| import gradio as gr | |
| import os | |
| os.system('CMAKE_ARGS="-DLLAMA_OPENBLAS=on" FORCE_CMAKE=1 pip install llama-cpp-python') | |
| import wget | |
| from llama_cpp import Llama | |
| import random | |
| url = 'https://huggingface.co/TheBloke/WizardLM-7B-uncensored-GGML/resolve/main/WizardLM-7B-uncensored.ggmlv3.q2_K.bin' | |
| filename = wget.download(url) | |
| llm2 = Llama(model_path=filename, seed=random.randint(1, 2**31)) | |
| filename = wget.download(url) | |
| theme = gr.themes.Soft( | |
| primary_hue=gr.themes.Color("#ededed", "#fee2e2", "#fecaca", "#fca5a5", "#f87171", "#ef4444", "#dc2626", "#b91c1c", "#991b1b", "#7f1d1d", "#6c1e1e"), | |
| neutral_hue="red", | |
| ) | |
| title = """<h1 align="center">Chat with awesome WizardLM 7b model!</h1><br>""" | |
| with gr.Blocks(theme=theme) as demo: | |
| gr.HTML(title) | |
| gr.HTML("This model is awesome for its size! It is only 20th the size of Chatgpt but is around 90% as good as Chatgpt. However, please don't rely on WizardLM to provide 100% true information as it might be wrong sometimes. ") | |
| chatbot = gr.Chatbot() | |
| msg = gr.Textbox() | |
| clear = gr.ClearButton([msg, chatbot]) | |
| instruction = gr.Textbox(label="Instruction", placeholder=) | |
| temperature = gr.Slider(minimum=0.1, maximum=1.0, default=0.72, step=0.01, label="Temperature") | |
| top_p = gr.Slider(minimum=0.1, maximum=1.0, default=0.73, step=0.01, label="Top-p") | |
| top_k = gr.Slider(minimum=1, maximum=100, default=50, step=1, label="Top-k") | |
| repeat_penalty = gr.Slider(minimum=0.1, maximum=2.0, default=1.1, step=0.1, label="Repeat Penalty") | |
| def user(user_message, history): | |
| return gr.update(value="", interactive=True), history + [[user_message, None]] | |
| def bot(history): | |
| instruction = history[-1][1] or "" | |
| user_message = history[-1][0] | |
| token1 = llm.tokenize(b"### Instruction: ") | |
| token2 = llm.tokenize(instruction.encode()) | |
| token3 = llm2.tokenize(b"USER: ") | |
| tokens3 = llm2.tokenize(user_message.encode()) | |
| token4 = llm2.tokenize(b"\n\n### Response:") | |
| tokens = tokens3 + token4 | |
| history[-1][1] = "" | |
| count = 0 | |
| output = "" | |
| for token in llm2.generate(tokens, top_k=top_k, top_p=top_p, temp=temperature, repeat_penalty=repeat_penalty): | |
| text = llm2.detokenize([token]) | |
| output += text.decode() | |
| count += 1 | |
| if count >= 500 or (token == llm2.token_eos()): | |
| break | |
| history[-1][1] += text.decode() | |
| yield history | |
| response = msg.submit(user, [msg, chatbot], [msg, chatbot], queue=False).then( | |
| bot, chatbot, chatbot | |
| ) | |
| response.then(lambda: gr.update(interactive=True), None, [msg], queue=False) | |
| gr.HTML("Thanks for checking out this app!") | |
| demo.queue() | |
| demo.launch(debug=True) |