Fix pruns prunes (#61383)

co63oc · web-flow · commit 4eb14b61ebd6 · 2024-02-01T10:41:19.000+08:00
diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt
@@ -22,7 +22,7 @@ add_subdirectory(fluid)
 # (2) calling `cc_test()` in each `CMakeLists.txt` will not `exactly` add test, but
 # record all tests and its source files, the action of add tests is defered to HERE.
 # Why doing so? since the target of `libpaddle.so` is mostly the last target, and
-# the tests should be added after that accroding to dependency.
+# the tests should be added after that according to dependency.
 # (3) the tests links dynamic libraries, `libpaddle.so`
 # (4) the tests are generated to the same directory, i.e., `CC_TESTS_DIR` defined above.
 
diff --git a/paddle/fluid/framework/ir/conv_bn_fuse_pass_tester.cc b/paddle/fluid/framework/ir/conv_bn_fuse_pass_tester.cc
@@ -98,7 +98,7 @@ void TestMain(const std::string& conv_type) {
 
 TEST(ConvBNFusePass, conv2d) { TestMain("conv"); }
 
-TEST(ConvBNFusePass, conv2d_tranpose) { TestMain("conv_transpose"); }
+TEST(ConvBNFusePass, conv2d_transpose) { TestMain("conv_transpose"); }
 
 }  // namespace ir
 }  // namespace framework
diff --git a/paddle/fluid/framework/ir/cudnn_placement_pass_tester.cc b/paddle/fluid/framework/ir/cudnn_placement_pass_tester.cc
@@ -120,7 +120,7 @@ TEST(CUDNNPlacementPass, enable_relu_pool) {
 
 TEST(CUDNNPlacementPass, enable_all) {
   // 1 conv2d + 1 pool2d
-  // depthwise_conv2d doesnot have CUDNN kernel.
+  // depthwise_conv2d does not have CUDNN kernel.
   PlacementPassTest().MainTest({}, 2);
 }
 
diff --git a/paddle/fluid/framework/ir/cutlass_teller.h b/paddle/fluid/framework/ir/cutlass_teller.h
@@ -157,7 +157,7 @@ class CutlassTeller {
     }
 
     // To prevent generating too many cutlass code,
-    // we only allow oc and ic is divisable by CUTLASS_NHWC_ALIGNMENT
+    // we only allow oc and ic is divisible by CUTLASS_NHWC_ALIGNMENT
     if (groups == 1) {
       if (oc % CUTLASS_NHWC_ALIGNMENT != 0 ||
           ic % CUTLASS_NHWC_ALIGNMENT != 0) {
diff --git a/paddle/fluid/framework/ir/gpu_cpu_map_matmul_to_mul_pass.h b/paddle/fluid/framework/ir/gpu_cpu_map_matmul_to_mul_pass.h
@@ -59,7 +59,7 @@ class GpuCpuMapMatmulV2ToMulPass : public FusePassBase {
 };
 
 /*
- * Map matmul_v2 to matmul, not supoort broadcast.
+ * Map matmul_v2 to matmul, not support broadcast.
  */
 class GpuCpuMapMatmulV2ToMatmulPass : public FusePassBase {
  public:
diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.h b/paddle/fluid/framework/ir/graph_pattern_detector.h
@@ -2059,12 +2059,12 @@ struct FusionLSTM : public PatternBase {
   // declare op
   PATTERN_DECL_NODE(op);
 
-  // declate inputs
+  // declare inputs
   PATTERN_DECL_NODE(x);
   PATTERN_DECL_NODE(weight_h);
   PATTERN_DECL_NODE(weight_x);
 
