diff --git a/vllm_ascend/ops/linear_op.py b/vllm_ascend/ops/linear_op.py
index 57044f5..f6feadd 100644
--- a/vllm_ascend/ops/linear_op.py
+++ b/vllm_ascend/ops/linear_op.py
@@ -299,6 +299,7 @@ class OProjRowParallelOp(CustomRowParallelOp):
 
         # otp-specific: Combine partial results across devices
         output = self.comm_group.reduce_scatter(output_parallel, dim=0)
+        output = output.view(input_.shape[0], self.layer.output_size)
 
         # Handle bias return based on configuration
         output_bias = self.bias if self.skip_bias_add else None