PaddlePaddle
diff --git a/‎paddle/fluid/operators/math/beam_search.h
Copy file name to clipboard
+1-1Lines changed: 1 addition & 1 deletion b/‎paddle/fluid/operators/math/beam_search.h
Copy file name to clipboard
+1-1Lines changed: 1 addition & 1 deletion
diff --git a/‎paddle/phi/infermeta/spmd_rules/default_data_parallel.cc
Copy file name to clipboard
+5-5Lines changed: 5 additions & 5 deletions b/‎paddle/phi/infermeta/spmd_rules/default_data_parallel.cc
Copy file name to clipboard
+5-5Lines changed: 5 additions & 5 deletions
diff --git a/‎paddle/phi/infermeta/spmd_rules/default_data_parallel.h
Copy file name to clipboard
+2-2Lines changed: 2 additions & 2 deletions b/‎paddle/phi/infermeta/spmd_rules/default_data_parallel.h
Copy file name to clipboard
+2-2Lines changed: 2 additions & 2 deletions
diff --git a/‎paddle/phi/infermeta/spmd_rules/dim_trans.h
Copy file name to clipboard
+3-3Lines changed: 3 additions & 3 deletions b/‎paddle/phi/infermeta/spmd_rules/dim_trans.h
Copy file name to clipboard
+3-3Lines changed: 3 additions & 3 deletions
diff --git a/‎paddle/phi/infermeta/spmd_rules/elementwise.cc
Copy file name to clipboard
+8-8Lines changed: 8 additions & 8 deletions b/‎paddle/phi/infermeta/spmd_rules/elementwise.cc
Copy file name to clipboard
+8-8Lines changed: 8 additions & 8 deletions
diff --git a/‎paddle/phi/infermeta/spmd_rules/embedding.cc
Copy file name to clipboard
+4-4Lines changed: 4 additions & 4 deletions b/‎paddle/phi/infermeta/spmd_rules/embedding.cc
Copy file name to clipboard
+4-4Lines changed: 4 additions & 4 deletions
diff --git a/‎paddle/phi/infermeta/spmd_rules/fused_rope.cc
Copy file name to clipboard
+3-3Lines changed: 3 additions & 3 deletions b/‎paddle/phi/infermeta/spmd_rules/fused_rope.cc
Copy file name to clipboard
+3-3Lines changed: 3 additions & 3 deletions
diff --git a/‎paddle/phi/infermeta/spmd_rules/layer_norm.cc
Copy file name to clipboard
+11-11Lines changed: 11 additions & 11 deletions b/‎paddle/phi/infermeta/spmd_rules/layer_norm.cc
Copy file name to clipboard
+11-11Lines changed: 11 additions & 11 deletions
diff --git a/‎paddle/phi/infermeta/spmd_rules/matmul.cc
Copy file name to clipboard
+2-2Lines changed: 2 additions & 2 deletions b/‎paddle/phi/infermeta/spmd_rules/matmul.cc
Copy file name to clipboard
+2-2Lines changed: 2 additions & 2 deletions
diff --git a/‎paddle/phi/infermeta/spmd_rules/reduction.cc
Copy file name to clipboard
+4-4Lines changed: 4 additions & 4 deletions b/‎paddle/phi/infermeta/spmd_rules/reduction.cc
Copy file name to clipboard
+4-4Lines changed: 4 additions & 4 deletions
@@ -88,7 +88,7 @@ class BeamSearchFunctor {
    *
    * @selected_ids: a [None, 1]-shaped tensor with LoD.
    *   In a machine translation model, it might be the candidate term id sets,
-   *   each set stored as a varience-length sequence.
+   *   each set stored as a variance-length sequence.
    *   The format might be described with a two-level LoD
    *   - [[0 1],
    *      [0 1 2]]
 
@@ -26,7 +26,7 @@ namespace distributed {
 using phi::distributed::auto_parallel::str_join;
 