-  // decalre outputs
+  // declare outputs
   PATTERN_DECL_NODE(hidden);
   PATTERN_DECL_NODE(cell);
 };
diff --git a/paddle/fluid/framework/ir/layernorm_shift_partition_fuse_pass.h b/paddle/fluid/framework/ir/layernorm_shift_partition_fuse_pass.h
@@ -29,7 +29,7 @@ namespace ir {
 //  reshape2
 //     |
 //  reshape2                            |
-//     |           fuse       layernorm_shift_patition
+//     |           fuse       layernorm_shift_partition
 // transpose2       ->                  |
 //     |                             other_op
 //  reshape2
@@ -48,7 +48,7 @@ namespace ir {
 //    roll
 //     |
 //  reshape2                            |
-//     |           fuse       layernorm_shift_patition
+//     |           fuse       layernorm_shift_partition
 // transpose2       ->                  |
 //     |                             other_op
 //  reshape2
diff --git a/paddle/fluid/framework/ir/lock_free_optimize_pass.cc b/paddle/fluid/framework/ir/lock_free_optimize_pass.cc
@@ -101,7 +101,7 @@ void LockFreeOptimizePass::ApplyImpl(ir::Graph* graph) const {
             PADDLE_ENFORCE_NOT_NULL(
                 forward_op,
                 platform::errors::NotFound(
-                    "Can not find forward op for backword op(%s).",
+                    "Can not find forward op for backward op(%s).",
                     backward_op->Name()));
 
             Node* new_optimizer_node = CreateNewSGDNode(
diff --git a/paddle/fluid/framework/ir/trt_flash_multihead_matmul_fuse_pass.cc b/paddle/fluid/framework/ir/trt_flash_multihead_matmul_fuse_pass.cc
@@ -319,7 +319,7 @@ int TrtFlashMultiHeadMatmulFusePass::BuildFlashFusion(
     // check the scale
     int hidden_out = head_number * head_size;
     if (abs(scale_attr - 1.0f / sqrt(static_cast<float>(head_size))) > 1e-5) {
-      VLOG(3) << "scale of muilthead matmul do not fit the requirement of "
+      VLOG(3) << "scale of multi-head matmul do not fit the requirement of "
                  "flash attention plugin, Stop fusing.";
       return;
     }
diff --git a/paddle/fluid/framework/ir/trt_qk_multihead_matmul_fuse_pass.cc b/paddle/fluid/framework/ir/trt_qk_multihead_matmul_fuse_pass.cc
@@ -310,7 +310,7 @@ int TrtQkMultiHeadMatmulFusePass::BuildQkFusion(Graph* graph,
     int hidden_out = wq_tensor->dims()[1];
     int head_size = hidden_out / head_number;
     if (abs(scale_attr - 1.0f / sqrt(static_cast<float>(head_size))) > 1e-5) {
-      VLOG(3) << "scale of muilthead matmul do not fit the requirement of "
+      VLOG(3) << "scale of multi-head matmul do not fit the requirement of "
                  "qk attention plugin, Stop fusing.";
       return;
     }
diff --git a/paddle/fluid/framework/ir/xpu/multi_encoder_xpu_fuse_pass.cc b/paddle/fluid/framework/ir/xpu/multi_encoder_xpu_fuse_pass.cc
@@ -1252,7 +1252,7 @@ int MultiEncoderXPUFusePass::ApplySingleEncoderXPUFuse(
 static std::vector<Node*> GetSingleEncoders(ir::Graph* graph) {
   std::vector<Node*> single_encoders;
   for (auto* node : graph->Nodes()) {
-    // Find first singld_encoder_xpu
+    // Find first single_encoder_xpu
     if (node->IsVar() || node->Op()->Type() != "single_encoder_xpu") continue;
     bool is_first_encoder = true;
     for (auto* in_node : node->inputs) {
diff --git a/paddle/fluid/framework/ir/yolo_box_fuse_pass.h b/paddle/fluid/framework/ir/yolo_box_fuse_pass.h
@@ -33,7 +33,7 @@ transpose-|   transpose-|   transpose-|
    |------|-----|-------|------|      |
           |   concat    |             |
           |-----|-------|-------------|
-                |     cocnat
+                |     concat
                 |-------|
                        nms3
 
