Revert "Links most cuda libs to jnitorch_cuda only" and preload/link cleanup

Author: HGuillemet

pytorch/src/main/java/org/bytedeco/pytorch/presets/torch.java

@@ -86,7 +86,7 @@
 
             },
             exclude = {"openblas_config.h", "cblas.h", "lapacke_config.h", "lapacke_mangling.h", "lapack.h", "lapacke.h", "lapacke_utils.h"},
-            preload = { "asmjit", "fbgemm" }
+            link = { "torch" }
         ),
         @Platform(
             value = {"linux", "macosx", "windows"},
@@ -105,16 +105,27 @@
         ),
         @Platform(
             value = {"linux"},
-            link = { "c10", "torch_cpu" }
+            preload = { "c10", "torch_cpu" },
+
         ),
         @Platform(
             value = {"macosx"},
-            link = { "c10", "torch_cpu", "iomp5" }
+            preload = { "c10", "torch_cpu", "iomp5" }
         ),
         @Platform(
             value = "windows",
-            link = { "c10", "torch_cpu", "uv" }
+            preload = { "c10", "torch_cpu", "uv" }
+        ),
+        @Platform(
+            value = "linux",
+            extension = "-gpu",
+            preload = { "c10", "torch_cpu", "c10_cuda", "torch_cuda" }
         ),
+        @Platform(
+            value = "windows",
+            extension = "-gpu",
+            preload = { "c10", "torch_cpu", "uv", "c10_cuda", "torch_cuda" }
+        )
     },
     target = "org.bytedeco.pytorch",
     global = "org.bytedeco.pytorch.global.torch"
@@ -1959,11 +1970,11 @@ We need either to put an annotation info on each member, or javaName("@NoOffset
             "c10::DDPLoggingData::strs_map",
             "c10::DDPLoggingData::ints_map",
             "torch::dynamo::autograd::TensorArgs::inputs",
+            "torch::dynamo::autograd::AutogradCompilerCall::tensor_args",
             "torch::dynamo::autograd::AutogradCompilerCall::all_size_inputs",
             "torch::dynamo::autograd::AutogradCompilerCall::dyn_size_inputs",
             "torch::dynamo::autograd::AutogradCompilerCall::node_calls",
             "torch::dynamo::autograd::AutogradCompilerCall::default_dyn_type",
-            "torch::dynamo::autograd::AutogradCompilerCall::tensor_args",
             "torch::jit::Object::Property::setter_func",
             "torch::jit::Object::Property::getter_func",
             "torch::jit::Object::Property::name",

pytorch/src/main/java/org/bytedeco/pytorch/presets/torch_cuda.java

@@ -58,22 +58,12 @@
         @Platform(
             value = "linux",
             extension = "-gpu",
-            link = { "c10", "torch_cpu", "c10_cuda", "torch_cuda", "torch_cuda_linalg", "cudart", "cupti", "cusparse", "cudnn" }, // cuda_linalg built as separate lib on linux only
-            linkpath = {
-                "/usr/local/cuda-12.3/lib64/",
-                "/usr/local/cuda/lib64/",
-                "/usr/local/cuda-12.3/extras/CUPTI/lib64/",
-                "/usr/lib64/"
-            }
+            link = { "torch" , "c10_cuda", "torch_cuda_linalg" } // cuda_linalg built as separate lib on linux only
         ),
         @Platform(
             value = "windows",
             extension = "-gpu",
-            link = { "c10", "torch_cpu", "uv", "c10_cuda", "torch_cuda", "cudart", "cupti", "cusparse", "cudnn" },
-            linkpath = {
-                "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.3/lib/x64/",
-                "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.3/extras/CUPTI/lib64/"
-            }
+            link = { "torch" , "c10_cuda" }
         )
     },
     target = "org.bytedeco.pytorch.cuda",