app.py

# Importing libraries
from llama_cpp import Llama
from time import time
import gradio as gr
import psutil

# Initing things                
print("debug: init model")
llm = Llama(model_path="./model.bin")                             # LLaMa model
llama_model_name = "TheBloke/dolphin-2.2.1-AshhLimaRP-Mistral-7B-GGUF" 
print("! INITING DONE !")

# Preparing things to work
title = "llama.cpp API"
desc = '''<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><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> or <code>`DLLAMA_METAL`</code>.<br><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>" + '''
Powered by <a href="https://github.com/abetlen/llama-cpp-python">llama-cpp-python</a> and <a href="https://www.gradio.app/">Gradio</a>.<br><br>'''

# Loading prompt
prompt = ""
system_message = ""

def generate_answer(request: str, max_tokens: int = 256, custom_prompt: str = None):
    t0 = time()
    logs = f"Request: {request}\nMax tokens: {max_tokens}\nCustom prompt: {custom_prompt}\n"
    try:
        maxTokens = max_tokens if 16 <= max_tokens <= 256 else 64
        userPrompt = prompt.replace("{prompt}", request)
        userPrompt = userPrompt.replace(
            "{system_message}",
            custom_prompt if isinstance(custom_prompt, str) and len(custom_prompt.strip()) > 1 and custom_prompt.strip() not in ['', None, ' '] else system_message
        )
        logs += f"\nFinal prompt: {userPrompt}\n"
    except:
        return "Not enough data! Check that you passed all needed data.", logs
    
    try:
        # this shitty fix will be until i willnt figure out why sometimes there is empty output
        counter = 1
        while counter <= 3:
            logs += f"Attempt {counter} to generate answer...\n"
            output = llm(userPrompt, max_tokens=maxTokens, stop=["<|im_end|>"], echo=False)
            text = output["choices"][0]["text"]
            if len(text.strip()) > 1 and text.strip() not in ['', None, ' ']:
                break
            counter += 1
        logs += f"Final attempt: {counter}\n"
        if len(text.strip()) <= 1 or text.strip() in ['', None, ' ']:
            logs += f"Generated and aborted: {text}"
            text = "Sorry, but something went wrong while generating answer. Try again or fix code. If you are maintainer of this space, look into logs."
        
        logs += f"\nFinal: '''{text}'''"
        logs += f"\n\nTime spent: {time()-t0}"
        return text, logs
    except Exception as e:
        logs += str(e)
        logs += f"\n\nTime spent: {time()-t0}"
        return "Oops! Internal server error. Check the logs of space/instance.", logs

print("! LOAD GRADIO INTERFACE !")
demo = gr.Interface(
    fn=generate_answer,
    inputs=[
        gr.components.Textbox(label="Input"),
        gr.components.Number(value=256),
        gr.components.Textbox(label="Custom system prompt"),
    ],
    outputs=[
        gr.components.Textbox(label="Output"),
        gr.components.Textbox(label="Logs")
    ],
    title=title,
    description=desc,
    allow_flagging='never'
)
demo.queue()
print("! LAUNCHING GRADIO !")
demo.launch(server_name="0.0.0.0")