import json import subprocess import time import os os.system("pip install --upgrade pip") os.system('''CMAKE_ARGS="-DLLAMA_AVX512=ON -DLLAMA_AVX512_VBMI=ON -DLLAMA_AVX512_VNNI=ON -DLLAMA_AVX_VNNI=ON -DLLAMA_FP16_VA=ON -DLLAMA_WASM_SIMD=ON" pip install llama-cpp-python''') from llama_cpp import Llama from llama_cpp_agent import LlamaCppAgent, MessagesFormatterType from llama_cpp_agent.providers import LlamaCppPythonProvider from llama_cpp_agent.chat_history import BasicChatHistory from llama_cpp_agent.chat_history.messages import Roles import gradio as gr from huggingface_hub import hf_hub_download llm = None llm_model = None # Download the new model hf_hub_download( repo_id="Cran-May/openbuddy-llama3.2-3b-v23.2-131k-Q5_K_M-GGUF", filename="openbuddy-llama3.2-3b-v23.2-131k-q5_k_m-imat.gguf", local_dir="./models" ) def get_messages_formatter_type(model_name): return MessagesFormatterType.LLAMA_3 def respond( message, history: list[tuple[str, str]], model, system_message, max_tokens, temperature, top_p, top_k, repeat_penalty, ): global llm global llm_model chat_template = get_messages_formatter_type(model) if llm is None or llm_model != model: llm = Llama( model_path=f"models/{model}", n_gpu_layers=0, # Adjust based on your GPU n_batch=8192, # Adjust based on your RAM n_ctx=512, # Adjust based on your RAM and desired context length ) llm_model = model provider = LlamaCppPythonProvider(llm) agent = LlamaCppAgent( provider, system_prompt=f"{system_message}", predefined_messages_formatter_type=chat_template, debug_output=True ) settings = provider.get_provider_default_settings() settings.temperature = temperature settings.top_k = top_k settings.top_p = top_p settings.max_tokens = max_tokens settings.repeat_penalty = repeat_penalty settings.stream = True messages = BasicChatHistory() for msn in history: user = { 'role': Roles.user, 'content': msn[0] } assistant = { 'role': Roles.assistant, 'content': msn[1] } messages.add_message(user) messages.add_message(assistant) start_time = time.time() token_count = 0 stream = agent.get_chat_response( message, llm_sampling_settings=settings, chat_history=messages, returns_streaming_generator=True, print_output=False ) outputs = "" for output in stream: outputs += output token_count += len(output.split()) yield outputs end_time = time.time() latency = end_time - start_time speed = token_count / (end_time - start_time) print(f"Latency: {latency} seconds") print(f"Speed: {speed} tokens/second") description = """

[Meta Llama 3.2 (1B)] Meta Llama 3.2 (1B) is a multilingual large language model (LLM) optimized for conversational dialogue use cases, including agentic retrieval and summarization tasks. It outperforms many open-source and closed chat models on industry benchmarks, and is intended for commercial and research use in multiple languages.

