minor spelling tweaks

tensorflow/core/api_def/base_api/api_def_ApplyFtrl.pbtxt

@@ -33,13 +33,13 @@ END
   in_arg {
     name: "l1"
     description: <<END
-L1 regulariation. Must be a scalar.
+L1 regularization. Must be a scalar.
 END
   }
   in_arg {
     name: "l2"
     description: <<END
-L2 regulariation. Must be a scalar.
+L2 regularization. Must be a scalar.
 END
   }
   in_arg {

tensorflow/core/api_def/base_api/api_def_ApplyFtrlV2.pbtxt

@@ -33,13 +33,13 @@ END
   in_arg {
     name: "l1"
     description: <<END
-L1 regulariation. Must be a scalar.
+L1 regularization. Must be a scalar.
 END
   }
   in_arg {
     name: "l2"
     description: <<END
-L2 shrinkage regulariation. Must be a scalar.
+L2 shrinkage regularization. Must be a scalar.
 END
   }
   in_arg {

tensorflow/core/api_def/base_api/api_def_BoostedTreesCalculateBestFeatureSplit.pbtxt

@@ -35,7 +35,7 @@ END
   in_arg {
     name: "min_node_weight"
     description: <<END
-mininum avg of hessians in a node before required for the node to be considered for splitting.
+minimum avg of hessians in a node before required for the node to be considered for splitting.
 END
   }
   out_arg {

tensorflow/core/api_def/base_api/api_def_BoostedTreesCalculateBestGainsPerFeature.pbtxt

@@ -34,7 +34,7 @@ END
   in_arg {
     name: "min_node_weight"
     description: <<END
-mininum avg of hessians in a node before required for the node to be considered for splitting.
+minimum avg of hessians in a node before required for the node to be considered for splitting.
 END
   }
   out_arg {
@@ -90,4 +90,4 @@ In this manner, the output is the best split per features and per node, so that
 The length of output lists are all of the same length, `num_features`.
 The output shapes are compatible in a way that the first dimension of all tensors of all lists are the same and equal to the number of possible split nodes for each feature.
 END
-}
+}

tensorflow/core/api_def/base_api/api_def_BoostedTreesSparseCalculateBestFeatureSplit.pbtxt

@@ -47,7 +47,7 @@ END
   in_arg {
     name: "min_node_weight"
     description: <<END
-mininum avg of hessians in a node before required for the node to be considered for splitting.
+minimum avg of hessians in a node before required for the node to be considered for splitting.
 END
   }
   out_arg {

tensorflow/core/api_def/base_api/api_def_CudnnRNNCanonicalToParamsV2.pbtxt

@@ -18,7 +18,7 @@ weights: the canonical form of weights that can be used for saving
 biases: the canonical form of biases that can be used for saving
     and restoration. They are more likely to be compatible across different
     generations.
-num_params_weigths: number of weight parameter matrix for all layers.
+num_params_weights: number of weight parameter matrix for all layers.
 num_params_biases: number of bias parameter vector for all layers.
 rnn_mode: Indicates the type of the RNN model.
 input_mode: Indicate whether there is a linear projection between the input and

tensorflow/core/api_def/base_api/api_def_CudnnRNNParamsToCanonicalV2.pbtxt

@@ -12,7 +12,7 @@ biases.
 num_layers: Specifies the number of layers in the RNN model.
 num_units: Specifies the size of the hidden state.
 input_size: Specifies the size of the input state.
-num_params_weigths: number of weight parameter matrix for all layers.
+num_params_weights: number of weight parameter matrix for all layers.
 num_params_biases: number of bias parameter vector for all layers.
 weights: the canonical form of weights that can be used for saving
     and restoration. They are more likely to be compatible across different

tensorflow/core/api_def/base_api/api_def_DebugNumericSummaryV2.pbtxt

@@ -20,7 +20,7 @@ Supported values:
   element is a bit which is set to 1 if the input tensor has an
   infinity or nan value, or zero otherwise.
 
