add flash attention

requirements.txt

@@ -11,7 +11,10 @@ langchain-chroma
 unstructured[all-docs]
 libmagic
 gradio
-flash-attn
+torch==2.2.0+cu121   # Ensure you're using the right PyTorch version with CUDA support
+flash-attn==2.6.3    # Flash attention module
+numpy<2              # Downgrade to avoid NumPy 2.0.1 conflicts
+pybind11>=2.12       # Ensure compatibility for modules needing pybind11
 # poppler
 # tesseract
 # libxml2