added install flash attn

app.py

@@ -3,6 +3,48 @@ import torch
 from transformers import AutoProcessor, MusicgenForConditionalGeneration
 import scipy.io.wavfile
 import numpy as np
+import subprocess
+import sys
+import os
+
+def setup_flash_attention():
+    """One-time setup for flash-attention with special flags"""
+    # Check if flash-attn is already installed
+    try:
+        import flash_attn
+        print("flash-attn already installed")
+        return
+    except ImportError:
+        pass
+    
+    # Check if we've already tried to install it in this session
+    if os.path.exists("/tmp/flash_attn_installed"):
+        return
+        
+    try:
+        print("Installing flash-attn with --no-build-isolation...")
+        subprocess.run([
+            sys.executable, "-m", "pip", "install", 
+            "flash-attn==2.7.3", "--no-build-isolation"
+        ], check=True)
+        
+        # Uninstall apex if it exists
+        subprocess.run([
+            sys.executable, "-m", "pip", "uninstall", "apex", "-y"
+        ], check=False)  # Don't fail if apex isn't installed
+        
+        # Mark as installed
+        with open("/tmp/flash_attn_installed", "w") as f:
+            f.write("installed")
+            
+        print("flash-attn installation completed")
+        
+    except subprocess.CalledProcessError as e:
+        print(f"Warning: Failed to install flash-attn: {e}")
+        # Continue anyway - the model might work without it
+
+# Run setup once when the module is imported
+setup_flash_attention()
 
 # Load model and processor
 @gr.cache()

setup.sh

@@ -0,0 +1,8 @@
+#!/bin/bash
+set -e
+
+# Install flash-attn with no-build-isolation
+pip install flash-attn==2.7.3 --no-build-isolation
+
+# Uninstall apex if it exists (as in your original script)
+pip uninstall apex -y || true