diff --git a/examples/basics.rs b/examples/basics.rs
index 8c0b1fcc..255a646d 100644
--- a/examples/basics.rs
+++ b/examples/basics.rs
@@ -13,7 +13,6 @@ fn grad_example() {
 }
 
 fn main() {
-    tch::maybe_init_cuda();
     let t = Tensor::of_slice(&[3, 1, 4, 1, 5]);
     t.print();
     let t = Tensor::randn(&[5, 4], kind::FLOAT_CPU);
diff --git a/src/lib.rs b/src/lib.rs
index 1d5d81fc..a452ce25 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -28,9 +28,3 @@ pub use tensor::{
 
 pub mod nn;
 pub mod vision;
-
-pub fn maybe_init_cuda() {
-    unsafe {
-        torch_sys::dummy_cuda_dependency();
-    }
-}
diff --git a/src/tensor/mod.rs b/src/tensor/mod.rs
index 44f4459f..5e707ace 100644
--- a/src/tensor/mod.rs
+++ b/src/tensor/mod.rs
@@ -3,7 +3,6 @@ use crate::{Device, Kind, Scalar, TchError};
 use half::f16;
 use std::convert::{TryFrom, TryInto};
 use std::ops::{Add, AddAssign, Div, DivAssign, Mul, MulAssign, Neg, Sub, SubAssign};
-use torch_sys::*;
 
 pub mod index;
 mod iter;
@@ -750,6 +749,3 @@ try_from_impl!(i32);
 try_from_impl!(f64);
 try_from_impl!(i64);
 try_from_impl!(bool);
-
-#[used]
-static INIT_ARRAY: [unsafe extern "C" fn(); 1] = [dummy_cuda_dependency];
diff --git a/torch-sys/build.rs b/torch-sys/build.rs
index 715eb5fd..d9981708 100644
--- a/torch-sys/build.rs
+++ b/torch-sys/build.rs
@@ -133,14 +133,8 @@ fn prepare_libtorch_dir() -> PathBuf {
     }
 }
 
-fn make<P: AsRef<Path>>(libtorch: P, use_cuda: bool, use_hip: bool) {
+fn make<P: AsRef<Path>>(libtorch: P) {
     let os = env::var("CARGO_CFG_TARGET_OS").expect("Unable to get TARGET_OS");
-
-    let cuda_dependency = if use_cuda || use_hip {
-        "libtch/dummy_cuda_dependency.cpp"
-    } else {
-        "libtch/fake_cuda_dependency.cpp"
-    };
     println!("cargo:rerun-if-changed=libtch/torch_api.cpp");
     println!("cargo:rerun-if-changed=libtch/torch_api.h");
     println!("cargo:rerun-if-changed=libtch/torch_api_generated.cpp.h");
@@ -165,7 +159,6 @@ fn make<P: AsRef<Path>>(libtorch: P, use_cuda: bool, use_hip: bool) {
                 .flag("-std=c++14")
                 .flag(&format!("-D_GLIBCXX_USE_CXX11_ABI={}", libtorch_cxx11_abi))
                 .file("libtch/torch_api.cpp")
-                .file(cuda_dependency)
                 .compile("tch");
         }
         "windows" => {
@@ -179,7 +172,6 @@ fn make<P: AsRef<Path>>(libtorch: P, use_cuda: bool, use_hip: bool) {
                 .include(libtorch.as_ref().join("include"))
                 .include(libtorch.as_ref().join("include/torch/csrc/api/include"))
                 .file("libtch/torch_api.cpp")
-                .file(cuda_dependency)
                 .compile("tch");
         }
         _ => panic!("Unsupported OS"),
@@ -189,20 +181,6 @@ fn make<P: AsRef<Path>>(libtorch: P, use_cuda: bool, use_hip: bool) {
 fn main() {
     if !cfg!(feature = "doc-only") {
         let libtorch = prepare_libtorch_dir();
-        // use_cuda is a hacky way to detect whether cuda is available and
-        // if it's the case link to it by explicitly depending on a symbol
-        // from the torch_cuda library.
-        // It would be better to use -Wl,--no-as-needed but there is no way
-        // to specify arbitrary linker flags at the moment.
-        //
-        // Once https://github.com/rust-lang/cargo/pull/8441 is available
-        // we should switch to using rustc-link-arg instead e.g. with the
-        // following flags:
-        //   -Wl,--no-as-needed -Wl,--copy-dt-needed-entries -ltorch
-        //
-        // This will be available starting from cargo 1.50 but will be a nightly
-        // only option to start with.
-        // https://github.com/rust-lang/cargo/blob/master/CHANGELOG.md
         let use_cuda = libtorch.join("lib").join("libtorch_cuda.so").exists()
             || libtorch.join("lib").join("torch_cuda.dll").exists();
         let use_cuda_cu = libtorch.join("lib").join("libtorch_cuda_cu.so").exists()
@@ -216,8 +194,11 @@ fn main() {
             libtorch.join("lib").display()
         );
 
-        make(&libtorch, use_cuda, use_hip);
+        make(&libtorch);
 
+        // PyTorch uses two separate library for the cpu bits and the cuda bits.
+        // The following flags ensure that the linker does not remove the
+        // cuda dependencies.
         println!("cargo:rustc-link-lib=static=tch");
         if use_cuda {
             println!("cargo:rustc-link-lib=torch_cuda");
@@ -231,7 +212,9 @@ fn main() {
         if use_hip {
             println!("cargo:rustc-link-lib=torch_hip");
         }
-        println!("cargo:rustc-link-lib=torch");
+        println!("cargo:rustc-link-arg=-Wl,--no-as-needed");
+        println!("cargo:rustc-link-arg=-Wl,--copy-dt-needed-entries");
+        println!("cargo:rustc-link-arg=-ltorch");
         println!("cargo:rustc-link-lib=torch_cpu");
         println!("cargo:rustc-link-lib=c10");
         if use_hip {
diff --git a/torch-sys/libtch/dummy_cuda_dependency.cpp b/torch-sys/libtch/dummy_cuda_dependency.cpp
deleted file mode 100644
index f2495a82..00000000
--- a/torch-sys/libtch/dummy_cuda_dependency.cpp
+++ /dev/null
@@ -1,18 +0,0 @@
-#include<stdio.h>
-#include<stdint.h>
-using namespace std;
-extern "C" {
-    void dummy_cuda_dependency();
-}
-
-namespace at {
-    namespace cuda {
-        void *getCurrentCUDABlasHandle();
-        int warp_size();
-    }
-}
-char * magma_strerror(int err);
-void dummy_cuda_dependency() {
-    at::cuda::getCurrentCUDABlasHandle();
-    at::cuda::warp_size();
-}
diff --git a/torch-sys/libtch/fake_cuda_dependency.cpp b/torch-sys/libtch/fake_cuda_dependency.cpp
deleted file mode 100644
index 9db48039..00000000
--- a/torch-sys/libtch/fake_cuda_dependency.cpp
+++ /dev/null
@@ -1,6 +0,0 @@
-extern "C" {
-    void dummy_cuda_dependency();
-}
-
-void dummy_cuda_dependency() {
-}
diff --git a/torch-sys/src/lib.rs b/torch-sys/src/lib.rs
index 32cf5049..e6b692c1 100644
--- a/torch-sys/src/lib.rs
+++ b/torch-sys/src/lib.rs
@@ -254,7 +254,3 @@ extern "C" {
         noutputs: c_int,
     );
 }
-
-extern "C" {
-    pub fn dummy_cuda_dependency();
-}