Update server.py

server.py

@@ -1,9 +1,17 @@
 from fastapi import FastAPI
 import subprocess
+import os
 
 app = FastAPI()
 
-MODEL_PATH = "deepseek-coder-v2-lite.Q4_K_M.gguf"
+MODEL_URL = "https://huggingface.co/TheBloke/deepseek-coder-v2-lite-GGUF/resolve/main/deepseek-coder-v2-lite.Q4_K_M.gguf"
+MODEL_PATH = "deepseek-coder.gguf"
+
+def download_model():
+    if not os.path.exists(MODEL_PATH):
+        subprocess.run(["wget", MODEL_URL, "-O", MODEL_PATH])
+
+download_model()
 
 @app.get("/")
 def read_root():
@@ -13,12 +21,12 @@ def read_root():
 def generate(prompt: str):
 
     cmd = [
-    "/app/llama.cpp/build/bin/llama-cli",
-    "-m", MODEL_PATH,
-    "-p", prompt,
-    "-n", "300",
-    "--ctx-size", "8192"
-]
+        "/app/llama.cpp/build/bin/llama-cli",
+        "-m", MODEL_PATH,
+        "-p", prompt,
+        "-n", "200",
+        "--ctx-size", "8192"
+    ]
 
     result = subprocess.run(cmd, capture_output=True, text=True)