[XPU] Fusion of gather and assign operators to fused_mt op for reducing memory usage (#53262)

paddle/fluid/framework/ir/CMakeLists.txt

@@ -242,7 +242,7 @@ if(WITH_XPU)
   pass_library(one_beam_size_fuse_pass inference DIR xpu DEPS ${XPU_PASS_DEPS})
   pass_library(delete_isolated_node_pass inference DIR xpu DEPS
                ${XPU_PASS_DEPS})
-  pass_library(fused_multi_transformer_xpu_quant_pass inference DIR xpu DEPS
+  pass_library(fused_multi_transformer_xpu_pass inference DIR xpu DEPS
                ${XPU_PASS_DEPS})
   pass_library(stack_fuse_pass inference DIR xpu DEPS ${XPU_PASS_DEPS})
   pass_library(fused_multi_transformer_cachekv_layout_trans_pass inference DIR
@@ -519,9 +519,9 @@ if(WITH_XPU)
     SRCS xpu/delete_isolated_node_pass_test.cc
     DEPS delete_isolated_node_pass)
   cc_test(
-    test_fused_multi_transformer_xpu_quant_pass
-    SRCS xpu/fused_multi_transformer_xpu_quant_pass_tester.cc
-    DEPS fused_multi_transformer_xpu_quant_pass)
+    test_fused_multi_transformer_xpu_pass
+    SRCS xpu/fused_multi_transformer_xpu_pass_tester.cc
+    DEPS fused_multi_transformer_xpu_pass)
   cc_test(
     test_one_beam_size_fuse_pass
     SRCS xpu/one_beam_size_fuse_pass_test.cc

paddle/fluid/framework/ir/pass.cc

@@ -65,7 +65,7 @@ static const std::vector<std::string> xpu_support_subgraph_passes = {
     "fused_multi_transformer_cachekv_layout_trans_pass",
     "one_beam_size_fuse_pass",
     "stack_fuse_pass",
-    "fused_multi_transformer_xpu_quant_pass",
+    "fused_multi_transformer_xpu_pass",
     "fc_xpu_fuse_pass",
     "link_xpu_op_max_pass",
 };

paddle/fluid/framework/ir/xpu/fused_multi_transformer_xpu_quant_pass.cc → paddle/fluid/framework/ir/xpu/fused_multi_transformer_xpu_pass.cc

@@ -39,6 +39,32 @@ namespace framework {
 namespace ir {
 namespace patterns {
 
+struct FusedMultiTransformerAssignPattern : public PatternBase {
+  FusedMultiTransformerAssignPattern(PDPattern* pattern,
+                                     const std::string& name_scope);
+  // declare operator node's name
+  PATTERN_DECL_NODE(assign);
+  // declare variable node's name
+  PATTERN_DECL_NODE(assign_out);
+};
+
+FusedMultiTransformerAssignPattern::FusedMultiTransformerAssignPattern(
+    PDPattern* pattern, const std::string& name_scope)
+    : PatternBase(pattern, name_scope, name_scope) {
+  auto* assign =
+      pattern->NewNode(assign_repr())
+          ->assert_is_op("assign")
+          ->assert_more([&](Node* node) {
+            auto pre_op_nodes = node->inputs[0]->inputs;
+            return pre_op_nodes.size() == 1 &&
+                   pre_op_nodes[0]->Op()->Type() == "fused_multi_transformer";
+          });
+  auto* assign_out =
+      pattern->NewNode(assign_out_repr())->assert_is_op_output("assign", "Out");
+
+  assign->LinksTo({assign_out});
+}
+
 struct FusedMultiTransformerPattern : public PatternBase {
   FusedMultiTransformerPattern(PDPattern* pattern,
                                const std::string& name_scope,
@@ -47,7 +73,6 @@ struct FusedMultiTransformerPattern : public PatternBase {
                                bool with_time_step,
                                bool with_seq_lengths,
                                bool with_src_mask);
-
   // declare operator node's name
   PATTERN_DECL_NODE(fused_mt);
   // declare variable node's name
@@ -234,44 +259,106 @@ FusedMultiTransformerPattern::FusedMultiTransformerPattern(
 }  // namespace patterns
 
