[Prim][CINN] Fix group_norm_grad decomp sum dtype (#73037)

Author: lshpku

paddle/fluid/primitive/decomp_rule/decomp_vjp/details.h

@@ -2307,15 +2307,18 @@ void group_norm_grad(const Tensor& x,
       auto tmp1 = out_grad_data * (x_data - mean_new) * sqrt_var_1;
 
       auto scale_grad_tmp = reshape<T>(
-          tmp1.sum(reduce_axis_except_channel, scale->dtype(), false), {-1});
+          tmp1.sum(reduce_axis_except_channel, x_data.dtype(), false), {-1});
+      scale_grad_tmp = ConvertToOrig<T>(scale_grad_tmp, scale->dtype());
+
       set_output<T>(scale_grad_tmp, scale_grad);
     }
   }
 
   if (bias_grad) {
     if (bias) {
       auto bias_grad_tmp =
-          out_grad_data.sum(reduce_axis_except_channel, bias->dtype(), false);
+          out_grad_data.sum(reduce_axis_except_channel, x_data.dtype(), false);
+      bias_grad_tmp = ConvertToOrig<T>(bias_grad_tmp, bias->dtype());
 
       set_output<T>(reshape<T>(bias_grad_tmp, {-1}), bias_grad);
     }

test/deprecated/legacy_test/test_group_norm_op_deprecated.py

@@ -586,6 +586,7 @@ def setUp(self):
         self.python_out_sig = ["Y"]
         self.data_format = "NHWC"
         self.prim_op_type = "comp"
+        self.channel_last = True
 
         self.dtype = np.uint16
         self.shape = (1, 3, 5, 512)