install flash attn on runtime

app.py

@@ -7,6 +7,20 @@ import tempfile
 import spaces
 import gradio as gr
 
+import subprocess
+import sys
+
+def install_flash_attn_wheel():
+    flash_attn_wheel_url = "https://github.com/Dao-AILab/flash-attention/releases/download/v2.6.3/flash_attn-2.6.3+cu123torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl"
+    try:
+        # Call pip to install the wheel file
+        subprocess.check_call([sys.executable, "-m", "pip", "install", flash_attn_wheel_url])
+        print("Wheel installed successfully!")
+    except subprocess.CalledProcessError as e:
+        print(f"Failed to install the flash attnetion wheel. Error: {e}")
+
+install_flash_attn_wheel()
+
 import cv2
 try:
     from mmengine.visualization import Visualizer