 /*
-1. transpose and quantify the weights of fused_multi_transformer op from fp32 to
+1. Remove gather and assign op to reduce graphics memory consumption
+2. transpose and quantify the weights of fused_multi_transformer op from fp32 to
 int16
 */
-class FusedMultiTransformerXPUQuantPass : public FusePassBase {
+class FusedMultiTransformerXPUPass : public FusePassBase {
  protected:
   void ApplyImpl(ir::Graph* graph) const override;
 
  private:
-  int ApplyImpl(ir::Graph* graph,
-                bool with_pre_caches,
-                bool with_rotary_pos_emb,
-                bool with_time_step,
-                bool with_seq_lengths,
-                bool with_src_mask) const;
-
-  const std::string name_scope_{"fused_multi_transformer_xpu_quant_pass"};
+  /*
+  Origin subgraph:
+              fused_multi_transformer
+               |        |        |
+             assign   assign    ...
+               |        |        |
+             gather   gather    ...
+
+  Fused subgraph:
+              fused_multi_transformer
+ */
+  void RemoveAssignGather(ir::Graph* graph) const;
+
+  /*
+  Origin subgraph:
+              fused_multi_transformer
+
+  Fused subgraph:
+              fused_multi_transformer_xpu
+ */
+  int FusedMultiTransformerXPUQuant(ir::Graph* graph,
+                                    bool with_pre_caches,
+                                    bool with_rotary_pos_emb,
+                                    bool with_time_step,
+                                    bool with_seq_lengths,
+                                    bool with_src_mask) const;
+
+  const std::string name_scope_{"fused_multi_transformer_xpu_pass"};
 };
 
-void FusedMultiTransformerXPUQuantPass::ApplyImpl(ir::Graph* graph) const {
+void FusedMultiTransformerXPUPass::ApplyImpl(ir::Graph* graph) const {
   PADDLE_ENFORCE_NOT_NULL(
       graph, platform::errors::PreconditionNotMet("graph should not be null."));
   Init(name_scope_, graph);
-  VLOG(3) << "in FusedMultiTransformerXPUQuantPass::ApplyImpl";
+  VLOG(3) << "in FusedMultiTransformerXPUPass::ApplyImpl";
 
   int found_subgraph_count = 0;
+  RemoveAssignGather(graph);
   for (bool with_time_step : {true, false}) {
-    found_subgraph_count +=
-        ApplyImpl(graph, false, false, with_time_step, false, true);
+    found_subgraph_count += FusedMultiTransformerXPUQuant(
+        graph, false, false, with_time_step, false, true);
   }
   AddStatis(found_subgraph_count);
 }
 