 ////////////////// Utils Functions //////////////////
-std::vector<int64_t> GetDefaultDataParallelDimsmapping(
+std::vector<int64_t> GetDefaultDataParallelDimsMapping(
     const int64_t batch_axis_dim, const int ndim) {
   std::vector<int64_t> dims_mapping(ndim, -1);
   dims_mapping[0] = batch_axis_dim;
@@ -61,7 +61,7 @@ SpmdInfo DefaultDataParallelInferSpmd(
     TensorDistAttr dist_attr_dst =
         CopyTensorDistAttrForOutput(ins[0]->dist_attr());
     std::vector<int64_t> dst_dims_maping =
-        GetDefaultDataParallelDimsmapping(batch_axis_dim, ndim);
+        GetDefaultDataParallelDimsMapping(batch_axis_dim, ndim);
     dist_attr_dst.set_dims_mapping(dst_dims_maping);
     output_dist_attrs.emplace_back(dist_attr_dst);
   }
@@ -73,7 +73,7 @@ SpmdInfo DefaultDataParallelInferSpmd(
     TensorDistAttr dist_attr_dst =
         CopyTensorDistAttrForOutput(ins[i]->dist_attr());
     std::vector<int64_t> dst_dims_maping =
-        GetDefaultDataParallelDimsmapping(batch_axis_dim, ndim);
+        GetDefaultDataParallelDimsMapping(batch_axis_dim, ndim);
     dist_attr_dst.set_dims_mapping(dst_dims_maping);
     dst_input_dist_attrs.emplace_back(dist_attr_dst);
   }
@@ -124,7 +124,7 @@ SpmdInfo DefaultDataParallelInferSpmdReverse(
     TensorDistAttr dist_attr_dst =
         CopyTensorDistAttrForOutput(outs[i]->dist_attr());
     std::vector<int64_t> dst_dims_maping =
-        GetDefaultDataParallelDimsmapping(batch_axis_dim, ndim);
+        GetDefaultDataParallelDimsMapping(batch_axis_dim, ndim);
     dist_attr_dst.set_dims_mapping(dst_dims_maping);
     output_dist_attrs.emplace_back(dist_attr_dst);
   }
@@ -136,7 +136,7 @@ SpmdInfo DefaultDataParallelInferSpmdReverse(
     TensorDistAttr dist_attr_dst =
         CopyTensorDistAttrForOutput(ins[i]->dist_attr());
     std::vector<int64_t> dst_dims_maping =
-        GetDefaultDataParallelDimsmapping(batch_axis_dim, ndim);
+        GetDefaultDataParallelDimsMapping(batch_axis_dim, ndim);
     dist_attr_dst.set_dims_mapping(dst_dims_maping);
     dst_input_dist_attrs.emplace_back(dist_attr_dst);
   }
 
@@ -23,7 +23,7 @@ limitations under the License. */
 namespace phi {
 namespace distributed {
 /**
- * A **hack** rule with a strong assumtion that the first dimension of
+ * A **hack** rule with a strong assumption that the first dimension of
  * all the input and ouput tensors is the batch dimension (broadcast dimension),
  * therefore, if any tensor's first dimension is sharded, the sharding would be
  * propagating to all the other tensors (for tensor first dimension). All the
@@ -34,7 +34,7 @@ namespace distributed {
  * once there is a specific rule for that op,  we should remove that op from
  * this rule.
  *
- * Vector of input tensors and output tensors used as argumnets (for both
+ * Vector of input tensors and output tensors used as arguments (for both
  * inferfw & inferbw) to support any kind of op.
  *
  */
 
