gradio_app.py

# Importing libraries
from transformers import M2M100Tokenizer, M2M100ForConditionalGeneration
from llama_cpp import Llama
import gradio as gr
import psutil

# Initing things
llm = Llama(model_path="./model.bin")                              # LLaMa model
llama_model_name = "TheBloke/Llama-2-13B-chat-GGUF"                 
translator_tokenizer = M2M100Tokenizer.from_pretrained(            # tokenizer for translator
    "facebook/m2m100_1.2B", cache_dir="translator/"
)
translator_model = M2M100ForConditionalGeneration.from_pretrained( # translator model
    "facebook/m2m100_1.2B", cache_dir="translator/"
)
translator_model.eval()

# Preparing things to work
translator_tokenizer.src_lang = "en"
title = "llama.cpp API"
desc = '''<style>a:visited{color:black;}</style>
<h1>Hello, world!</h1>
This is showcase how to make own server with Llama2 model.<br>
I'm using here 7b model just for example. Also here's only CPU power.<br>
But you can use GPU power as well!<br>
<h1>How to GPU?</h1>
Change <code>`CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS`</code> in Dockerfile on <code>`CMAKE_ARGS="-DLLAMA_CUBLAS=on"`</code>. Also you can try <code>`DLLAMA_CLBLAST`</code>, <code>`DLLAMA_METAL`</code> or <code>`DLLAMA_METAL`</code>.<br>
Powered by <a href="https://github.com/abetlen/llama-cpp-python">llama-cpp-python</a>, <a href="https://quart.palletsprojects.com/">Quart</a> and <a href="https://www.uvicorn.org/">Uvicorn</a>.<br>
<h1>How to test it on own machine?</h1>
You can install Docker, build image and run it. I made <code>`run-docker.sh`</code> for ya. To stop container run <code>`docker ps`</code>, find name of container and run <code>`docker stop _dockerContainerName_`</code><br>
Or you can once follow steps in Dockerfile and try it on your machine, not in Docker.<br>
<br>''' + f"Memory used: {psutil.virtual_memory()[2]}<br>" + '''
<script>document.write("<b>URL of space:</b> "+window.location.href);</script>'''

# Loading prompt
with open('system.prompt', 'r', encoding='utf-8') as f:
    prompt = f.read()

def generate_answer(request: str, max_tokens: int = 256, language: str = "en", custom_prompt: str = None):
    try:
        maxTokens = max_tokens if 16 <= max_tokens <= 256 else 64
        if isinstance(custom_prompt, str):
            userPrompt = custom_prompt + "\n\nUser: " + request + "\nAssistant: "
        else:
            userPrompt = prompt + "\n\nUser: " + request + "\nAssistant: "
    except:
        return "Not enough data! Check that you passed all needed data."
    
    try:
        output = llm(userPrompt, max_tokens=maxTokens, stop=["User:", "\n"], echo=False)
        text = output["choices"][0]["text"]
        # i allowed only certain languages (its not discrimination, its just other popular language on my opinion!!!):
        # russian (ru), ukranian (uk), chinese (zh)
        if language in ["ru", "uk", "zh"]:
            encoded_input = translator_tokenizer(output, return_tensors="pt")
            generated_tokens = translator_model.generate(
                **encoded_input, forced_bos_token_id=translator_tokenizer.get_lang_id(language)
            )
            translated_text = translator_tokenizer.batch_decode(
                generated_tokens, skip_special_tokens=True
            )[0]
            return translated_text
        return text
    except Exception as e:
        print(e)
        return "Oops! Internal server error. Check the logs of space/instance."


demo = gr.Interface(
    fn=generate_answer,
    inputs=[
        gr.components.Textbox(label="Input"),
        gr.components.Number(value=256),
        gr.components.Dropdown(label="Target Language", value="en", choices=["en", "ru", "uk", "zh"]),
        gr.components.Textbox(label="Custom system prompt"),
    ],
    outputs=["text"],
    title=title,
    description=desc
)
demo.queue()
demo.launch()