-int FusedMultiTransformerXPUQuantPass::ApplyImpl(ir::Graph* graph,
-                                                 bool with_pre_caches,
-                                                 bool with_rotary_pos_emb,
-                                                 bool with_time_step,
-                                                 bool with_seq_lengths,
-                                                 bool with_src_mask) const {
+void FusedMultiTransformerXPUPass::RemoveAssignGather(ir::Graph* graph) const {
+  // detect assign + gather
+  GraphPatternDetector gpd;
+  patterns::FusedMultiTransformerAssignPattern pattern(gpd.mutable_pattern(),
+                                                       name_scope_);
+  int found_subgraph_count = 0;
+
+  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
+                     Graph* graph) {
+    VLOG(1) << "handle RemoveAssignGather";
+    GET_IR_NODE(assign);
+    GET_IR_NODE(assign_out);
+    // Assign_out may not link to gather, so we find gather by input name.
+    auto next_ops = FindOpNodeByInputName(graph, assign_out->Name());
+    if (next_ops.size() != 1 || next_ops[0]->Name() != "gather") return;
+    auto* gather = next_ops[0];
+
+    // "assign_out" is used in multi blocks. "assign_out" should be reserved.
+    auto* gather_index = gather->inputs[0];
+    auto* assign_in = assign->inputs[0];
+    auto* fused_multi_transformer = assign_in->inputs[0];
+    fused_multi_transformer->Op()->Rename(assign_in->Name(),
+                                          assign_out->Name());
+    fused_multi_transformer->Op()->SetInput("gather_index",
+                                            gather->Op()->Input("Index"));
+    fused_multi_transformer->Op()->SetAttr("gather_axis",
+                                           gather->Op()->GetAttr("axis"));
+    IR_NODE_LINK_TO(gather_index, fused_multi_transformer);
+    IR_NODE_LINK_TO(fused_multi_transformer, assign_out);
+
+    std::unordered_set<const Node*> delete_nodes{assign, assign_in, gather};
+    GraphSafeRemoveNodes(graph, delete_nodes);
+    found_subgraph_count++;
+  };
+
+  gpd(graph, handler);
+  AddStatis(found_subgraph_count);
+}
+
+int FusedMultiTransformerXPUPass::FusedMultiTransformerXPUQuant(
+    ir::Graph* graph,
+    bool with_pre_caches,
+    bool with_rotary_pos_emb,
+    bool with_time_step,
+    bool with_seq_lengths,
+    bool with_src_mask) const {
   GraphPatternDetector gpd;
   patterns::FusedMultiTransformerPattern pattern(gpd.mutable_pattern(),
                                                  name_scope_,
@@ -286,7 +373,7 @@ int FusedMultiTransformerXPUQuantPass::ApplyImpl(ir::Graph* graph,
   int found_subgraph_count = 0;
   auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                      Graph* graph) {
-    VLOG(4) << "handle FusedMultiTransformerXPUQuantPass fuse";
+    VLOG(4) << "handle FusedMultiTransformerXPUQuant";
 
     GET_IR_NODE(x);
     GET_IR_NODE(ln_scale);
@@ -459,6 +546,13 @@ int FusedMultiTransformerXPUQuantPass::ApplyImpl(ir::Graph* graph,
     if (name_caches.count("CacheKV") > 0) {
       fused_mt_xpu_op_desc->SetInput("cache_kv", name_caches.at("CacheKV"));
     }
+    if (name_caches.count("gather_index") > 0) {
+      fused_mt_xpu_op_desc->SetInput("gather_index",
+                                     name_caches.at("gather_index"));
+    }
+    if (!fused_mt_xpu_op_desc->HasAttr("gather_axis")) {
+      fused_mt_xpu_op_desc->SetAttr("gather_axis", 0);
+    }
     if (pre_caches) {
       fused_mt_xpu_op_desc->SetInput("pre_caches", name_caches.at("PreCaches"));
     }
@@ -529,5 +623,5 @@ int FusedMultiTransformerXPUQuantPass::ApplyImpl(ir::Graph* graph,
 }  // namespace framework
 }  // namespace paddle
 
-REGISTER_PASS(fused_multi_transformer_xpu_quant_pass,
-              paddle::framework::ir::FusedMultiTransformerXPUQuantPass);
+REGISTER_PASS(fused_multi_transformer_xpu_pass,
+              paddle::framework::ir::FusedMultiTransformerXPUPass);

paddle/fluid/framework/ir/xpu/fused_multi_transformer_xpu_quant_pass_tester.cc → paddle/fluid/framework/ir/xpu/fused_multi_transformer_xpu_pass_tester.cc