diff --git a/paddle/fluid/operators/array_to_lod_tensor_op.cc b/paddle/fluid/operators/array_to_lod_tensor_op.cc
@@ -119,7 +119,7 @@ class ArrayToLoDTensorOp : public framework::OperatorBase {
           platform::errors::InvalidArgument(
               "The dimension of the %zu'th element in LoDTensorArray "
               "differs from previous ones."
-              "The current dimension is %d, and the previous dimesion is %d.",
+              "The current dimension is %d, and the previous dimension is %d.",
               i,
               ins_i_dims,
               ins_dims));
diff --git a/paddle/fluid/operators/detection/locality_aware_nms_op.cc b/paddle/fluid/operators/detection/locality_aware_nms_op.cc
@@ -489,10 +489,10 @@ Firstly, this operator merge box and score according their IOU(intersection over
 In the NMS step, this operator greedily selects a subset of detection bounding
 boxes that have high scores larger than score_threshold, if providing this
 threshold, then selects the largest nms_top_k confidences scores if nms_top_k
-is larger than -1. Then this operator pruns away boxes that have high IOU
+is larger than -1. Then this operator prunes away boxes that have high IOU
 (intersection over union) overlap with already selected boxes by adaptive
 threshold NMS based on parameters of nms_threshold and nms_eta.
-Aftern NMS step, at most keep_top_k number of total bboxes are to be kept
+After NMS step, at most keep_top_k number of total bboxes are to be kept
 per image if keep_top_k is larger than -1.
 This operator support multi-class and batched inputs. It applying NMS
 independently for each class. The outputs is a 2-D LoDTensor, for each
diff --git a/paddle/fluid/operators/detection/multiclass_nms_op.cc b/paddle/fluid/operators/detection/multiclass_nms_op.cc
@@ -556,10 +556,10 @@ of boxes and scores.
 In the NMS step, this operator greedily selects a subset of detection bounding
 boxes that have high scores larger than score_threshold, if providing this
 threshold, then selects the largest nms_top_k confidences scores if nms_top_k
-is larger than -1. Then this operator pruns away boxes that have high IOU
+is larger than -1. Then this operator prunes away boxes that have high IOU
 (intersection over union) overlap with already selected boxes by adaptive
 threshold NMS based on parameters of nms_threshold and nms_eta.
-Aftern NMS step, at most keep_top_k number of total bboxes are to be kept
+After NMS step, at most keep_top_k number of total bboxes are to be kept
 per image if keep_top_k is larger than -1.
 This operator support multi-class and batched inputs. It applying NMS
 independently for each class. The outputs is a 2-D LoDTensor, for each
diff --git a/paddle/fluid/operators/detection/prior_box_op.h b/paddle/fluid/operators/detection/prior_box_op.h
@@ -25,25 +25,25 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-inline void ExpandAspectRatios(const std::vector<float>& input_aspect_ratior,
+inline void ExpandAspectRatios(const std::vector<float>& input_aspect_ratio,
                                bool flip,
-                               std::vector<float>* output_aspect_ratior) {
+                               std::vector<float>* output_aspect_ratio) {
   constexpr float epsilon = 1e-6;
-  output_aspect_ratior->clear();
-  output_aspect_ratior->push_back(1.0f);
-  for (size_t i = 0; i < input_aspect_ratior.size(); ++i) {
-    float ar = input_aspect_ratior[i];
+  output_aspect_ratio->clear();
+  output_aspect_ratio->push_back(1.0f);
+  for (size_t i = 0; i < input_aspect_ratio.size(); ++i) {
+    float ar = input_aspect_ratio[i];
     bool already_exist = false;
-    for (size_t j = 0; j < output_aspect_ratior->size(); ++j) {
-      if (fabs(ar - output_aspect_ratior->at(j)) < epsilon) {
+    for (size_t j = 0; j < output_aspect_ratio->size(); ++j) {
+      if (fabs(ar - output_aspect_ratio->at(j)) < epsilon) {
         already_exist = true;
         break;
       }
     }
     if (!already_exist) {
-      output_aspect_ratior->push_back(ar);
+      output_aspect_ratio->push_back(ar);
       if (flip) {
-        output_aspect_ratior->push_back(1.0f / ar);
+        output_aspect_ratio->push_back(1.0f / ar);
       }
     }
   }
diff --git a/paddle/fluid/operators/dlnne/dlnne_engine_op.cc b/paddle/fluid/operators/dlnne/dlnne_engine_op.cc
@@ -147,9 +147,9 @@ class DlnneEngineOpMaker : public framework::OpProtoAndCheckerMaker {
                          "'01', '23', '0123' ");
     // when use_calib_mode is true and enable_int8 is true,
     // the calibration_runtime start,
-    // when calibration_mode is true, the calibration_runtiime
+    // when calibration_mode is true, the calibration_runtime
     // go to the first stage of calibration, and when finish
-    // fisrt stage, the calibration_mode is set false, the
+    // first stage, the calibration_mode is set false, the
     // calibration_runtime go to the second stage
     AddAttr<bool>("use_calib_mode", "dlnne use calib mode");
     AddAttr<bool>("enable_int8", "dlnne enable int8");
diff --git a/paddle/fluid/operators/fake_dequantize_op.cc b/paddle/fluid/operators/fake_dequantize_op.cc
@@ -300,7 +300,7 @@ PD_REGISTER_STRUCT_KERNEL(fake_channel_wise_dequantize_max_abs,
 REGISTER_OP_VERSION(fake_channel_wise_dequantize_max_abs)
     .AddCheckpoint(
         R"ROC(add new attributes [quant_axis] for applying per-channel "
-        "dequantization to conv2d_tranpose and mul ops.)ROC",
+        "dequantization to conv2d_transpose and mul ops.)ROC",
         paddle::framework::compatible::OpVersionDesc().NewAttr(
             "quant_axis", "The axis for dequantization.", 0))
     .AddCheckpoint(
diff --git a/paddle/fluid/operators/fake_quantize_op.cc b/paddle/fluid/operators/fake_quantize_op.cc
@@ -979,7 +979,7 @@ PD_REGISTER_STRUCT_KERNEL(fake_channel_wise_quantize_dequantize_abs_max,
 REGISTER_OP_VERSION(fake_channel_wise_quantize_abs_max)
     .AddCheckpoint(
         R"ROC(add new attributes [quant_axis] for applying per-channel "
-        "quantization to conv2d_tranpose and mul ops.)ROC",
+        "quantization to conv2d_transpose and mul ops.)ROC",
         paddle::framework::compatible::OpVersionDesc().NewAttr(
             "quant_axis", "The axis for quantization.", 0));
 REGISTER_OP_VERSION(moving_average_abs_max_scale)
diff --git a/paddle/fluid/operators/fused/fused_transformer_op.h b/paddle/fluid/operators/fused/fused_transformer_op.h
@@ -54,7 +54,7 @@ class FusedAttention {
   void ComputeBackward();
 
