from fastapi import FastAPI
from pydantic import BaseModel
import requests
from llama_cpp import Llama 
import threading
import gc

llms = {
  "TinyLLama 1b 4_K_M 2048": {
    "nctx": 2048, 
    "file": "tinyllama-1.1b-chat-v0.3.Q4_K_M.gguf", 
    "prefix": "### Human:", 
    "suffix": "### Assistant:"
  },
  "TinyLLama 1b OpenOrca 4_K_M 2048": {
    "nctx": 2048, 
    "file": "tinyllama-1.1b-1t-openorca.Q4_K_M.gguf", 
    "prefix": "<|im_start|>system You are a helpfull assistant<|im_end|><|im_start|>user", 
    "suffix": "<|im_end|><|im_start|>assistant"
  },
  "OpenLLama 3b 4_K_M 196k": {
    "nctx": 60000, 
    "file": "open-llama-3b-v2-wizard-evol-instuct-v2-196k.Q4_K_M.gguf", 
    "prefix": "### HUMAN:", 
    "suffix": "### RESPONSE:"
  },
  "Phi-2 2.7b 4_K_M 2048": {
    "nctx": 2048, 
    "file": "phi-2.Q4_K_M.gguf", 
    "prefix": "Instruct:", 
    "suffix": "Output:"
  },
  "Stable Zephyr 3b 4_K_M 4096": {
    "nctx": 4096, 
    "file": "stablelm-zephyr-3b.Q4_K_M.gguf", 
    "prefix": "<|user|>", 
    "suffix": "<|endoftext|><|assistant|>"
  }
}
    
    
#Fast API
app = FastAPI()
llm = None

def clean_memory():
    llm = None
    gc.collect()
    

@app.post("/llm_on_cpu")
async def stream(item: dict):

    model = llms[item['llm']]
    prefix=model['prefix']
    suffix=model['suffix']
    nctx =  item['nctx'] if 'nctx' in item.keys() else model['nctx']
    max_tokens = item['max_tokens'] if 'max_tokens' in item.keys() else 512
    user="""
    {prompt}"""
    
    llm = Llama(model_path="./code/"+model['file'], n_ctx=nctx, verbose=True,  n_threads=8)

    prompt = f"{prefix}{user.replace('{prompt}', item['prompt'])}{suffix}"
    result = llm(prompt, max_tokens=max_tokens)
    thread = threading.Thread(target=clean_memory)
    thread.start()
    return result