@@ -64,7 +64,62 @@ Scope* CreateParamScope() {
   return param_scope;
 }
 
-TEST(FusedMultiTransformerXPUQuantPass, context_stage) {
+VarDesc* Data(paddle::framework::BlockDesc* block,
+              std::string name,
+              std::vector<int64_t> shape = {},
+              bool is_persistable = false,
+              proto::VarType::Type data_type = proto::VarType::FP32) {
+  auto* var = block->Var(name);
+  var->SetType(proto::VarType::LOD_TENSOR);
+  var->SetDataType(data_type);
+  var->SetShape(shape);
+  var->SetPersistable(is_persistable);
+  return var;
+}
+
+TEST(RemoveAssignGather, basic) {
+  paddle::framework::ProgramDesc program;
+  auto* block = program.MutableBlock(0);
+
+  auto* x = Data(block, "fused_multi_transformer_x", {1, 1, 1536});
+  auto* cache_kv =
+      Data(block, "fused_multi_transformer_cache_kv", {2, 1, 24, 512, 64});
+  OpDesc* fused_multi_transformer_op = block->AppendOp();
+  fused_multi_transformer_op->SetType("fused_multi_transformer");
+  fused_multi_transformer_op->SetInput("X", {x->Name()});
+  fused_multi_transformer_op->SetInput("CacheKV", {cache_kv->Name()});
+  fused_multi_transformer_op->SetOutput("CacheKVOut", {cache_kv->Name()});
+
+  auto* assign_out = Data(block, "assign_out", cache_kv->GetShape());
+  OpDesc* assign_op = block->AppendOp();
+  assign_op->SetType("assign");
+  assign_op->SetInput("X", {cache_kv->Name()});
+  assign_op->SetOutput("Out", {assign_out->Name()});
+
+  OpDesc* gather_op = block->AppendOp();
+  auto gather_index = Data(block, "gather_index", {10});
+  gather_op->SetType("gather");
+  gather_op->SetInput("X", {assign_out->Name()});
+  gather_op->SetInput("Index", {gather_index->Name()});
+  gather_op->SetAttr("axis", {1});
+  gather_op->SetOutput("Out", {cache_kv->Name()});
+
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(program));
+  auto pass = PassRegistry::Instance().Get("fused_multi_transformer_xpu_pass");
+  pass->Apply(graph.get());
+  auto assign_num = GetNumOpNodes(graph, "assign");
+  auto gather_num = GetNumOpNodes(graph, "gather");
+  PADDLE_ENFORCE_EQ(assign_num,
+                    0,
+                    platform::errors::PreconditionNotMet(
+                        "assign op should be removed from the graph."));
+  PADDLE_ENFORCE_EQ(gather_num,
+                    0,
+                    platform::errors::PreconditionNotMet(
+                        "gather op should be removed from the graph."));
+}
+
+TEST(FusedMultiTransformerXPUPass, context_stage) {
   DEF_INPUT_DATA
 
   auto* cache_kv = layers.fill_constant_batch_size_like(
@@ -95,10 +150,9 @@ TEST(FusedMultiTransformerXPUQuantPass, context_stage) {
   std::unique_ptr<ir::Graph> graph(new ir::Graph(layers.main_program()));
   graph->Set("__param_scope__", CreateParamScope());
 
-  auto pass =
-      PassRegistry::Instance().Get("fused_multi_transformer_xpu_quant_pass");
+  auto pass = PassRegistry::Instance().Get("fused_multi_transformer_xpu_pass");
   if (pass.get() == nullptr) {
-    LOG(INFO) << "get fused_multi_transformer_xpu_quant_pass failed";
+    LOG(INFO) << "get fused_multi_transformer_xpu_pass failed";
   }
 
   graph.reset(pass->Apply(graph.release()));
@@ -114,7 +168,7 @@ TEST(FusedMultiTransformerXPUQuantPass, context_stage) {
           num_nodes_after));
 }
 
