Before remove redundant output from group op: 
() = "builtin.module" [id:181] () {program:0x55e308966350} : () -> 
{
    (%0) = "pd_op.data" [id:183] () {dtype:int32,name:"var_1",place:Place(undefined:0),shape:[-1,1],stop_gradient:[true]} : () -> tensor<-1x1xi32>
    (%1) = "pd_op.data" [id:184] () {dtype:float32,name:"var_0",place:Place(undefined:0),shape:[-1,80],stop_gradient:[true]} : () -> tensor<-1x80xf32>
    (%2, %3, %4, %5, %6, %7, %8, %9, %10) = "cinn_op.group" [id:279] () -> tensor<-1x80xf32>, tensor<-1x80xf32>, tensor<-1x80xf32>, tensor<-1x80xf32>, tensor<80x-1xf32>, tensor<-1x1xi32>, tensor<80x-1xf32>, tensor<-1x1xi32>, tensor<-1x-1xf32> {
        (%11) = "pd_op.full" [id:185] () {dtype:float32,place:Place(undefined:0),shape:[1],stop_gradient:[true],value:1} : () -> tensor<1xf32>
        (%12) = "cinn_op.scale" [id:259] (%1) {bias:0,bias_after_scale:true,scale:-1,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%13) = "cinn_op.scale" [id:260] (%12) {bias:0,bias_after_scale:true,scale:1,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%14) = "pd_op.exp" [id:190] (%13) {stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%15) = "cinn_op.generate_shape" [id:293] (%1, %0) {output_dim_exprs:["S1",80],stop_gradient:[true],symbol_bindings:[["ShapeSymbolBinding","S1",0,0]]} : (tensor<-1x80xf32>, tensor<-1x1xi32>) -> tensor<2xi64>
        (%16) = "pd_op.expand" [id:284] (%11, %15) {stop_gradient:[true]} : (tensor<1xf32>, tensor<2xi64>) -> tensor<-1x-1xf32>
        (%17) = "pd_op.add" [id:191] (%16, %14) {stop_gradient:[true]} : (tensor<-1x-1xf32>, tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%18) = "cinn_op.generate_shape" [id:294] (%1, %0) {output_dim_exprs:["S1",80],stop_gradient:[true],symbol_bindings:[["ShapeSymbolBinding","S1",0,0]]} : (tensor<-1x80xf32>, tensor<-1x1xi32>) -> tensor<2xi64>
        (%19) = "pd_op.expand" [id:288] (%11, %18) {stop_gradient:[true]} : (tensor<1xf32>, tensor<2xi64>) -> tensor<-1x-1xf32>
        (%20) = "pd_op.divide" [id:192] (%19, %17) {stop_gradient:[true]} : (tensor<-1x-1xf32>, tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%21) = "cinn_op.scale" [id:261] (%20) {bias:0,bias_after_scale:true,scale:-1,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%22) = "cinn_op.scale" [id:262] (%21) {bias:1,bias_after_scale:true,scale:1,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%23) = "cinn_op.scale" [id:263] (%22) {bias:0,bias_after_scale:true,scale:1,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%24) = "cinn_op.scale" [id:264] (%23) {bias:1e-12,bias_after_scale:true,scale:1,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%25) = "pd_op.log" [id:201] (%24) {stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%26) = "cinn_op.scale" [id:265] (%25) {bias:0,bias_after_scale:true,scale:-1,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%27) = "cinn_op.scale" [id:266] (%26) {bias:0,bias_after_scale:true,scale:1,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%28) = "cinn_op.scale" [id:267] (%27) {bias:0,bias_after_scale:true,scale:0.75,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%29) = "cinn_op.scale" [id:268] (%28) {bias:0,bias_after_scale:true,scale:1,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%30) = "pd_op.multiply" [id:210] (%20, %20) {stop_gradient:[true]} : (tensor<-1x80xf32>, tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%31) = "pd_op.multiply" [id:211] (%29, %30) {stop_gradient:[true]} : (tensor<-1x80xf32>, tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%32) = "cinn_op.scale" [id:269] (%20) {bias:0,bias_after_scale:true,scale:1,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