main.py

from fastapi import FastAPI
from fastapi.responses import StreamingResponse
from pydantic import BaseModel
from huggingface_hub import InferenceClient
import uvicorn
from typing import Generator
import json  # Asegúrate de que esta línea esté al principio del archivo

app = FastAPI()

# Initialize the InferenceClient with your model
client = InferenceClient("mistralai/Mistral-7B-Instruct-v0.2")

class Item(BaseModel):
    prompt: str
    history: list
    system_prompt: str
    temperature: float = 0.8
    max_new_tokens: int = 9000
    top_p: float = 0.15
    repetition_penalty: float = 1.0

def format_prompt(current_prompt, history):
    formatted_history = "<s>"
    for entry in history:
        if entry["role"] == "user":
            formatted_history += f"[USER] {entry['content']} [/USER]"
        elif entry["role"] == "assistant":
            formatted_history += f"[ASSISTANT] {entry['content']} [/ASSISTANT]"
    formatted_history += f"[USER] {current_prompt} [/USER]</s>"
    return formatted_history


def generate_stream(item: Item) -> Generator[bytes, None, None]:
    formatted_prompt = format_prompt(f"{item.system_prompt}, {item.prompt}", item.history)
    generate_kwargs = {
        "temperature": item.temperature,
        "max_new_tokens": item.max_new_tokens,
        "top_p": item.top_p,
        "repetition_penalty": item.repetition_penalty,
        "do_sample": True,
        "seed": 42,  # Adjust or omit the seed as needed
    }

    # Stream the response from the InferenceClient
    for response in client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True):
        # This assumes 'details=True' gives you a structure where you can access the text like this
        chunk = {
            "text": response.token.text,
            "complete": response.generated_text is not None  # Adjust based on how you detect completion
        }
        yield json.dumps(chunk).encode("utf-8") + b"\n"

@app.post("/generate/")
async def generate_text(item: Item):
    # Stream response back to the client
    return StreamingResponse(generate_stream(item), media_type="application/x-ndjson")

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=8000)