-TEST(FusedMultiTransformerXPUQuantPass, decoder_stage) {
+TEST(FusedMultiTransformerXPUPass, decoder_stage) {
   DEF_INPUT_DATA
 
   auto* cache_kv = layers.fill_constant_batch_size_like(
@@ -146,10 +200,9 @@ TEST(FusedMultiTransformerXPUQuantPass, decoder_stage) {
   std::unique_ptr<ir::Graph> graph(new ir::Graph(layers.main_program()));
   graph->Set("__param_scope__", CreateParamScope());
 
-  auto pass =
-      PassRegistry::Instance().Get("fused_multi_transformer_xpu_quant_pass");
+  auto pass = PassRegistry::Instance().Get("fused_multi_transformer_xpu_pass");
   if (pass.get() == nullptr) {
-    LOG(INFO) << "get fused_multi_transformer_xpu_quant_pass failed";
+    LOG(INFO) << "get fused_multi_transformer_xpu_pass failed";
   }
 
   graph.reset(pass->Apply(graph.release()));
@@ -169,4 +222,4 @@ TEST(FusedMultiTransformerXPUQuantPass, decoder_stage) {
 }  // namespace framework
 }  // namespace paddle
 
-USE_PASS(fused_multi_transformer_xpu_quant_pass);
+USE_PASS(fused_multi_transformer_xpu_pass);

paddle/fluid/framework/ir/xpu/one_beam_size_fuse_pass.cc

@@ -259,23 +259,6 @@ bool OnlyOneBeamSearchAndOneBeamSize(ir::Graph* graph) {
          beam_search_nodes[0]->Op()->GetAttrIfExists<int>("beam_size") == 1;
 }
 
-std::vector<Node*> FindOpNodeByInputName(Graph* graph,
-                                         const std::string& var_name) {
-  std::vector<Node*> ret;
-  for (auto* node : graph->Nodes()) {
-    if (!node->IsOp()) continue;
-    auto inputs = node->Op()->Inputs();
-    for (auto input : inputs) {
-      auto in_names = input.second;
-      if (std::count(in_names.begin(), in_names.end(), var_name) > 0) {
-        ret.push_back(node);
-        break;
-      }
-    }
-  }
-  return ret;
-}
-
 void OneBeamSizeFusePass::RemoveAssignGather(ir::Graph* graph) const {
   // detect assign + gather
   GraphPatternDetector gpd;

paddle/fluid/framework/ir/xpu/pass_utils.cc

@@ -71,6 +71,23 @@ Node* FindNodeWithName(Graph* graph, std::string name) {
   return nullptr;
 }
 
+std::vector<Node*> FindOpNodeByInputName(Graph* graph,
+                                         const std::string& var_name) {
+  std::vector<Node*> ret;
+  for (auto* node : graph->Nodes()) {
+    if (!node->IsOp()) continue;
+    auto inputs = node->Op()->Inputs();
+    for (auto input : inputs) {
+      auto in_names = input.second;
+      if (std::count(in_names.begin(), in_names.end(), var_name) > 0) {
+        ret.push_back(node);
+        break;
+      }
+    }
+  }
+  return ret;
+}
+
 template <typename T>
 std::string IntTypeToString() {
   LOG(FATAL) << "Not support type.";

paddle/fluid/framework/ir/xpu/pass_utils.h

@@ -51,6 +51,9 @@ int ConvertActivationType(std::string act_type);
 
 Node* FindNodeWithName(Graph* graph, std::string name);
 
+std::vector<Node*> FindOpNodeByInputName(Graph* graph,
+                                         const std::string& var_name);
+
 template <typename T>
 size_t HashTensor(const phi::DenseTensor& in);
 