@@ -31,12 +31,12 @@ namespace distributed {
 // transformed from the input shape. Each element in vector<DimTrans*>
 // describes the transformation of one output axis. For example, when
 // a reshape operator reshapes a tensor from the shape of (6, 12, 48)
-// to (72, 6, 8), this transfromation can be described as:
+// to (72, 6, 8), this transformation can be described as:
 // [Flatten(Dim(0), Dim(1)), Split(Dim(2), (6,8), 0), Split(Dim(2), (6,8), 1)]
 // meaning that dim0 in output is flattened from dim0 and dim1 in input,
 // dim1 and dim2 in output are obtained by splitting dim2 in input, the
-// splitted shape is (6, 8), dim1 referes to the first shape value in (6, 8)
-// and dim2 referes to the second shape value in (6, 8).
+// splitted shape is (6, 8), dim1 refers to the first shape value in (6, 8)
+// and dim2 refers to the second shape value in (6, 8).
 class DimTrans {
  public:
   enum class Type { INPUTDIM, SINGLETON, FLATTEN, SPLIT };
 
@@ -98,7 +98,7 @@ SpmdInfo ElementwiseUnaryInferSpmd(const DistMetaTensor& x) {
   std::string x_axes = GetBroadcastAxes(x_ndim, x_ndim, alphabet);
   std::string out_axes = x_axes;
 
-  // Step2: Sharding Propogation
+  // Step2: Sharding Propagation
   // Step2.1: Merge input shardings
   std::pair<std::string, std::vector<int64_t>> axes_sharding_info(
       x_axes, x_dims_mapping);
@@ -125,7 +125,7 @@ SpmdInfo ElementwiseUnaryInferSpmd(const DistMetaTensor& x) {
 }
 
 // NOTE(lizhiyu): This function is only for `cast` right now to support partial
-// propagatation
+// propagation
 SpmdInfo ElementwiseUnaryWithPartialInferSpmd(const DistMetaTensor& x) {
   // Step0: Verify Input Args Based on Elementwise Logic
   auto x_shape = common::vectorize(x.dims());
@@ -145,7 +145,7 @@ SpmdInfo ElementwiseUnaryWithPartialInferSpmd(const DistMetaTensor& x) {
   std::string x_axes = GetBroadcastAxes(x_ndim, x_ndim, alphabet);
   std::string out_axes = x_axes;
 
-  // Step2: Sharding Propogation
+  // Step2: Sharding Propagation
   // Step2.1: Merge input shardings
   std::pair<std::string, std::vector<int64_t>> axes_sharding_info(
       x_axes, x_dims_mapping);
@@ -204,7 +204,7 @@ SpmdInfo ElementwiseUnaryInferSpmdReverse(const DistMetaTensor& x,
   std::string x_axes = GetBroadcastAxes(x_ndim, x_ndim, alphabet);
   std::string out_axes = x_axes;
 
-  // Step2: Sharding Propogation
+  // Step2: Sharding Propagation
   // Step2.1: Merge output shardings
   std::pair<std::string, std::vector<int64_t>> axes_sharding_info(
       out_axes, out_dims_mapping);
@@ -259,13 +259,13 @@ SpmdInfo ElementwiseBinaryInferSpmd(const DistMetaTensor& x,
   std::string x_axes, y_axes, out_axes;
   GetBinaryNotations(x_shape, y_shape, &x_axes, &y_axes, &out_axes);
 
-  // Step2: Sharding Propogation
+  // Step2: Sharding Propagation
   // Step2.1: Merge input shardings
   std::unordered_map<std::string, int64_t> axis_to_dim_map =
       ShardingMergeForTensors(
           {{x_axes, x_dims_mapping}, {y_axes, y_dims_mapping}});
 
-  // Step2.2: Infer output dimsmapping from merged input dimsmapping
+  // Step2.2: Infer output dims mapping from merged input dims mapping
   std::vector<int64_t> out_dims_mapping =
       GetDimsMappingForAxes(out_axes, axis_to_dim_map);
 
@@ -332,7 +332,7 @@ SpmdInfo ElementwiseBinaryInferSpmdReverse(const DistMetaTensor& x,
   std::string x_axes, y_axes, out_axes;
   GetBinaryNotations(x_shape, y_shape, &x_axes, &y_axes, &out_axes);
 
