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
import torch


app = FastAPI()

# Initialize the InferenceClient with your model
client = InferenceClient("meta-llama/Llama-2-7b-chat")

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(message, history):
    # Simple structure: alternating lines of dialogue, no special tokens unless specified by the model documentation
    conversation = ""
    for user_prompt, bot_response in history:
        conversation += f"User: {user_prompt}\nBot: {bot_response}\n"
    conversation += f"User: {message}"
    return conversation



# No changes needed in the format_prompt function unless the new model requires different prompt formatting

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):
        # Check if the 'details' flag and response structure are the same for the new model
        chunk = {
            "text": response.token.text,
            "complete": response.generated_text is not None
        }
        yield json.dumps(chunk).encode("utf-8") + b"\n"

@app.post("/generate/")
async def generate_text(item: Item):
    return StreamingResponse(generate_stream(item), media_type="application/x-ndjson")

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