%33) = "cinn_op.scale" [id:270] (%32) {bias:1e-12,bias_after_scale:true,scale:1,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%34) = "pd_op.log" [id:216] (%33) {stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%35) = "cinn_op.scale" [id:271] (%34) {bias:0,bias_after_scale:true,scale:-1,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%36) = "cinn_op.scale" [id:272] (%35) {bias:0,bias_after_scale:true,scale:1,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%37) = "cinn_op.scale" [id:273] (%36) {bias:0,bias_after_scale:true,scale:0.25,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%38) = "cinn_op.scale" [id:274] (%37) {bias:0,bias_after_scale:true,scale:1,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%39) = "cinn_op.scale" [id:275] (%20) {bias:0,bias_after_scale:true,scale:-1,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%40) = "cinn_op.scale" [id:276] (%39) {bias:1,bias_after_scale:true,scale:1,stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%41) = "pd_op.multiply" [id:229] (%40, %40) {stop_gradient:[true]} : (tensor<-1x80xf32>, tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%42) = "pd_op.multiply" [id:230] (%38, %41) {stop_gradient:[true]} : (tensor<-1x80xf32>, tensor<-1x80xf32>) -> tensor<-1x80xf32>
        (%43) = "pd_op.transpose" [id:233] (%42) {perm:[1,0],stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<80x-1xf32>
        (%44) = "pd_op.gather_nd" [id:236] (%43, %0) {stop_gradient:[true]} : (tensor<80x-1xf32>, tensor<-1x1xi32>) -> tensor<-1x-1xf32>
        (%45) = "pd_op.transpose" [id:237] (%44) {perm:[1,0],stop_gradient:[true]} : (tensor<-1x-1xf32>) -> tensor<-1x-1xf32>
        (%46) = "pd_op.transpose" [id:240] (%31) {perm:[1,0],stop_gradient:[true]} : (tensor<-1x80xf32>) -> tensor<80x-1xf32>
        (%47) = "pd_op.gather_nd" [id:243] (%46, %0) {stop_gradient:[true]} : (tensor<80x-1xf32>, tensor<-1x1xi32>) -> tensor<-1x-1xf32>
        (%48) = "pd_op.transpose" [id:244] (%47) {perm:[1,0],stop_gradient:[true]} : (tensor<-1x-1xf32>) -> tensor<-1x-1xf32>
        (%49) = "pd_op.subtract" [id:245] (%45, %48) {stop_gradient:[true]} : (tensor<-1x-1xf32>, tensor<-1x-1xf32>) -> tensor<-1x-1xf32>
        (%50) = "cinn_op.scale" [id:277] (%49) {bias:0,bias_after_scale:true,scale:2,stop_gradient:[true]} : (tensor<-1x-1xf32>) -> tensor<-1x-1xf32>
        (%51) = "cinn_op.scale" [id:278] (%50) {bias:0,bias_after_scale:true,scale:1,stop_gradient:[true]} : (tensor<-1x-1xf32>) -> tensor<-1x-1xf32>
        () = "cf.yield" [id:280] (%29, %30, %38, %41, %43, %0, %46, %0, %51) {} : (tensor<-1x80xf32>, tensor<-1x80xf32>, tensor<-1x80xf32>, tensor<-1x80xf32>, tensor<80x-1xf32>, tensor<-1x1xi32>, tensor<80x-1xf32>, tensor<-1x1xi32>, tensor<-1x-1xf32>) -> 
    }
    () = "builtin.shadow_output" [id:250] (%2) {output_name:"output_0"} : (tensor<-1x80xf32>) -> 
    () = "builtin.shadow_output" [id:251] (%3) {output_name:"output_1"} : (tensor<-1x80xf32>) -> 
    () = "builtin.shadow_output" [id:252] (%4) {output_name:"output_2"} : (tensor<-1x80xf32>) -> 
    () = "builtin.shadow_output" [id:253] (%5) {output_name:"output_3"} : (tensor<-1x80xf32>) -> 
    () = "builtin.shadow_output" [id:254] (%6) {output_name:"output_4"} : (tensor<80x-1xf32>) -> 
    () = "builtin.shadow_output" [id:255] (%7) {output_name:"output_5"} : (tensor<-1x1xi32>) -> 
    () = "builtin.shadow_output" [id:256] (%8) {output_name:"output_6"} : (tensor<80x-1xf32>) -> 
    () = "builtin.shadow_output" [id:257] (%9) {output_name:"output_7"} : (tensor<-1x1xi32>) -> 
    () = "builtin.shadow_output" [id:258] (%10) {output_name:"output_8"} : (tensor<-1x-1xf32>) -> 
}