-  // Step2: Sharding Propogation
+  // Step2: Sharding Propagation
   // Step2.1: Merge output shardings
   std::unordered_map<std::string, int64_t> axis_to_dim_map =
       ShardingMergeForTensors({{out_axes, out_dims_mapping}});
@@ -436,7 +436,7 @@ SpmdInfo ElementwiseBinaryGradInferSpmd(const DistMetaTensor& x,
                                    "the rank of input as [%d].",
                                    out_grad.dims().size(),
                                    y.dims().size()));
-  // The backward rule of elementwise follows the princple: the dist_attr
+  // The backward rule of elementwise follows the principle: the dist_attr
   // of input should equal to out_grad.
   // Caution the special case when the inputs calculate together with different
   // shape it means one of the input is broadcast to same shape with the other
 
@@ -123,7 +123,7 @@ SpmdInfo EmbeddingInferSpmd(const DistMetaTensor& x,
   std::string weight_axes = "jk";
   std::string out_axes = x_axes + "k";
 
-  // Step2: Sharding Propogation
+  // Step2: Sharding Propagation
   // Step2.1: merge input shardings
   auto axis_to_dim_map = ShardingMergeForTensors(
       {{x_axes, x_dims_mapping}, {weight_axes, weight_dims_mapping}}, false);
@@ -196,7 +196,7 @@ SpmdInfo EmbeddingInferSpmdReverse(const DistMetaTensor& x,
   std::string weight_axes = "jk";
   std::string out_axes = x_axes + "k";
 
-  // step2: Sharding Propogation
+  // step2: Sharding Propagation
   // should not use input dims mapping for backward sharding merge
   auto axis_to_dim_map =
       ShardingMergeForTensors({{out_axes, out_dims_mapping}}, false);
@@ -280,7 +280,7 @@ SpmdInfo EmbeddingGradInferSpmd(const DistMetaTensor& x,
   std::string out_grad_dst_axes = t0_axes.substr(0, t0_axes.length() - 1) + "k";
   std::string w_grad_axes = t0_axes.substr(t0_axes.length() - 1, 1) + "k";
 
-  // Step2.2: Sharding Propogation
+  // Step2.2: Sharding Propagation
   // Step2.2.1: merge input shardings
   auto axis_to_dim_map = ShardingMergeForTensors(
       {{t0_axes, t0.dist_attr().dims_mapping()},
@@ -308,7 +308,7 @@ SpmdInfo EmbeddingGradInferSpmd(const DistMetaTensor& x,
   w_grad_dist_attr.set_partial_status(partial_on_dims);
 
   // Step2.3: Update inputs info.
-  // NOTE: Reshard happend on intemediate operators must be ensure propagated
+  // NOTE: Reshard happened on intermediate operators must be ensure propagated
   // back to first inputs.
   t0 = DistMetaTensor(t0.dims(), t0_dist_attr);
   const auto& t0_dims = t0.dist_attr().dims_mapping();
 
@@ -27,7 +27,7 @@ namespace distributed {
 using auto_parallel::str_join;
 const int kBatchDimIndex = 0;
 const int kSeqlenDimIndex = 1;
-const int kNumheadsDimIndex = 2;
+const int kNumHeadsDimIndex = 2;
 const int kHeadDimIndex = 3;
 
 void check_q(const DistMetaTensor& q) {
@@ -107,12 +107,12 @@ void check_sin_cos(const DistMetaTensor& sin,
   int head_dim_index = ndim == 2 ? 1 : 3;
   if (ndim == 4) {
     PADDLE_ENFORCE_EQ(
-        (shape[kBatchDimIndex] == 1 && shape[kNumheadsDimIndex] == 1),
+        (shape[kBatchDimIndex] == 1 && shape[kNumHeadsDimIndex] == 1),
         true,
         phi::errors::InvalidArgument("The batch_size and num_heads of sin/cos "
                                      "must be 1, but given [%d], [%d]",
                                      shape[kBatchDimIndex],
-                                     shape[kNumheadsDimIndex]));
+                                     shape[kNumHeadsDimIndex]));
   }
 
   const std::vector<int64_t> position_ids_shape =
 
@@ -82,7 +82,7 @@ SpmdInfo LayerNormInferSpmd(const DistMetaTensor& x,
   // get output notation
   std::string out_axes = x_axes;
 
