Merge pull request #642 from LaurentMazare/remove-cuda-hack

Another attempt at removing the cuda hack.

examples/stable-diffusion/main.rs

@@ -2448,7 +2448,6 @@ impl DDIMScheduler {
 }
 
 fn main() -> anyhow::Result<()> {
-    tch::maybe_init_cuda();
     println!("Cuda available: {}", tch::Cuda::is_available());
     println!("Cudnn available: {}", tch::Cuda::cudnn_is_available());
     // TODO: Switch to using claps to allow more flags?

src/lib.rs

@@ -30,9 +30,3 @@ pub use tensor::{
 
 pub mod nn;
 pub mod vision;
-
-pub fn maybe_init_cuda() {
-    unsafe {
-        torch_sys::dummy_cuda_dependency();
-    }
-}

src/tensor/mod.rs

@@ -1,6 +1,5 @@
 //! A Torch tensor.
 use crate::{Device, Kind, TchError};
-use torch_sys::*;
 
 mod convert;
 pub mod display;
@@ -252,6 +251,3 @@ impl Tensor {
         self.g_to_mkldnn(self.kind())
     }
 }
-
-#[used]
-static INIT_ARRAY: [unsafe extern "C" fn(); 1] = [dummy_cuda_dependency];

torch-sys/build.rs

@@ -349,13 +349,7 @@ impl SystemInfo {
         }
     }
 
-    fn make(&self, use_cuda: bool, use_hip: bool) {
-        let cuda_dependency = if use_cuda || use_hip {
-            "libtch/dummy_cuda_dependency.cpp"
-        } else {
-            "libtch/fake_cuda_dependency.cpp"
-        };
-        println!("cargo:rerun-if-changed={}", cuda_dependency);
+    fn make(&self) {
         println!("cargo:rerun-if-changed=libtch/torch_python.cpp");
         println!("cargo:rerun-if-changed=libtch/torch_python.h");
         println!("cargo:rerun-if-changed=libtch/torch_api_generated.cpp");
@@ -365,8 +359,7 @@ impl SystemInfo {
         println!("cargo:rerun-if-changed=libtch/stb_image_write.h");
         println!("cargo:rerun-if-changed=libtch/stb_image_resize.h");
         println!("cargo:rerun-if-changed=libtch/stb_image.h");
-        let mut c_files =
-            vec!["libtch/torch_api.cpp", "libtch/torch_api_generated.cpp", cuda_dependency];
+        let mut c_files = vec!["libtch/torch_api.cpp", "libtch/torch_api_generated.cpp"];
         if cfg!(feature = "python-extension") {
             c_files.push("libtch/torch_python.cpp")
         }
@@ -440,6 +433,9 @@ fn main() -> anyhow::Result<()> {
         // if this issue.
         // TODO: Try out the as-needed native link modifier when it lands.
         // https://github.com/rust-lang/rust/issues/99424
+        //
+        // Update: it seems that the dummy dependency is not necessary anymore, so just
+        // removing it and keeping this comment around for legacy.
         let si_lib = &system_info.libtorch_lib_dir;
         let use_cuda =
             si_lib.join("libtorch_cuda.so").exists() || si_lib.join("torch_cuda.dll").exists();
@@ -451,7 +447,7 @@ fn main() -> anyhow::Result<()> {
             si_lib.join("libtorch_hip.so").exists() || si_lib.join("torch_hip.dll").exists();
         println!("cargo:rustc-link-search=native={}", si_lib.display());
 
-        system_info.make(use_cuda, use_hip);
+        system_info.make();
 
         println!("cargo:rustc-link-lib=static=tch");
         if use_cuda {

torch-sys/src/lib.rs

@@ -333,7 +333,3 @@ extern "C" {
     pub fn atm_set_tensor_expr_fuser_enabled(enabled: c_int);
     pub fn atm_get_tensor_expr_fuser_enabled() -> bool;
 }
-
-extern "C" {
-    pub fn dummy_cuda_dependency();
-}