3

Dockerfile

@@ -1,6 +1,6 @@
 FROM python:3.9
 
-#RUN wget -q zephyr-7b-beta.Q4_K_S.gguf https://huggingface.co/TheBloke/zephyr-7B-beta-GGUF/resolve/main/zephyr-7b-beta.Q4_K_S.gguf
+RUN wget -q zephyr-7b-beta.Q4_K_S.gguf https://huggingface.co/TheBloke/zephyr-7B-beta-GGUF/resolve/main/zephyr-7b-beta.Q4_K_S.gguf
 
 COPY requirements.txt ./requirements.txt
 

main.py

@@ -1,5 +1,3 @@
-import os
-
 from ctransformers import AutoModelForCausalLM
 from fastapi import FastAPI
 from pydantic import BaseModel
@@ -17,16 +15,6 @@ class validation(BaseModel):
 #Fast API
 
 app = FastAPI()
-file_name = "zephyr-7b-beta.Q4_K_S.gguf"
-
-if not os.path.exists(file_name):
-    print("Downloading model...")
-    url = "https://huggingface.co/TheBloke/zephyr-7B-beta-GGUF/resolve/main/zephyr-7b-beta.Q4_K_S.gguf"
-    response = requests.get(url)
-    print(response.status_code)
-    with open(file_name, 'wb') as file:
-        file.write(response.content)
-
 
 @app.post("/llm_on_cpu")
 async def stream(item: validation):