-  // Step2: Sharding Propogation
+  // Step2: Sharding Propagation
   // Step2.1: merge input sharding
   // As the mean and variance in outputs are `flattened` from
   // x[0:begin_norm_axis], only the first axis can be sharded,
@@ -94,13 +94,13 @@ SpmdInfo LayerNormInferSpmd(const DistMetaTensor& x,
   // Step2.2: infer output dims mapping
   TensorDistAttr out_dist_attr = CopyTensorDistAttrForOutput(x_dist_attr_src);
   TensorDistAttr mean_dist_attr = CopyTensorDistAttrForOutput(x_dist_attr_src);
-  TensorDistAttr varience_dist_attr =
+  TensorDistAttr variance_dist_attr =
       CopyTensorDistAttrForOutput(x_dist_attr_src);
   out_dist_attr.set_dims_mapping(
       GetDimsMappingForAxes(out_axes, axis_to_dim_map));
   mean_dist_attr.set_dims_mapping(
       GetDimsMappingForAxes(mean_axes, axis_to_dim_map));
-  varience_dist_attr.set_dims_mapping(
+  variance_dist_attr.set_dims_mapping(
       GetDimsMappingForAxes(variance_axes, axis_to_dim_map));
 
   // Step2.3: update input dims mapping
@@ -110,7 +110,7 @@ SpmdInfo LayerNormInferSpmd(const DistMetaTensor& x,
   TensorDistAttr bias_dist_attr_dst =
       CopyTensorDistAttrForOutput(bias.dist_attr());
   x_dist_attr_dst.set_dims_mapping(x_dims_mapping);
-  // TODO(zhiqiu): support shardding on scale and bias
+  // TODO(zhiqiu): support sharding on scale and bias
   // Now, apply replicating.
   scale_dist_attr_dst.set_dims_mapping({-1});
   bias_dist_attr_dst.set_dims_mapping({-1});
@@ -142,11 +142,11 @@ SpmdInfo LayerNormInferSpmd(const DistMetaTensor& x,
   VLOG(4) << "Mean dims mapping: [" << str_join(mean_dist_attr.dims_mapping())
           << "]";
   VLOG(4) << "Variance dims mapping: ["
-          << str_join(varience_dist_attr.dims_mapping()) << "]";
+          << str_join(variance_dist_attr.dims_mapping()) << "]";
   VLOG(4) << std::endl;
 
   return {{x_dist_attr_dst, scale_dist_attr_dst, bias_dist_attr_dst},
-          {out_dist_attr, mean_dist_attr, varience_dist_attr}};
+          {out_dist_attr, mean_dist_attr, variance_dist_attr}};
 }
 
 SpmdInfo LayerNormInferSpmdReverse(const DistMetaTensor& x,
@@ -215,7 +215,7 @@ SpmdInfo LayerNormInferSpmdReverse(const DistMetaTensor& x,
   std::string bias_axes(1, x_axes[x_ndim - 1]);
   std::string out_axes = x_axes;
 
-  // Step2: Sharding Propogation
+  // Step2: Sharding Propagation
   // For the axes after norm_axis in both input and output tensors,
   // set their dims mappings to -1. For the other axes, set input
   // tensor's dims mapping the same as output tensor's dims mapping.
@@ -356,12 +356,12 @@ SpmdInfo LayerNormGradInferSpmd(const DistMetaTensor& x,
   if (begin_norm_axis > 0) {
     std::vector<std::vector<int64_t>> shapes = {
         x_shape, mean_shape, variance_shape, x_shape};
-    std::vector<std::string> anotations;
-    std::string align_anotation;
-    std::tie(anotations, align_anotation) =
+    std::vector<std::string> annotations;
+    std::string align_annotation;
+    std::tie(annotations, align_annotation) =
         BuildLayerNormGradEinsum(x_shape.size(), begin_norm_axis);
     AlignDimsSharding(
-        &dist_attrs, shapes, anotations, {}, align_anotation, false);
+        &dist_attrs, shapes, annotations, {}, align_annotation, false);
     x_dist_attr = std::move(dist_attrs[0]);
     mean_dist_attr = std::move(dist_attrs[1]);
     variance_dist_attr = std::move(dist_attrs[2]);
 
