app.py

from llama_cpp import Llama
import gradio as gr

import os
import json

GRADIO_SHOW_API_INFO = os.getenv("AGS_SHOW_API_INFO") or True

AGS_REPO = os.getenv("AGS_REPO") or "lmstudio-community/gemma-1.1-2b-it-GGUF"
AGS_FILENAME = os.getenv("AGS_FILENAME") or "gemma-1.1-2b-it-Q4_K_M.gguf"
AGS_LLAMA_CONFIG = {
    "prompt_format": "raw"
}
AGS_TITLE = os.getenv("AGS_TITLE") or "API GGUF Space"

try:
    AGS_LLAMA_CONFIG = json.loads(os.getenv("AGS_LLAMA_CONFIG"))
except Exception as e:
    if AGS_LLAMA_CONFIG and AGS_LLAMA_CONFIG is not None:
        print("Invalid Llama config. Config must be valid JSON. Got:\n", AGS_LLAMA_CONFIG)
    ARGS_LLAMA_CONFIG = {
        "n_gpu_layers": 0,
    }

def main():
    llm = Llama.from_pretrained(AGS_REPO, filename=AGS_FILENAME, **ARGS_LLAMA_CONFIG)
    def api_chat(inpt, settings):
        res = None
        try:
            inpt = json.loads(inpt)
            settings = json.loads(settings)
            print("Request:\n" + json.dumps(inpt, indent=2) + "\n\n" + ("*_"*24))
        except Exception as e:
            res = llm(inpt, **settings)
        if "@execute" in inpt and inpt['@execute']:
            inpt.pop("@execute", None)
            res = llm(json.dumps(inpt), **settings)
        if "messages" in inpt and inpt['messages']:
            res = llm.create_chat_completion(messages=inpt, **settings)
        if "prompt" in settings and settings["prompt"]:
            res = llm(inpt, **settings)
        if res is None:
            res = llm(json.dumps(inpt), **settings)
        if settings and "full_output" in settings and settings["full_output"]:
            return res
        if "content" in res['choices'][0]:
            return res['choices'][0]['content']
        if "text" in res['choices'][0]:
            return res['choices'][0]['text']
        if "message" in res['choices'][0] and "content" in res['choices'][0]['message']:
            return res['choices'][0]['message']['content']
        return res
    def chat(inpt):
        if not inpt:
            return ""
        return llm.create_chat_completion(messages=[{
            "role": "user",
            "content": inpt
        }])['choices'][0]['message']['content']
    with gr.Interface(fn=chat, inputs=[inpt:=gr.Textbox()], outputs="text") as interface:
        with gr.Row(visible=False):
            shadow_input = gr.Textbox(visible=False)
            shadow_input.submit(api_chat, inputs=[inpt, shadow_input], api_name="api_chat", server_name=AGS_TITLE)
        interface.launch(debug=True, show_api=GRADIO_SHOW_API_INFO)

if __name__ == "__main__":
    main()