Blocks(fill_height=True, title="λͺ¨λ LLM λͺ¨λΈμ μΆλ‘ λ₯λ ₯ λΆμ¬νκΈ°") as demo:
with gr.Row(scale=1):
with gr.Column(scale=2):
gr.Markdown("## Before (Original)")
chatbot_original = gr.Chatbot(
scale=1,
type="messages",
latex_delimiters=latex_delimiters,
label="Original Model (No Reasoning)"
)
with gr.Column(scale=2):
gr.Markdown("## After (Thinking)")
chatbot_thinking = gr.Chatbot(
scale=1,
type="messages",
latex_delimiters=latex_delimiters,
label="Model with Reasoning"
)
with gr.Row():
msg = gr.Textbox(
submit_btn=True,
label="",
show_label=False,
placeholder="μ¬κΈ°μ μ§λ¬Έμ μ
λ ₯νμΈμ.",
autofocus=True,
)
with gr.Row():
with gr.Column():
gr.Markdown("""## λ§€κ°λ³μ μ‘°μ """)
num_tokens = gr.Slider(
50,
4000,
2000,
step=1,
label="μΆλ‘ λ¨κ³λΉ μ΅λ ν ν° μ",
interactive=True,
)
final_num_tokens = gr.Slider(
50,
4000,
2000,
step=1,
label="μ΅μ’
λ΅λ³μ μ΅λ ν ν° μ",
interactive=True,
)
do_sample = gr.Checkbox(True, label="μνλ§ μ¬μ©")
temperature = gr.Slider(0.1, 1.0, 0.7, step=0.1, label="μ¨λ")
# μ¬μ©μκ° λ©μμ§λ₯Ό μ μΆνλ©΄ λ λ΄μ΄ λμμ μλ΅ν©λλ€
msg.submit(
user_input,
[msg, chatbot_original, chatbot_thinking], # μ
λ ₯
[msg, chatbot_original, chatbot_thinking], # μΆλ ₯
).then(
bot_original,
[
chatbot_original,
num_tokens,
do_sample,
temperature,
],
chatbot_original, # μΆλ ₯μμ μ νμ€ν 리 μ μ₯
).then(
bot_thinking,
[
chatbot_thinking,
num_tokens,
final_num_tokens,
do_sample,
temperature,
],
chatbot_thinking, # μΆλ ₯μμ μ νμ€ν 리 μ μ₯
)
if __name__ == "__main__":
demo.queue().launch() |