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -523,7 +523,7 @@ XpuPassStrategy::XpuPassStrategy() : PassStrategy({}) {
       "one_beam_size_fuse_pass",
       "delete_cast_op_pass",
       "stack_fuse_pass",
-      "fused_multi_transformer_xpu_quant_pass",
+      "fused_multi_transformer_xpu_pass",
       "fc_xpu_fuse_pass",
       "conv2d_xpu_fuse_pass",
       "link_xpu_op_max_pass",

paddle/phi/api/yaml/fused_ops.yaml

@@ -58,14 +58,14 @@
   support_dygraph_mode : true
 
 - op : fused_multi_transformer_xpu
-  args : (Tensor x, Tensor[] ln_scale, Tensor[] ln_bias, Tensor[] qkvw, Tensor[] qkvw_max, Tensor[] qkv_bias, Tensor[] out_linear_w, Tensor[] out_linear_wmax, Tensor[] out_linear_bias, Tensor[] ffn_ln_scale, Tensor[] ffn_ln_bias, Tensor[] ffn1_weight, Tensor[] ffn1_weight_max, Tensor[] ffn1_bias, Tensor[] ffn2_weight, Tensor[] ffn2_weight_max, Tensor[] ffn2_bias, Tensor[] cache_kv, Tensor[] pre_caches, Tensor rotary_pos_emb, Tensor time_step, Tensor seq_lengths, Tensor src_mask, bool pre_layer_norm, int rotary_emb_dims, float epsilon, float dropout_rate, bool is_test, str dropout_implementation, str act_method, bool trans_qkvw, int ring_id)
+  args : (Tensor x, Tensor[] ln_scale, Tensor[] ln_bias, Tensor[] qkvw, Tensor[] qkvw_max, Tensor[] qkv_bias, Tensor[] out_linear_w, Tensor[] out_linear_wmax, Tensor[] out_linear_bias, Tensor[] ffn_ln_scale, Tensor[] ffn_ln_bias, Tensor[] ffn1_weight, Tensor[] ffn1_weight_max, Tensor[] ffn1_bias, Tensor[] ffn2_weight, Tensor[] ffn2_weight_max, Tensor[] ffn2_bias, Tensor[] cache_kv, Tensor[] pre_caches, Tensor rotary_pos_emb, Tensor time_step, Tensor seq_lengths, Tensor src_mask, Tensor gather_index, bool pre_layer_norm, int rotary_emb_dims, float epsilon, float dropout_rate, bool is_test, str dropout_implementation, str act_method, bool trans_qkvw, int ring_id, int gather_axis)
   output : Tensor(out), Tensor[](cache_kv_out){out_linear_w.size()}
   infer_meta :
     func : FusedMultiTransformerXpuInferMeta
   kernel :
     func : fused_multi_transformer_xpu
     data_type : x
-  optional : cache_kv, pre_caches, rotary_pos_emb, time_step, seq_lengths, src_mask
+  optional : cache_kv, pre_caches, rotary_pos_emb, time_step, seq_lengths, src_mask, gather_index
 
 - op : generate_sequence_xpu
   args : (Tensor x, DataType dtype)

paddle/phi/infermeta/fusion.cc

@@ -278,6 +278,7 @@ void FusedMultiTransformerXpuInferMeta(
     const std::vector<const MetaTensor*>& time_step,
     const std::vector<const MetaTensor*>& seq_lengths,
     const std::vector<const MetaTensor*>& src_mask,
+    const std::vector<const MetaTensor*>& gather_index,
     bool pre_layer_norm,
     int rotary_emb_dims,
     float epsilon,
@@ -287,6 +288,7 @@ void FusedMultiTransformerXpuInferMeta(
     const std::string& act_method,
     bool trans_qkvw,
     int ring_id,
+    int gather_axis,
     MetaTensor* out,
     std::vector<MetaTensor*> cache_kv_out) {
   auto x_dim = x.dims();
@@ -325,13 +327,6 @@ void FusedMultiTransformerXpuInferMeta(
                       phi::errors::InvalidArgument(
                           "The first dim of CacheKV must be 2, but got %d",
                           c_dim[0]));  // 2
-    PADDLE_ENFORCE_EQ(
-        c_dim[2],
-        x_dim[0],
-        phi::errors::InvalidArgument("The third dim of CacheKV must be equal "
-                                     "with batch size %d, but got %d",
-                                     x_dim[0],
-                                     c_dim[2]));  // batch_size
     PADDLE_ENFORCE_EQ(
         c_dim[3],
         trans_qkvw ? y_dim[1] : y_dim[2],

