update

Author: chen2016013

test/ir/pir/cinn/symbolic/simple_llama.config

@@ -94,11 +94,11 @@
     (%93) = "pd_op.matmul" (%91, %3) {is_persistable:[false],stop_gradient:[false],transpose_x:false,transpose_y:false} : (builtin.tensor<-1x-1x4096xf16>, builtin.tensor<4096x4096xf16>) -> builtin.tensor<-1x-1x4096xf16>
     (%94) = "pd_op.matmul" (%91, %4) {is_persistable:[false],stop_gradient:[false],transpose_x:false,transpose_y:false} : (builtin.tensor<-1x-1x4096xf16>, builtin.tensor<4096x4096xf16>) -> builtin.tensor<-1x-1x4096xf16>
     (%95) = "pd_op.full_int_array" () {dtype:(pd_op.DataType)int64,place:(pd_op.Place)Place(cpu),stop_gradient:[true],value:[(Int64)0,(Int64)0,(Int64)32,(Int64)128]} : () -> builtin.tensor<4xi64>
-    (%96, %97) = "pd_op.reshape" (%92, %95) {is_persistable:[false,false],stop_gradient:[false,false]} : (builtin.tensor<-1x-1x4096xf16>, builtin.tensor<4xi64>) -> builtin.tensor<-1x-1x32x128xf16>, builtin.tensor<0x-1x-1x4096xf16>
+    (%96) = "pd_op.reshape" (%92, %95) {is_persistable:[false,false],stop_gradient:[false,false]} : (builtin.tensor<-1x-1x4096xf16>, builtin.tensor<4xi64>) -> builtin.tensor<-1x-1x32x128xf16>
     (%98) = "pd_op.full_int_array" () {dtype:(pd_op.DataType)int64,place:(pd_op.Place)Place(cpu),stop_gradient:[true],value:[(Int64)0,(Int64)0,(Int64)32,(Int64)128]} : () -> builtin.tensor<4xi64>
-    (%99, %100) = "pd_op.reshape" (%93, %98) {is_persistable:[false,false],stop_gradient:[false,false]} : (builtin.tensor<-1x-1x4096xf16>, builtin.tensor<4xi64>) -> builtin.tensor<-1x-1x32x128xf16>, builtin.tensor<0x-1x-1x4096xf16>
+    (%99) = "pd_op.reshape" (%93, %98) {is_persistable:[false,false],stop_gradient:[false,false]} : (builtin.tensor<-1x-1x4096xf16>, builtin.tensor<4xi64>) -> builtin.tensor<-1x-1x32x128xf16>
     (%101) = "pd_op.full_int_array" () {dtype:(pd_op.DataType)int64,place:(pd_op.Place)Place(cpu),stop_gradient:[true],value:[(Int64)0,(Int64)0,(Int64)32,(Int64)128]} : () -> builtin.tensor<4xi64>
-    (%102, %103) = "pd_op.reshape" (%94, %101) {is_persistable:[false,false],stop_gradient:[false,false]} : (builtin.tensor<-1x-1x4096xf16>, builtin.tensor<4xi64>) -> builtin.tensor<-1x-1x32x128xf16>, builtin.tensor<0x-1x-1x4096xf16>
+    (%102) = "pd_op.reshape" (%94, %101) {is_persistable:[false,false],stop_gradient:[false,false]} : (builtin.tensor<-1x-1x4096xf16>, builtin.tensor<4xi64>) -> builtin.tensor<-1x-1x32x128xf16>
     (%104) = "pd_op.shape" (%99) {is_persistable:[false],stop_gradient:[false]} : (builtin.tensor<-1x-1x32x128xf16>) -> builtin.tensor<4xi32>
     (%105) = "pd_op.full_int_array" () {dtype:(pd_op.DataType)int64,place:(pd_op.Place)Place(cpu),stop_gradient:[true],value:[(Int64)1]} : () -> builtin.tensor<1xi64>
     (%106) = "pd_op.full_int_array" () {dtype:(pd_op.DataType)int64,place:(pd_op.Place)Place(cpu),stop_gradient:[true],value:[(Int64)2]} : () -> builtin.tensor<1xi64>
@@ -173,7 +173,7 @@
     (%181) = "pd_op.matmul" (%179, %180) {is_persistable:[false],stop_gradient:[false],transpose_x:false,transpose_y:false} : (builtin.tensor<-1x32x-1x128xf16>, builtin.tensor<-1x32x128x-1xf16>) -> builtin.tensor<-1x32x-1x-1xf16>
     (%182) = "pd_op.full" () {dtype:(pd_op.DataType)int32,is_persistable:[false],place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[1],stop_gradient:[false],value:(Float)1} : () -> builtin.tensor<1xi32>
     (%183) = "builtin.combine" (%167, %182, %170, %174) {} : (builtin.tensor<i32>, builtin.tensor<1xi32>, builtin.tensor<i32>, builtin.tensor<i32>) -> vec[builtin.tensor<i32>,builtin.tensor<1xi32>,builtin.tensor<i32>,builtin.tensor<i32>]
