Spaces:
Runtime error
Runtime error
| 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 | |
| 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 | |
| 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() |