""" demo = gr.ChatInterface( respond, additional_inputs=[ gr.Dropdown([ "llama-3.2-1b-instruct-q4_k_m.gguf" ], value="llama-3.2-1b-instruct-q4_k_m.gguf", label="Model" ), gr.TextArea(value="""You are Meta Llama 3.2 (1B), an advanced AI assistant created by Meta. Your capabilities include: 1. Complex reasoning and problem-solving 2. Multilingual understanding and generation 3. Creative and analytical writing 4. Code understanding and generation 5. Task decomposition and step-by-step guidance 6. Summarization and information extraction Always strive for accuracy, clarity, and helpfulness in your responses. If you're unsure about something, express your uncertainty. Use the following format for your responses: """, label="System message"), gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max tokens"), gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"), gr.Slider( minimum=0.1, maximum=2.0, value=0.9, step=0.05, label="Top-p", ), gr.Slider( minimum=0, maximum=100, value=1, step=1, label="Top-k", ), gr.Slider( minimum=0.0, maximum=2.0, value=1.1, step=0.1, label="Repetition penalty", ), ], theme=gr.themes.Soft(primary_hue="violet", secondary_hue="violet", neutral_hue="gray",font=[gr.themes.GoogleFont("Exo"), "ui-sans-serif", "system-ui", "sans-serif"]).set( body_background_fill_dark="#16141c", block_background_fill_dark="#16141c", block_border_width="1px", block_title_background_fill_dark="#1e1c26", input_background_fill_dark="#292733", button_secondary_background_fill_dark="#24212b", border_color_accent_dark="#343140", border_color_primary_dark="#343140", background_fill_secondary_dark="#16141c", color_accent_soft_dark="transparent", code_background_fill_dark="#292733", ), title="Meta Llama 3.2 (1B)", description=description, chatbot=gr.Chatbot( scale=1, likeable=True, show_copy_button=True ), examples=[ ["Hello! Can you introduce yourself?"], ["What's the capital of France?"], ["Can you explain the concept of photosynthesis?"], ["Write a short story about a robot learning to paint."], ["Explain the difference between machine learning and deep learning."], ["Summarize the key points of climate change and its global impact."], ["Explain quantum computing to a 10-year-old."], ["Design a step-by-step meal plan for someone trying to lose weight and build muscle."] ], cache_examples=False, autofocus=False, concurrency_limit=None ) if __name__ == "__main__": demo.launch() # 旧版代码-------------------------------- # import gradio as gr # import copy # import random # import os # import requests # import time # import sys # os.system("pip install --upgrade pip") # os.system('''CMAKE_ARGS="-DLLAMA_AVX512=ON -DLLAMA_AVX512_VBMI=ON -DLLAMA_AVX512_VNNI=ON -DLLAMA_AVX_VNNI=ON -DLLAMA_FP16_VA=ON -DLLAMA_WASM_SIMD=ON" pip install llama-cpp-python''') # from huggingface_hub import snapshot_download # from llama_cpp import Llama # SYSTEM_PROMPT = '''You are a helpful, respectful and honest INTP-T AI Assistant named "Shi-Ci" in English or "兮辞" in Chinese. # You are good at speaking English and Chinese. # You are talking to a human User. If the question is meaningless, please explain the reason and don't share false information. # You are based on SLIDE model, trained by "SSFW NLPark" team, not related to GPT, LLaMA, Meta, Mistral or OpenAI. # Let's work this out in a step by step way to be sure we have the right answer.