Update prebuilt.sh

prebuilt.sh

@@ -1,13 +1,23 @@
 #!/bin/bash
 set -e
 
-echo "=== Building llama-cpp-python with CUDA for ZeroGPU ==="
-export CMAKE_ARGS="-DGGML_CUDA=on"
+echo "==============================================="
+echo " Building llama-cpp-python with CUDA for ZeroGPU"
+echo "==============================================="
+
+# Your Space has CUDA 13.0.2 toolkit (visible in build logs).
+# Build llama-cpp-python from source linked against it.
+export CMAKE_ARGS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=native"
 export FORCE_CMAKE=1
+export GGML_CUDA_FORCE_MMQ=on       # force CUDA for all layers
+export GGML_CUDA_FULL_TENSORS=on    # use CUDA tensors throughout
+
+echo "CMAKE_ARGS=$CMAKE_ARGS"
+echo "FORCE_CMAKE=$FORCE_CMAKE"
 
-# The Gradio base image has CUDA 12.x installed.
-# Build llama-cpp-python from source so it links against the actual GPU toolkit.
-pip install --no-cache-dir llama-cpp-python
-pip install --no-cache-dir python-docx pdfplumber chardet
+pip install --no-cache-dir --no-build-isolation llama-cpp-python
 
-echo "=== Build complete ==="
+echo ""
+echo "==============================================="
+echo " llama-cpp-python CUDA build complete!"
+echo "==============================================="