paddle/phi/infermeta/fusion.h

@@ -108,6 +108,7 @@ void FusedMultiTransformerXpuInferMeta(
     const std::vector<const MetaTensor*>& time_step,
     const std::vector<const MetaTensor*>& seq_lengths,
     const std::vector<const MetaTensor*>& src_mask,
+    const std::vector<const MetaTensor*>& gather_index,
     bool pre_layer_norm,
     int rotary_emb_dims,
     float epsilon,
@@ -117,6 +118,7 @@ void FusedMultiTransformerXpuInferMeta(
     const std::string& act_method,
     bool trans_qkvw,
     int ring_id,
+    int gather_axis,
     MetaTensor* out,
     std::vector<MetaTensor*> cache_kv_out);
 }  // namespace phi

paddle/phi/kernels/fusion/xpu/fused_multi_transformer_xpu_kernel.cc

@@ -17,6 +17,8 @@
 #include "glog/logging.h"
 
 #include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/core/meta_tensor.h"
+#include "paddle/phi/infermeta/binary.h"
 #include "paddle/phi/kernels/memcpy_kernel.h"
 #ifdef PADDLE_WITH_XPU_XFT
 #include "models/fused_multi_transformer_op.h"
@@ -52,6 +54,7 @@ void FusedMultiTransformerXpuKernel(
     const paddle::optional<DenseTensor>& time_step,
     const paddle::optional<DenseTensor>& seq_lengths,
     const paddle::optional<DenseTensor>& src_mask,
+    const paddle::optional<DenseTensor>& gather_index,
     bool pre_layer_norm,
     int rotary_emb_dims,
     float epsilon,
@@ -61,6 +64,7 @@ void FusedMultiTransformerXpuKernel(
     const std::string& act_method,
     bool trans_qkvw,
     int ring_id,
+    int gather_axis,
     DenseTensor* out,
     std::vector<DenseTensor*> cache_kv_out) {
 #ifdef PADDLE_WITH_XPU_XFT
@@ -160,6 +164,21 @@ void FusedMultiTransformerXpuKernel(
   std::vector<xft::xftTensor<XPUTypeT, 5>> xft_cache_kv;
   std::vector<xft::xftTensor<XPUTypeT, 5>> xft_cache_kv_out;
 