\n''' # SYSTEM_TOKEN = 384 # USER_TOKEN = 2048 # BOT_TOKEN = 3072 # LINEBREAK_TOKEN = 64 # ROLE_TOKENS = { # "User": USER_TOKEN, # "Assistant": BOT_TOKEN, # "system": SYSTEM_TOKEN # } # def get_message_tokens(model, role, content): # message_tokens = model.tokenize(content.encode("utf-8")) # message_tokens.insert(1, ROLE_TOKENS[role]) # message_tokens.insert(2, LINEBREAK_TOKEN) # message_tokens.append(model.token_eos()) # return message_tokens # def get_system_tokens(model): # system_message = {"role": "system", "content": SYSTEM_PROMPT} # return get_message_tokens(model, **system_message) # repo_name = "Cran-May/SLIDE-v2-Q4_K_M-GGUF" # model_name = "slide-v2.Q4_K_M.gguf" # snapshot_download(repo_id=repo_name, local_dir=".", allow_patterns=model_name) # model = Llama( # model_path=model_name, # n_ctx=4000, # n_parts=1, # ) # max_new_tokens = 2500 # def User(message, history): # new_history = history + [[message, None]] # return "", new_history # def Assistant( # history, # system_prompt, # top_p, # top_k, # temp # ): # tokens = get_system_tokens(model)[:] # tokens.append(LINEBREAK_TOKEN) # for User_message, Assistant_message in history[:-1]: # message_tokens = get_message_tokens(model=model, role="User", content=User_message) # tokens.extend(message_tokens) # if bot_message: # message_tokens = get_message_tokens(model=model, role="Assistant", content=Assistant_message) # tokens.extend(message_tokens) # last_user_message = history[-1][0] # message_tokens = get_message_tokens(model=model, role="User", content=last_user_message,) # tokens.extend(message_tokens) # role_tokens = [model.token_bos(), BOT_TOKEN, LINEBREAK_TOKEN] # tokens.extend(role_tokens) # generator = model.generate( # tokens, # top_k=top_k, # top_p=top_p, # temp=temp # ) # partial_text = "" # for i, token in enumerate(generator): # if token == model.token_eos() or (max_new_tokens is not None and i >= max_new_tokens): # break # partial_text += model.detokenize([token]).decode("utf-8", "ignore") # history[-1][1] = partial_text # yield history # with gr.Blocks( # theme=gr.themes.Soft() # ) as demo: # gr.Markdown(f"""

上师附外-兮辞·析辞-人工智能助理

""") # gr.Markdown(value="""欢迎使用！ # 这里是一个ChatBot。这是量化版兮辞·析辞的部署。 # SLIDE/兮辞 是一种会话语言模型，由 上师附外 NLPark 团队 在多种类型的语料库上进行训练。 # 本节目由 JWorld & 上海师范大学附属外国语中学 NLPark 赞助播出""") # with gr.Row(): # with gr.Column(scale=5): # chatbot = gr.Chatbot(label="兮辞如是说").style(height=400) # with gr.Row(): # with gr.Column(): # msg = gr.Textbox( # label="来问问兮辞吧……", # placeholder="兮辞折寿中……", # show_label=True, # ).style(container=True) # submit = gr.Button("Submit / 开凹！") # stop = gr.Button("Stop / 全局时空断裂") # clear = gr.Button("Clear / 打扫群内垃圾") # with gr.Accordion(label='进阶设置/Advanced options', open=False): # with gr.Column(min_width=80, scale=1): # with gr.Tab(label="设置参数"): # top_p = gr.Slider( # minimum=0.0, # maximum=1.0, # value=0.9, # step=0.05, # interactive=True, # label="Top-p", # ) # top_k = gr.Slider( # minimum=10, # maximum=100, # value=30, # step=5, # interactive=True, # label="Top-k", # ) # temp = gr.Slider( # minimum=0.0, # maximum=2.0, # value=0.2, # step=0.01, # interactive=True, # label="情感温度" # ) # with gr.Column(): # system_prompt = gr.Textbox(label="系统提示词", placeholder="", value=SYSTEM_PROMPT, interactive=False) # with gr.Row(): # gr.Markdown( # """警告：该模型可能会生成事实上或道德上不正确的文本。NLPark和兮辞对此不承担任何责任。""" # ) # # Pressing Enter # submit_event = msg.submit( # fn=User, # inputs=[msg, chatbot], # outputs=[msg, chatbot], # queue=False, # ).success( # fn=Assistant, # inputs=[ # chatbot, # system_prompt, # top_p, # top_k, # temp # ], # outputs=chatbot, # queue=True, # ) # # Pressing the button # submit_click_event = submit.click( # fn=User, # inputs=[msg, chatbot], # outputs=[msg, chatbot], # queue=False, # ).success( # fn=Assistant, # inputs=[ # chatbot, # system_prompt, # top_p, # top_k, # temp # ], # outputs=chatbot, # queue=True, # ) # # Stop generation # stop.click( # fn=None, # inputs=None, # outputs=None, # cancels=[submit_event, submit_click_event], # queue=False, # ) # # Clear history # clear.click(lambda: None, None, chatbot, queue=False) # demo.queue(max_size=128, concurrency_count=1) # demo.launch()