main.py

from fastapi import FastAPI
from pydantic import BaseModel
from huggingface_hub import InferenceClient
import uvicorn
import asyncio

app = FastAPI()

client = InferenceClient(model="Qwen/Qwen2.5-7B-Instruct-GGUF")

class Item(BaseModel):
    prompt: str
    history: list
    system_prompt: str
    temperature: float = 0.0
    max_new_tokens: int = 1048
    top_p: float = 0.15
    repetition_penalty: float = 1.0

def format_prompt(message, history):
    prompt = "<s>"
    for user_prompt, bot_response in history:
        prompt += f"[INST] {user_prompt} [/INST]"
        prompt += f" {bot_response}</s> "
    prompt += f"[INST] {message} [/INST]"
    return prompt

def generate(item: Item):
    temperature = max(float(item.temperature), 1e-2)
    top_p = float(item.top_p)

    generate_kwargs = dict(
        temperature=temperature,
        max_new_tokens=item.max_new_tokens,
        top_p=top_p,
        repetition_penalty=item.repetition_penalty,
        do_sample=True,
        seed=42,
    )

    formatted_prompt = format_prompt(f"{item.system_prompt}, {item.prompt}", item.history)
    stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)
    output = ""

    for response in stream:
        output += response.token.text
    return output

@app.post("/generate/")
async def generate_text(item: Item):
    return {"response": generate(item)}

@app.on_event("startup")
async def preload_model():
    # Check if the model is already loaded
    status = client.get_model_status()
    if not status.loaded:
        # Trigger model loading by making a dummy request
        dummy_prompt = "This is a dummy prompt to load the model."
        client.text_generation(dummy_prompt, max_new_tokens=1)
        # Optionally, wait until the model is loaded
        while not client.get_model_status().loaded:
            await asyncio.sleep(5)  # Wait for 5 seconds before checking again