  private:
-  FusedMHA<T>* fmha;  // fused multihead attention
+  FusedMHA<T>* fmha;  // fused multi-head attention
 
   int batch_size;
   int max_seq_len;
diff --git a/paddle/fluid/operators/fused_token_prune_op.cc b/paddle/fluid/operators/fused_token_prune_op.cc
@@ -110,19 +110,19 @@ class FusedTokenPruneOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(mask_dim.size(),
                       4,
                       platform::errors::InvalidArgument(
-                          "The input mask must be 4-dimention"));
+                          "The input mask must be 4-dimension"));
     PADDLE_ENFORCE_EQ(attn_dim.size(),
                       4,
                       platform::errors::InvalidArgument(
-                          "The input attn must be 4-dimention"));
+                          "The input attn must be 4-dimension"));
     PADDLE_ENFORCE_EQ(
         x_dim.size(),
         3,
-        platform::errors::InvalidArgument("The input x must be 4-dimention"));
+        platform::errors::InvalidArgument("The input x must be 4-dimension"));
     PADDLE_ENFORCE_EQ(new_mask_dim.size(),
                       4,
                       platform::errors::InvalidArgument(
-                          "The input attn must be 4-dimention"));
+                          "The input attn must be 4-dimension"));
 
     // check input dims relations
     PADDLE_ENFORCE_EQ(mask_dim[0],
diff --git a/paddle/fluid/operators/gru_op.cc b/paddle/fluid/operators/gru_op.cc
@@ -135,7 +135,7 @@ class GRUOpMaker : public framework::OpProtoAndCheckerMaker {
         "the update gate and reset gate with shape (D x 2D), and the second "
         "part are weights of output candidate with shape (D x D).");
     AddInput("Bias",
-             "(Tensor, optional) Bias vector with shape (1 x 3D) concating "
+             "(Tensor, optional) Bias vector with shape (1 x 3D) concatenating "
              "bias of the update gate, reset gate and output candidate.")
         .AsDispensable();
     AddOutput(
@@ -407,7 +407,7 @@ class GRUCPUKernel : public framework::OpKernel<T> {
       PADDLE_ENFORCE_NOT_NULL(
           packed_gate,
           platform::errors::NotFound(
-              "The caculation result of packed_gate by "
+              "The calculation result of packed_gate by "
               "GEMM_ALLOC should not be null when using MKL."));
       blas.GEMM_PACK(CblasBMatrix,
                      CblasNoTrans,
@@ -425,7 +425,7 @@ class GRUCPUKernel : public framework::OpKernel<T> {
       PADDLE_ENFORCE_NOT_NULL(
           packed_state,
           platform::errors::NotFound(
-              "The caculation result of packed_state by "
+              "The calculation result of packed_state by "
               "GEMM_ALLOC should not be null when using MKL."));
       blas.GEMM_PACK(CblasBMatrix,
                      CblasNoTrans,
diff --git a/paddle/fluid/operators/im2sequence_op.cc b/paddle/fluid/operators/im2sequence_op.cc
@@ -37,14 +37,14 @@ class Im2SequenceOp : public framework::OperatorWithKernel {
                           "The output 'Out' of Im2SequenceOp is not found."));
     auto in_dim = ctx->GetInputDim("X");
 
