main.py

from fastapi import FastAPI
from fastapi.responses import StreamingResponse
from pydantic import BaseModel
from huggingface_hub import InferenceClient
import uvicorn
import json  # Make sure to import json


app = FastAPI()

client = InferenceClient("mistralai/Mixtral-8x7B-Instruct-v0.1")

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

import json  # Import the JSON module

def generate(item: Item):
    temperature = float(item.temperature)
    if temperature < 1e-2:
        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)
    
    # Convert stream to a list to check if it's the last element
    responses = list(stream)
    for i, response in enumerate(responses):
        # Prepare the chunk as a JSON object
        chunk = {
            "text": response.token.text,
            "complete": i == len(responses) - 1  # True if this is the last chunk
        }
        # Yield the JSON-encoded string with a newline to separate chunks
        yield json.dumps(chunk).encode("utf-8") + b"\n"

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