nccl dep + default for cuda (#2526)

python/scripts/repair_cuda.sh

@@ -6,6 +6,7 @@ auditwheel repair dist/* \
   --exclude libnvrtc* \
   --exclude libcuda* \
   --exclude libcudnn* \
+  --exclude libnccl* \
   -w wheel_tmp
 
 
@@ -17,7 +18,7 @@ rm "${repaired_wheel}"
 mlx_so="mlx/lib/libmlx.so"
 rpath=$(patchelf --print-rpath "${mlx_so}")
 base="\$ORIGIN/../../nvidia"
-rpath=$rpath:${base}/cublas/lib:${base}/cuda_nvrtc/lib:${base}/cudnn/lib
+rpath=$rpath:${base}/cublas/lib:${base}/cuda_nvrtc/lib:${base}/cudnn/lib:${base}/nccl/lib
 patchelf --force-rpath --set-rpath "$rpath" "$mlx_so"
 python ../python/scripts/repair_record.py ${mlx_so}