Update app.py

app.py

@@ -14,6 +14,14 @@ from vcoder_llava.utils import (build_logger, server_error_msg,
                           violates_moderation, moderation_msg)
 from chat import Chat
 
+import subprocess
+# Install flash attention, skipping CUDA build if necessary
+subprocess.run(
+    "pip install flash-attn --no-build-isolation",
+    env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
+    shell=True,
+)
+
 
 logger = build_logger("gradio_app", "gradio_web_server.log")