-    PADDLE_ENFORCE_EQ(
-        in_dim.size(),
-        4,
-        platform::errors::InvalidArgument(
-            "The dimesions size of input 'X' in Im2SequenceOp should be 4. But "
-            "received dimesions size=[%d], dimesions=[%s].",
-            in_dim.size(),
-            in_dim));
+    PADDLE_ENFORCE_EQ(in_dim.size(),
+                      4,
+                      platform::errors::InvalidArgument(
+                          "The dimensions size of input 'X' in Im2SequenceOp "
+                          "should be 4. But "
+                          "received dimensions size=[%d], dimensions=[%s].",
+                          in_dim.size(),
+                          in_dim));
     auto img_channels = in_dim[1];
 
     auto kernels = ctx->Attrs().Get<std::vector<int>>("kernels");
@@ -91,7 +91,7 @@ class Im2SequenceOpMaker : public framework::OpProtoAndCheckerMaker {
                               "the attribute is valid only when input(Y)"
                               "is not NULL.this attribute represents the"
                               "scaling of the pic through the CNN"
-                              "(vector<int> dedault:{1,1}),the out_stride"
+                              "(vector<int> default:{1,1}),the out_stride"
                               " (out_stride_height, out_stride_width)")
         .SetDefault({1, 1});
     AddComment(R"DOC(
diff --git a/paddle/fluid/operators/interpolate_op.cc b/paddle/fluid/operators/interpolate_op.cc
@@ -39,7 +39,7 @@ static void Interpolate1DInferShapeCheck(framework::InferShapeContext* ctx) {
       common::StringToDataLayout(ctx->Attrs().Get<std::string>("data_layout"));
 
   if (ctx->HasInputs("SizeTensor")) {
-    // top prority size
+    // top priority size
     auto inputs_name = ctx->Inputs("SizeTensor");
     PADDLE_ENFORCE_EQ(
         inputs_name.size(),
@@ -91,7 +91,7 @@ static void Interpolate1DInferShapeCheck(framework::InferShapeContext* ctx) {
         out_size_dim.size(),
         1,
         platform::errors::InvalidArgument(
-            "OutSize's dimension size must be 1, but got dimention = %d .",
+            "OutSize's dimension size must be 1, but got dimension = %d .",
             out_size_dim.size()));
     PADDLE_ENFORCE_EQ(
         out_size_dim[0],
@@ -128,7 +128,7 @@ static void Interpolate2DInferShapeCheck(framework::InferShapeContext* ctx) {
       common::StringToDataLayout(ctx->Attrs().Get<std::string>("data_layout"));
 
   if (ctx->HasInputs("SizeTensor")) {
-    // top prority size
+    // top priority size
     auto inputs_name = ctx->Inputs("SizeTensor");
     PADDLE_ENFORCE_EQ(
         inputs_name.size(),
@@ -223,7 +223,7 @@ static void Interpolate3DInferShapeCheck(framework::InferShapeContext* ctx) {
       common::StringToDataLayout(ctx->Attrs().Get<std::string>("data_layout"));
 
   if (ctx->HasInputs("SizeTensor")) {
-    // top prority size
+    // top priority size
     auto inputs_name = ctx->Inputs("SizeTensor");
     PADDLE_ENFORCE_EQ(
         inputs_name.size(),
diff --git a/paddle/fluid/operators/partial_sum_op.cc b/paddle/fluid/operators/partial_sum_op.cc
@@ -48,15 +48,15 @@ class PartialSumOp : public framework::OperatorWithKernel {
     int start_index = ctx->Attrs().Get<int>("start_index");
     int length = ctx->Attrs().Get<int>("length");
 