-  3 (CONCISE_HEALTH): Ouput a float32/64 tensor of shape [5]. The 1st
+  3 (CONCISE_HEALTH): Output a float32/64 tensor of shape [5]. The 1st
   element is the tensor_id, if provided, and -1 otherwise. The
   remaining four slots are the total number of elements, -infs,
   +infs, and nans in the input tensor respectively.

tensorflow/core/api_def/base_api/api_def_IsBoostedTreesEnsembleInitialized.pbtxt

@@ -4,7 +4,7 @@ op {
   in_arg {
     name: "tree_ensemble_handle"
     description: <<END
-Handle to the tree ensemble resouce.
+Handle to the tree ensemble resource.
 END
   }
   out_arg {

tensorflow/core/api_def/base_api/api_def_ResourceApplyFtrl.pbtxt

@@ -33,13 +33,13 @@ END
   in_arg {
     name: "l1"
     description: <<END
-L1 regulariation. Must be a scalar.
+L1 regularization. Must be a scalar.
 END
   }
   in_arg {
     name: "l2"
     description: <<END
-L2 regulariation. Must be a scalar.
+L2 regularization. Must be a scalar.
 END
   }
   in_arg {

tensorflow/core/api_def/base_api/api_def_ResourceApplyFtrlV2.pbtxt

@@ -33,13 +33,13 @@ END
   in_arg {
     name: "l1"
     description: <<END
-L1 regulariation. Must be a scalar.
+L1 regularization. Must be a scalar.
 END
   }
   in_arg {
     name: "l2"
     description: <<END
-L2 shrinkage regulariation. Must be a scalar.
+L2 shrinkage regularization. Must be a scalar.
 END
   }
   in_arg {

tensorflow/core/api_def/base_api/api_def_ResourceSparseApplyFtrlV2.pbtxt

@@ -45,7 +45,7 @@ END
   in_arg {
     name: "l2"
     description: <<END
-L2 shrinkage regulariation. Must be a scalar.
+L2 shrinkage regularization. Must be a scalar.
 END
   }
   in_arg {

tensorflow/core/api_def/base_api/api_def_SparseApplyFtrlV2.pbtxt

@@ -45,7 +45,7 @@ END
   in_arg {
     name: "l2"
     description: <<END
-L2 shrinkage regulariation. Must be a scalar.
+L2 shrinkage regularization. Must be a scalar.
 END
   }
   in_arg {

tensorflow/core/api_def/base_api/api_def_TridiagonalMatMul.pbtxt

@@ -5,7 +5,7 @@ op {
     name: "superdiag"
     description: <<END
 Tensor of shape `[..., 1, M]`, representing superdiagonals of
-tri-diagonal matrices to the left of multiplication. Last element is ingored.
+tri-diagonal matrices to the left of multiplication. Last element is ignored.
 END
   }
   in_arg {
@@ -19,7 +19,7 @@ END
     name: "subdiag"
     description: <<END
 Tensor of shape `[..., 1, M]`, representing subdiagonals of tri-diagonal
-matrices to the left of multiplication. First element is ingored.
+matrices to the left of multiplication. First element is ignored.
 END
   }
   in_arg {

tensorflow/core/common_runtime/colocation_graph.cc

@@ -1004,7 +1004,7 @@ void ColocationGraph::GetSoftDeviceCandidates(
 
   // Failed to find supported devices that don't violate resource devices.
   // Try finding some devices that violated resource devices.
-  // If we succceed, we will log a warning below.
+  // If we succeed, we will log a warning below.
   soft_device_name = root_member.GetSoftDeviceName();
   device_set_.FindMatchingDevices(soft_device_name, possible_devices);
   if (!possible_devices->empty()) {

tensorflow/core/common_runtime/constant_folding_test.cc

@@ -704,7 +704,7 @@ class TestTFFileSystem : public ::tensorflow::NullFileSystem {
   ::tensorflow::Tensor data_tensor_;
 };
 
-// A test TF environent that checks that the environment was used.
+// A test TF environment that checks that the environment was used.
 class TestTFEnvironment : public ::tensorflow::EnvWrapper {
  public:
   using tf_base = ::tensorflow::EnvWrapper;

tensorflow/core/common_runtime/dynamic_device_mgr.cc

@@ -125,7 +125,7 @@ Status DynamicDeviceMgr::AddDevices(
     if (device_map_.find(d->name()) != device_map_.end()) {
       return errors::InvalidArgument(
           "Trying to add device ", d->name(),
-          " to manager but its name conflicts with an existing deivce.");
+          " to manager but its name conflicts with an existing device.");
     }
     // Register under the (1) full name and (2) canonical name.
     for (const string& name :

tensorflow/core/common_runtime/dynamic_device_mgr_test.cc

@@ -101,7 +101,7 @@ TEST(DynamicDeviceMgrTest, AddRepeatedDeviceToMgr) {
   added_devices.emplace_back(std::move(d1));
   Status s = dm->AddDevices(std::move(added_devices));
   EXPECT_TRUE(absl::StrContains(s.error_message(),
-                                "name conflicts with an existing deivce"));
+                                "name conflicts with an existing device"));
 }
 
