main.py

from ctransformers import AutoModelForCausalLM
from fastapi import FastAPI, Form
from pydantic import BaseModel

#Model loading
llm = AutoModelForCausalLM.from_pretrained("pt_merge_model_v3.Q4_K_M.gguf",
model_type='llama',
max_new_tokens = 512,
threads = 3,
)
   

#Pydantic object
class validation(BaseModel):
    prompt: str
    
#Fast API
app = FastAPI()

#Zephyr completion
@app.post("/llama_on_cpu")
async def stream(item: validation):
    system_prompt = 'Below is an instruction that describes a task. Write a response that appropriately completes the request.'
    prompt =  f"<s>[INST]<<SYS>>\n + {system_prompt} + <</SYS>>\n{item.prompt.strip()}[/INST]"
    return llm(prompt)