diff --git a/torch_npu/meta/_meta_registrations.py b/torch_npu/meta/_meta_registrations.py
index 8f8caac40a753364a1e890c25f045921c38309bf..3da24df98b3339e1b1abe9c9cf60b637dc2c7993 100644
--- a/torch_npu/meta/_meta_registrations.py
+++ b/torch_npu/meta/_meta_registrations.py
@@ -694,6 +694,32 @@ def npu_apply_rotary_pos_emb_meta(query, key, cos, sin, layout=1):
     return (torch.empty_like(query, dtype=query.dtype), torch.empty_like(key, dtype=key.dtype))
 
 
+@impl(m, "npu_moe_finalize_routing_v2_backward")
+def npu_moe_finalize_routing_v2_backward_meta(grad_y, expanded_row_idx, expanded_x=None, scales=None, expert_idx=None,
+                                              bias=None, drop_pad_mode=0, active_num=0, expert_num=0,
+                                              expert_capacity=0):
+    grad_expanded_x_dim0 = expanded_row_idx.size(0)
+    if drop_pad_mode == 0 and active_num > 0 and active_num < grad_expanded_x_dim0:
+        grad_expanded_x_dim0 = active_num
+    elif drop_pad_mode == 1:
+        if expert_num == 0 or expert_capacity == 0:
+            raise RuntimeError("When drop_pad_mode is 1, expert_num and expert_capacity should be greater than 0.")
+        grad_expanded_x_dim0 = expert_num * expert_capacity
+    grad_expanded_x_dim1 = grad_y.size(1)
+    grad_expanded_x = torch.empty([grad_expanded_x_dim0, grad_expanded_x_dim1], dtype=grad_y.dtype, device='meta')
+
+    grad_scales_dim0 = grad_y.size(0)
+    grad_scales_dim1 = 1
+    if grad_y.size(0) != 0:
+        if expanded_row_idx.size(0) % grad_y.size(0) == 0:
+            grad_scales_dim1 = expanded_row_idx.size(0) / grad_y.size(0)
+        else:
+            raise RuntimeError("The first dim of expanded_row_idx should be a multiple of the first dim of grad_y.")
+    grad_scales = torch.empty([grad_scales_dim0, grad_scales_dim1], dtype=grad_y.dtype, device='meta')
+    
+    return (torch.empty_like(grad_expanded_x), torch.empty_like(grad_scales))
+
+
 @impl(m, "npu_quant_conv2d")
 def npu_quant_conv2d(input_, weight, scale, strides, pads, dilations,
                      groups=1, offset_x=0, round_mode='rint', output_dtype=None, bias=None, offset=None):