Update app.py

app.py

@@ -2,21 +2,20 @@ import spaces
 @spaces.GPU(duration=15)
 def gpu():
     print("[GPU] | GPU maintained.")
-    
+
 import os
 import sys
 import subprocess
+import urllib.request
 import gradio as gr
 
-def find_run_script():
-    for root, dirs, files in os.walk(os.getcwd()):
-        if "run_inference.py" in files:
-            return os.path.join(root, "run_inference.py")
-    return None
-
-SCRIPT_PATH = find_run_script()
-if not SCRIPT_PATH:
-    raise FileNotFoundError("run_inference.py not found in repo")
+# download run_inference.py at startup if it’s missing
+SCRIPT_PATH = os.path.join(os.getcwd(), "run_inference.py")
+if not os.path.isfile(SCRIPT_PATH):
+    urllib.request.urlretrieve(
+        "https://raw.githubusercontent.com/microsoft/BitNet/main/run_inference.py",
+        SCRIPT_PATH
+    )
 
 MODEL_PATH = os.environ.get("MODEL_PATH", "models/BitNet-b1.58-2B-4T/ggml-model-i2_s.gguf")
 
@@ -30,9 +29,7 @@ def generate(prompt, max_tokens=128, temperature=0.7):
         "-temp", str(temperature)
     ]
     proc = subprocess.run(cmd, capture_output=True, text=True)
-    if proc.returncode:
-        return proc.stderr.strip()
-    return proc.stdout.strip()
+    return proc.stdout.strip() if proc.returncode == 0 else proc.stderr.strip()
 
 iface = gr.Interface(
     fn=generate,
@@ -43,7 +40,7 @@ iface = gr.Interface(
     ],
     outputs=gr.Textbox(label="completion"),
     title="bitnet.cpp completion demo",
-    description="auto-detects run_inference.py so it won’t 404"
+    description="downloads inference script via python so no bash needed"
 )
 
 if __name__ == "__main__":