bugfix: correct rotary_dim calculation in apply_rotary function to use sin tensor size.

Author: XuZhang99

xllm/core/kernels/mlu/rope.cpp

@@ -28,7 +28,7 @@ void apply_rotary(torch::Tensor& q,
                   bool discrete,
                   bool dynamic_ntk,
                   int max_query_len) {
-  const int64_t rotary_dim = q.size(-1);
+  const int64_t rotary_dim = sin.size(-1);
   const int64_t T = q.size(0);
   q = q.view({T, -1});
   k = k.view({T, -1});