Improve CUDA/XLA setup and CPU fallback

cli/setup_python_venv

@@ -178,6 +178,26 @@ echo "   ===== Installing keras ====="
 # keras 3.13 has "issues" so we need to back down to 3.12.
 pip_install "keras==3.12.0"
 
+CUDA_DATA_DIR="${DATA_DIR}/cuda"
+LIBDEVICE_DIR="${CUDA_DATA_DIR}/nvvm/libdevice"
+mkdir -p "${LIBDEVICE_DIR}"
+TRITON_LIBDEVICE="$(
+    python - <<'PY'
+import glob
+import sys
+
+paths = glob.glob("**/site-packages/triton/backends/nvidia/lib/libdevice.10.bc", recursive=True)
+print(paths[0] if paths else "", end="")
+PY
+)"
+
+if [ -n "${TRITON_LIBDEVICE}" ] ; then
+    ln -sf "${TRITON_LIBDEVICE}" "${LIBDEVICE_DIR}/libdevice.10.bc"
+    echo "   Linked Triton libdevice.10.bc to ${LIBDEVICE_DIR}"
+else
+    echo "   ⚠️  Triton libdevice.10.bc not found; XLA may require --xla_gpu_cuda_data_dir"
+fi
+
 "${PROGDIR}/test_python" --data-dir="${DATA_DIR}"
 
 touch .mww-data-dir
@@ -185,4 +205,4 @@ END_TS=$EPOCHSECONDS
 
 echo "Run 'source ${VENV}/bin/activate' to activate the new virtualenv in the current shell."
 
-print_elapsed_time "${START_TS}" "${END_TS}" "Python package installation complete"
+print_elapsed_time "${START_TS}" "${END_TS}" "Python package installation complete"