 TEST(DynamicDeviceMgrTest, RemoveNonExistingDeviceFromMgr) {

tensorflow/core/common_runtime/eager/context.h

@@ -609,7 +609,7 @@ class EagerContext : public core::RefCounted {
 
   uint64 context_id_ GUARDED_BY(remote_state_mu_);
   // The view id of an eager context should be set to 0 when context is created,
-  // and continously incremented when context with the same context_id gets
+  // and continuously incremented when context with the same context_id gets
   // updated. The view id should be consistent between master and workers.
   uint64 context_view_id_ GUARDED_BY(remote_state_mu_);
   std::vector<string> remote_contexts_;

tensorflow/core/common_runtime/eager/eager_executor.h

@@ -190,7 +190,7 @@ class EagerExecutor {
   void NotifyWaiters(uint64 id) EXCLUSIVE_LOCKS_REQUIRED(node_queue_mutex_);
 
   // Starts execution of pending EagerNodes. This function loops till executor
-  // state_ is set to kShutDown. If any errors are encontered, these are set
+  // state_ is set to kShutDown. If any errors are encountered, these are set
   // inside `status_`. The loop blocks anytime there are no pending nodes, or if
   // `status_` is not ok.
   void Run();

tensorflow/core/common_runtime/eager/eager_op_rewrite_registry.h

@@ -34,7 +34,7 @@ class EagerOpRewrite {
 
   virtual ~EagerOpRewrite() {}
 
-  // To be implemnted by an Eager op rewrite pass.
+  // To be implemented by an Eager op rewrite pass.
   virtual Status Run(EagerOperation* orig_op,
                      std::unique_ptr<tensorflow::EagerOperation>* out_op) = 0;
 

tensorflow/core/common_runtime/eager/kernel_and_device.cc

@@ -326,7 +326,7 @@ Status KernelAndDeviceFunc::Run(
   } else {
     opts = absl::make_unique<FunctionLibraryRuntime::Options>();
     if (get_op_id_ && is_cross_process_) {
-      // If the function is a cross-process function and the remote excution
+      // If the function is a cross-process function and the remote execution
       // goes through eager service, create an eager op id for the function.
       opts->op_id = get_op_id_();
     }

tensorflow/core/common_runtime/executor_test.cc

@@ -255,7 +255,7 @@ void BuildConcurrentAddAssign(Graph* g) {
   auto one = test::graph::Constant(g, V(1.0));
   // A variable holds one float.
   auto var = test::graph::Var(g, DT_FLOAT, TensorShape({}));
-  // Initilize the variable with 1.0.
+  // Initialize the variable with 1.0.
   auto init = test::graph::Assign(g, var, one);
   // Output
   auto out = test::graph::Send(g, var, "out", ALICE, kIncarnation, BOB);

tensorflow/core/common_runtime/function.cc

@@ -1544,7 +1544,7 @@ std::vector<string> InputDevices(const Node& caller) {
   return input_devices;
 }
 
-// Place input nodes on the same device as the correspinding caller input
+// Place input nodes on the same device as the corresponding caller input
 // node. Do not specify any placement for all other nodes.
 class DefaultFunctionBodyPlacer : public InlinedFunctionBodyPlacer {
  public:
@@ -1593,7 +1593,7 @@ class SingleDeviceFunctionBodyPlacer : public InlinedFunctionBodyPlacer {
   const string caller_device_;
 };
 
-// Place input nodes on the same device as the correspinding caller input
+// Place input nodes on the same device as the corresponding caller input
 // node. Do not place output node. Place control nodes on the same device as
 // caller node. For all function body nodes overrides job, replica and task
 // parts of the device assignment to match function caller node.