-    (%184, %185) = "pd_op.reshape" (%82, %183) {is_persistable:[false,false],stop_gradient:[false,false]} : (builtin.tensor<-1x1x-1x-1xf16>, vec[builtin.tensor<i32>,builtin.tensor<1xi32>,builtin.tensor<i32>,builtin.tensor<i32>]) -> builtin.tensor<-1x1x-1x-1xf16>, builtin.tensor<0x-1x1x-1x-1xf16>
+    (%184) = "pd_op.reshape" (%82, %183) {is_persistable:[false,false],stop_gradient:[false,false]} : (builtin.tensor<-1x1x-1x-1xf16>, vec[builtin.tensor<i32>,builtin.tensor<1xi32>,builtin.tensor<i32>,builtin.tensor<i32>]) -> builtin.tensor<-1x1x-1x-1xf16>
     (%186) = "pd_op.add" (%181, %184) {is_persistable:[false],stop_gradient:[false]} : (builtin.tensor<-1x32x-1x-1xf16>, builtin.tensor<-1x1x-1x-1xf16>) -> builtin.tensor<-1x32x-1x-1xf16>
     (%187) = "pd_op.cast" (%186) {dtype:(pd_op.DataType)float32,is_persistable:[false],stop_gradient:[false]} : (builtin.tensor<-1x32x-1x-1xf16>) -> builtin.tensor<-1x32x-1x-1xf32>
     (%188) = "pd_op.softmax" (%187) {axis:(Int32)-1,is_persistable:[false],stop_gradient:[false]} : (builtin.tensor<-1x32x-1x-1xf32>) -> builtin.tensor<-1x32x-1x-1xf32>
@@ -182,7 +182,7 @@
     (%191) = "pd_op.transpose" (%190) {is_persistable:[false],perm:[(Int32)0,(Int32)2,(Int32)1,(Int32)3],stop_gradient:[false]} : (builtin.tensor<-1x32x-1x128xf16>) -> builtin.tensor<-1x-1x32x128xf16>
     (%192) = "pd_op.full" () {dtype:(pd_op.DataType)int32,is_persistable:[false],place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[1],stop_gradient:[false],value:(Float)4096} : () -> builtin.tensor<1xi32>
     (%193) = "builtin.combine" (%167, %170, %192) {} : (builtin.tensor<i32>, builtin.tensor<i32>, builtin.tensor<1xi32>) -> vec[builtin.tensor<i32>,builtin.tensor<i32>,builtin.tensor<1xi32>]
-    (%194, %195) = "pd_op.reshape" (%191, %193) {is_persistable:[false,false],stop_gradient:[false,false]} : (builtin.tensor<-1x-1x32x128xf16>, vec[builtin.tensor<i32>,builtin.tensor<i32>,builtin.tensor<1xi32>]) -> builtin.tensor<-1x-1x4096xf16>, builtin.tensor<0x-1x-1x32x128xf16>
+    (%194) = "pd_op.reshape" (%191, %193) {is_persistable:[false,false],stop_gradient:[false,false]} : (builtin.tensor<-1x-1x32x128xf16>, vec[builtin.tensor<i32>,builtin.tensor<i32>,builtin.tensor<1xi32>]) -> builtin.tensor<-1x-1x4096xf16>
     (%196) = "pd_op.matmul" (%194, %7) {is_persistable:[false],stop_gradient:[false],transpose_x:false,transpose_y:false} : (builtin.tensor<-1x-1x4096xf16>, builtin.tensor<4096x4096xf16>) -> builtin.tensor<-1x-1x4096xf16>
     (%197) = "pd_op.add" (%49, %196) {is_persistable:[false],stop_gradient:[false]} : (builtin.tensor<-1x-1x4096xf16>, builtin.tensor<-1x-1x4096xf16>) -> builtin.tensor<-1x-1x4096xf16>
     (%198) = "pd_op.cast" (%197) {dtype:(pd_op.DataType)float32,is_persistable:[false],stop_gradient:[false]} : (builtin.tensor<-1x-1x4096xf16>) -> builtin.tensor<-1x-1x4096xf32>

test/ir/pir/cinn/symbolic/test_simple_llama_dy.py

@@ -57,16 +57,6 @@ def __init__(self):
     def run(self, file):
         program = self.load_from(file)
         for op in program.global_block().ops:
-            if op.name() == "pd_op.reshape":
-                if (
-                    op.result(1).initialized()
-                    and not op.result(1).use_empty()
-                    and op.result(1).first_use().owner().name() == "pd_op.fetch"
-                ):
-                    program.global_block().remove_op(
-                        op.result(1).first_use().owner()
-                    )
-
             if op.name() == "pd_op.squeeze":
                 if (
                     op.result(1).initialized()