more fixes on dtypes

csrc/multi_tensor_lamb.cu

@@ -387,7 +387,7 @@ void multi_tensor_lamb_cuda(
         epsilon,
         (adamMode_t) mode,
         weight_decay,
-        global_grad_norm.data<float>(),
+        global_grad_norm.DATA_PTR<float>(),
         max_grad_norm); )
 
   // Compute update norms

csrc/multi_tensor_lamb_stage_1.cu

@@ -123,7 +123,7 @@ void multi_tensor_lamb_stage1_cuda(
 {
   using namespace at;
 
-  auto g_grad_norm = global_grad_norm.data<float>();
+  const float* g_grad_norm = global_grad_norm.DATA_PTR<float>();
   float clipped_global_grad_norm = *(g_grad_norm) > max_global_grad_norm ? *(g_grad_norm) / max_global_grad_norm : 1.0f;
   float next_step = float(step+1);
   float beta1_correction = 1.0f - std::pow(beta1, next_step);