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
import nltk
import os
from transformers import pipeline

# Set up the environment for NLTK
nltk.data.path.append(os.getenv('NLTK_DATA'))

# Initialize the FastAPI app
app = FastAPI()

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

# Initialize the summarization pipeline
summarizer = pipeline("summarization", model="sshleifer/distilbart-cnn-12-6")

class Item(BaseModel):
    prompt: str
    history: list
    system_prompt: str
    temperature: float = 0.8
    max_new_tokens: int = 12000
    top_p: float = 0.15
    repetition_penalty: float = 1.0

def summarize_history(history):
    # Concatenate all history entries into a single string
    full_history = " ".join(entry['content'] for entry in history if entry['role'] == 'user')
    # Summarize the history
    summarized_history = summarizer(full_history, max_length=1024, truncation=True)
    return summarized_history[0]['summary_text']

def format_prompt(current_prompt, history):
    formatted_history = "<s>"
    formatted_history += f"[HISTORY] {history} [/HISTORY]"
    formatted_history += f"[USER] {current_prompt} [/USER]</s>"
    return formatted_history

def generate_stream(item: Item) -> Generator[bytes, None, None]:
    summarized_history = summarize_history(item.history)
    formatted_prompt = format_prompt(item.prompt, summarized_history)
    input_token_count = len(nltk.word_tokenize(formatted_prompt))

    max_tokens_allowed = 32768
    max_new_tokens_adjusted = max(1, min(item.max_new_tokens, max_tokens_allowed - input_token_count))

    generate_kwargs = {
        "temperature": item.temperature,
        "max_new_tokens": max_new_tokens_adjusted,
        "top_p": item.top_p,
        "repetition_penalty": item.repetition_penalty,
        "do_sample": True,
        "seed": 42,
    }

    for response in client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True):
        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)