issue/314/fix: cuda compilation and test

Signed-off-by: YdrMaster <ydrml@hotmail.com>

test/infiniop/rope.py

@@ -94,7 +94,7 @@ def rotary_embedding(ans, t, sin, cos, device):
 def sin_cos_table(pos, dim, device, theta, dtype):
     assert dim % 2 == 0, "Embedding dimension must be even."
     freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
-    angles = torch.outer(pos, freqs)
+    angles = torch.outer(pos.cpu(), freqs)
     return (
         TestTensor.from_torch(torch.sin(angles), dtype, device),
         TestTensor.from_torch(torch.cos(angles), dtype, device),

xmake/cuda.lua

@@ -15,8 +15,7 @@ target("infiniop-cuda")
 
     set_policy("build.cuda.devlink", true)
     set_toolchains("cuda")
-    add_links("cuda", "cublas")
-    add_linkdirs(CUDA_ROOT .. "/lib64/stubs")
+    add_links("cudart", "cublas")
     if has_config("cudnn") then
         add_links("cudnn")
     end