-    // Only suppert two dimensions now, should be extended later
+    // Only support two dimensions now, should be extended later
     // when length is -1, need make sure all dimensions to be added are the same
     int64_t batch_size = -1;
     int64_t input_len = -1;
     for (size_t i = 0; i < inputs_num; ++i) {
       PADDLE_ENFORCE_EQ(inputs_dims[i].size(),
                         2,
                         platform::errors::InvalidArgument(
-                            "Only suppert two dimensions input now."));
+                            "Only support two dimensions input now."));
       if (i == 0) {
         batch_size = inputs_dims[0][0];
         input_len = inputs_dims[0][1];
diff --git a/paddle/fluid/operators/reduce_ops/reduce_op.h b/paddle/fluid/operators/reduce_ops/reduce_op.h
diff --git a/paddle/fluid/operators/reshape_op.cc b/paddle/fluid/operators/reshape_op.cc
diff --git a/paddle/fluid/operators/sequence_ops/sequence_expand_op.cc b/paddle/fluid/operators/sequence_ops/sequence_expand_op.cc
diff --git a/paddle/fluid/operators/sequence_ops/sequence_pad_op.cc b/paddle/fluid/operators/sequence_ops/sequence_pad_op.cc

Original file line number	Diff line number	Diff line change
`@@ -120,7 +120,7 @@ TEST(CUDNNPlacementPass, enable_relu_pool) {`
`120`	`120`
`121`	`121`	`TEST(CUDNNPlacementPass, enable_all) {`
`122`	`122`	`// 1 conv2d + 1 pool2d`
`123`		`- // depthwise_conv2d doesnot have CUDNN kernel.`
	`123`	`+ // depthwise_conv2d does not have CUDNN kernel.`
`124`	`124`	`PlacementPassTest().MainTest({}, 2);`
`125`	`125`	`}`
`126`	`126`
Original file line number	Diff line number	Diff line change
`@@ -157,7 +157,7 @@ class CutlassTeller {`
`157`	`157`	`}`
`158`	`158`
`159`	`159`	`// To prevent generating too many cutlass code,`
`160`		`- // we only allow oc and ic is divisable by CUTLASS_NHWC_ALIGNMENT`
	`160`	`+ // we only allow oc and ic is divisible by CUTLASS_NHWC_ALIGNMENT`
`161`	`161`	`if (groups == 1) {`
`162`	`162`	`if (oc % CUTLASS_NHWC_ALIGNMENT != 0 \|\|`
`163`	`163`	`ic % CUTLASS_NHWC_ALIGNMENT != 0) {`
Original file line number	Diff line number	Diff line change
`@@ -319,7 +319,7 @@ int TrtFlashMultiHeadMatmulFusePass::BuildFlashFusion(`
`319`	`319`	`// check the scale`
`320`	`320`	`int hidden_out = head_number * head_size;`
`321`	`321`	`if (abs(scale_attr - 1.0f / sqrt(static_cast<float>(head_size))) > 1e-5) {`
`322`		`- VLOG(3) << "scale of muilthead matmul do not fit the requirement of "`
	`322`	`+ VLOG(3) << "scale of multi-head matmul do not fit the requirement of "`
`323`	`323`	`"flash attention plugin, Stop fusing.";`
`324`	`324`	`return;`
`325`	`325`	`}`
Original file line number	Diff line number	Diff line change
`@@ -310,7 +310,7 @@ int TrtQkMultiHeadMatmulFusePass::BuildQkFusion(Graph* graph,`
`310`	`310`	`int hidden_out = wq_tensor->dims()[1];`
`311`	`311`	`int head_size = hidden_out / head_number;`
`312`	`312`	`if (abs(scale_attr - 1.0f / sqrt(static_cast<float>(head_size))) > 1e-5) {`
`313`		`- VLOG(3) << "scale of muilthead matmul do not fit the requirement of "`
	`313`	`+ VLOG(3) << "scale of multi-head matmul do not fit the requirement of "`
`314`	`314`	`"qk attention plugin, Stop fusing.";`
`315`	`315`	`return;`
`316`	`316`	`}`