import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from fastapi import FastAPI
from pydantic import BaseModel

# Load the model and tokenizer
model_name = "databricks/dolly-v2-3b"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16, device_map="auto")

# FastAPI app
app = FastAPI()

# Input format
class ChatInput(BaseModel):
    user_input: str

@app.post("/chat")
async def chat(chat_input: ChatInput):
    inputs = tokenizer(chat_input.user_input, return_tensors="pt").to("cuda")
    outputs = model.generate(**inputs, max_length=200, do_sample=True)
    response_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return {"response": response_text}