Install flash-attn at runtime for HF Spaces

- Remove flash-attn from requirements.txt to avoid build issues
- Install flash-attn at startup only when on HF Spaces (SPACE_ID env var)
- Use --no-build-isolation flag to use pre-built wheels
- Made optional: continues without flash-attn if installation fails
- Installs after torch is available, avoiding build dependency issues

app.py

@@ -21,6 +21,18 @@ from mosaic.gradio_app import download_and_process_models
 from mosaic.ui import launch_gradio
 
 if __name__ == "__main__":
+    # Install flash-attn for H100 acceleration (if on Spaces)
+    if os.environ.get("SPACE_ID"):
+        try:
+            subprocess.run(
+                "pip install flash-attn --no-build-isolation",
+                shell=True,
+                check=False,
+                capture_output=True,
+            )
+        except Exception:
+            pass  # Flash-attn is optional, continue without it
+    
     # Install Paladin from GitHub
     GIT_TOKEN = os.environ.get("GH_TOKEN")
     if GIT_TOKEN:

requirements.txt

@@ -1,7 +1,6 @@
 --extra-index-url https://download.pytorch.org/whl/cu121
 torch>=2.0.0,<2.6
 torchvision>=0.15.0
-flash-attn>=2.5.0
 open-clip-torch
 gradio>=5.49.0
 loguru>=0.7.3