@@ -157,7 +157,7 @@ SpmdInfo MatmulInferSpmd(const DistMetaTensor& x,
   std::string out_axes;
   FillMatmulOperandNotation(x_ndim, y_ndim, &x_axes, &y_axes, &out_axes);
 
-  // Step2: Sharding Propogation
+  // Step2: Sharding Propagation
   if (trans_x) {
     PADDLE_ENFORCE_GE(x_ndim,
                       2,
@@ -251,7 +251,7 @@ SpmdInfo MatmulInferSpmdReverse(const DistMetaTensor& x,
   std::string out_axes;
   FillMatmulOperandNotation(x_ndim, y_ndim, &x_axes, &y_axes, &out_axes);
 
-  // step2: Sharding Propogation
+  // step2: Sharding Propagation
   // should not use input dims mapping for backward sharding merge
   auto axis_to_dim_map =
       ShardingMergeForTensors({{out_axes, out_dims_mapping}}, false);
 
@@ -90,14 +90,14 @@ SpmdInfo ReductionInferSpmdBase(const DistMetaTensor& x,
   // get einsum notation for output
   std::string out_axes = GetOutputNotation(x_ndim, alphabet, axis, keep_dim);
 
-  // Step2: Sharding Propogation
+  // Step2: Sharding Propagation
   // Step2.1: Merge input shardings
   std::pair<std::string, std::vector<int64_t>> x_sharding_info(x_axes,
                                                                x_dims_mapping);
   std::unordered_map<std::string, int64_t> axis_to_dim_map =
       ShardingMergeForTensors({x_sharding_info});
 
-  // Step2.2: Infer output dimsmapping from merged input dimsmapping
+  // Step2.2: Infer output dims mapping from merged input dims mapping
   std::vector<int64_t> out_dims_mapping =
       GetDimsMappingForAxes(out_axes, axis_to_dim_map);
 
@@ -119,7 +119,7 @@ SpmdInfo ReductionInferSpmdBase(const DistMetaTensor& x,
   // If the op is a linear op, i.e. `linearity` is true, it supports
   // the input to be partial. Otherwise, the input cannot be partial
   // on reduced axes, we should reshard the input when the reduced
-  // axes are parital.
+  // axes are partial.
   VLOG(4) << "ReductionInferSpmd:";
   VLOG(4) << "axis: " << str_join(axis) << ", keep_dim: " << keep_dim;
   VLOG(4) << "Einsum Notation: " << x_axes << " --> " << out_axes;
@@ -195,7 +195,7 @@ SpmdInfo ReductionInferSpmdReverse(const DistMetaTensor& x,
   // get einsum notation for output
   std::string out_axes = GetOutputNotation(x_ndim, alphabet, axis, keep_dim);
 
-  // Step2: Sharding propogation
+  // Step2: Sharding propagation
   // Step2.1: Merge input shardings
   std::unordered_map<std::string, int64_t> axis_to_dim_map =
       ShardingMergeForTensors({{out_axes, out_dims_mapping}});
Original file line number	Diff line number	Diff line change
`@@ -88,7 +88,7 @@ class BeamSearchFunctor {`
`88`	`88`	`*`
`89`	`89`	`* @selected_ids: a [None, 1]-shaped tensor with LoD.`
`90`	`90`	`* In a machine translation model, it might be the candidate term id sets,`
`91`		`- * each set stored as a varience-length sequence.`
	`91`	`+ * each set stored as a variance-length sequence.`
`92`	`92`	`* The format might be described with a two-level LoD`
`93`	`93`	`* - [[0 1],`
`94`	`94`	`* [0 1 2]]`