+  // Create a temporary Tensor to store the gather output of cache_kv
+  auto gather_index_t = gather_index.get_ptr();
+  auto cache_kv_dims = cache_kv.get_ptr()->at(0)->dims();
+  auto cache_kv_gather_dims = cache_kv_dims;
+  phi::DenseTensor cache_kv_gather_tensor;
+  if (gather_index_t) {
+    MetaTensor cache_kv_gather_meta(&cache_kv_gather_tensor);
+    phi::GatherInferMeta(*cache_kv.get_ptr()->at(0),
+                         *gather_index_t,
+                         Scalar(gather_axis),
+                         &cache_kv_gather_meta);
+    cache_kv_gather_dims = cache_kv_gather_meta.dims();
+    ctx.template Alloc<T>(&cache_kv_gather_tensor);
+  }
+
   int layers = qkvw.size();
   for (int i = 0; i < layers; ++i) {
     // step1. layer_norm
@@ -211,27 +230,55 @@ void FusedMultiTransformerXpuKernel(
     xft_ffn2_bias.emplace_back(const_cast<float*>(ffn2_bias[i]->data<float>()),
                                std::array<int64_t, 1>{ffn2_bias[i]->dims()[0]});
     // cache kv in
-    if (time_step_value > 0) {
-      auto cachekv_dims = cache_kv.get_ptr()->at(i)->dims();
-      xft_cache_kv.emplace_back(reinterpret_cast<XPUTypeT*>(const_cast<T*>(
-                                    cache_kv.get_ptr()->at(i)->data<T>())),
-                                std::array<int64_t, 5>{cachekv_dims[0],
-                                                       cachekv_dims[1],
-                                                       cachekv_dims[2],
-                                                       cachekv_dims[3],
-                                                       cachekv_dims[4]});
+    auto cache_kv_data = reinterpret_cast<XPUTypeT*>(
+        const_cast<T*>(cache_kv.get_ptr()->at(i)->data<T>()));
+    if (gather_index_t) {
+      const auto& index_type = gather_index_t->dtype();
+      if (index_type == DataType::INT32) {
+        r = xpu::gather<XPUTypeT, int32_t>(
+            ctx.x_context(),
+            cache_kv_data,
+            gather_index_t->data<int32_t>(),
+            reinterpret_cast<XPUTypeT*>(cache_kv_gather_tensor.data<T>()),
+            phi::vectorize<int32_t>(cache_kv_dims),
+            gather_index_t->dims().size() == 0 ? 1 : gather_index_t->dims()[0],
+            gather_axis);
+      } else {
+        r = xpu::gather<XPUTypeT, int64_t>(
+            ctx.x_context(),
+            cache_kv_data,
+            gather_index_t->data<int64_t>(),
+            reinterpret_cast<XPUTypeT*>(cache_kv_gather_tensor.data<T>()),
+            phi::vectorize<int32_t>(cache_kv_dims),
+            gather_index_t->dims().size() == 0 ? 1 : gather_index_t->dims()[0],
+            gather_axis);
+      }
+      PADDLE_ENFORCE_XDNN_SUCCESS(r, "xpu::gather");
+      cache_kv_out[i]->ResizeAndAllocate(cache_kv_gather_dims);
+      r = xpu::copy<XPUTypeT>(
+          ctx.x_context(),
+          reinterpret_cast<XPUTypeT*>(cache_kv_gather_tensor.data<T>()),
+          reinterpret_cast<XPUTypeT*>(ctx.template Alloc<T>(cache_kv_out[i])),
+          cache_kv_out[i]->numel());
+      PADDLE_ENFORCE_XDNN_SUCCESS(r, "xpu::copy");
     }
-    // cache kv out
-    auto cachekv_out_dims = cache_kv_out[i]->dims();
+    cache_kv_data = reinterpret_cast<XPUTypeT*>(
+        const_cast<T*>(cache_kv.get_ptr()->at(i)->data<T>()));
+    xft_cache_kv.emplace_back(cache_kv_data,
+                              std::array<int64_t, 5>{cache_kv_gather_dims[0],
+                                                     cache_kv_gather_dims[1],
+                                                     cache_kv_gather_dims[2],
+                                                     cache_kv_gather_dims[3],
+                                                     cache_kv_gather_dims[4]});
+    // cache kv out direct use cache_kv_data
     xft_cache_kv_out.emplace_back(
-        reinterpret_cast<XPUTypeT*>(ctx.template Alloc<T>(cache_kv_out[i])),
-        std::array<int64_t, 5>{cachekv_out_dims[0],
-                               cachekv_out_dims[1],
-                               cachekv_out_dims[2],
-                               cachekv_out_dims[3],
-                               cachekv_out_dims[4]});
+        cache_kv_data,
+        std::array<int64_t, 5>{cache_kv_gather_dims[0],
+                               cache_kv_gather_dims[1],
+                               cache_kv_gather_dims[2],
+                               cache_kv_gather_dims[3],
+                               cache_kv_gather_dims[4]});
   }
-
   xft::NlpParam param;
   param.num_layer = layers;
   param.n_head = num_head;