tensorflow/core/common_runtime/function_test.cc

@@ -1312,7 +1312,7 @@ int GetConstantFoldingCounter() {
       return counter;
     }
   }
-  LOG(FATAL) << "Should have found a node that replcaed add";
+  LOG(FATAL) << "Should have found a node that replaced add";
 }
 
 TEST_F(FunctionLibraryRuntimeTest, OptimizeGraph) {
@@ -2129,7 +2129,7 @@ TEST(OptimizationTest, RemoveListArrayConverter) {
   }
 }
 
-TEST(OptimizationTest, RemoveListArrayConverter_WithContolDeps) {
+TEST(OptimizationTest, RemoveListArrayConverter_WithControlDeps) {
   auto func = FDH::Create(
       // Name
       "Test",

tensorflow/core/common_runtime/gpu/gpu_device.cc

@@ -1837,7 +1837,7 @@ void GPUKernelTracker::RecordTerminated(uint64 queued_count) {
   // advance the completed frontier to the just-completed PendingKernel.  In
   // practice we occasionally see the termination callbacks arrive out of
   // order probably because of thread scheduling.  Eventually we may support
-  // out-of- order completion involving multple compute streams so here we
+  // out-of- order completion involving multiple compute streams so here we
   // follow a conservative approach and wait for every single callback to
   // arrive before advancing the frontier.
   while (true) {

tensorflow/core/common_runtime/gpu/gpu_device.h

@@ -157,7 +157,7 @@ class BaseGPUDevice : public LocalDevice {
   int32 pending_cap_ = 0;
   bool timestamped_allocator_ = false;
 
-  // Initialize scractch buffers used by Eigen.
+  // Initialize scratch buffers used by Eigen.
   Status InitScratchBuffers();
 
   void ReinitializeDevice(OpKernelContext* context, PerOpGpuDevice* device,

tensorflow/core/common_runtime/gpu/gpu_id.h

@@ -56,7 +56,7 @@ namespace tensorflow {
 //
 // Assuming we configure the Session to create one BaseGPUDevice per GPU
 // hardware, then setting GPUOptions::visible_device_list to "2,0" will create
-// the following mappting between TF GPU id and platform GPU id:
+// the following mapping between TF GPU id and platform GPU id:
 //
 //                  TF GPU id  ->  platform GPU ID
 //      0 (i.e. /device:GPU:0) ->  2
@@ -67,7 +67,7 @@ namespace tensorflow {
 //
 // On the other hand, if we configure it to create 2 BaseGPUDevice per GPU
 // hardware, then setting GPUOptions::visible_device_list to "2,0" will create
-// the following mappting between TF GPU id and platform GPU id:
+// the following mapping between TF GPU id and platform GPU id:
 //
 //                  TF GPU id  ->  platform GPU ID
 //      0 (i.e. /device:GPU:0) ->  2

tensorflow/core/common_runtime/graph_execution_state.cc

@@ -615,7 +615,7 @@ Status GraphExecutionState::InitBaseGraph(std::unique_ptr<Graph>&& new_graph) {
                     session_options_->config.allow_soft_placement(),
                 session_options_ != nullptr &&
                     session_options_->config.log_device_placement());
-  // TODO(mrry): Consider making the Placer cancelable.
+  // TODO(mrry): Consider making the Placer cancellable.
   TF_RETURN_IF_ERROR(placer.Run());
 
   TF_RETURN_IF_ERROR(OptimizationPassRegistry::Global()->RunGrouping(

tensorflow/core/common_runtime/metrics.h

@@ -53,7 +53,7 @@ void RecordTFDataFingerprint(const string& name);
 // Records the number of independent graph changes resulting from the
 // application of a tf.data optimization.
 //
-// The `name` argument identifies the optimization (e.g. "noop_eliminiation").
+// The `name` argument identifies the optimization (e.g. "noop_elimination").
 void RecordTFDataOptimization(const string& name, int64 num_changes);
 
 // Records parsing of dense tensor features.

tensorflow/core/common_runtime/optimization_registry.h

@@ -73,7 +73,7 @@ class GraphOptimizationPass {
   string name() const { return name_; }
 
  private:
-  // The name of the opitimization pass, which is the same as the inherited
+  // The name of the optimization pass, which is the same as the inherited
   // class name.
   string name_;
 };

tensorflow/core/common_runtime/partitioning_utils.h

@@ -41,7 +41,7 @@ Status PartitionFunctionGraph(
 //
 // More specifically, this function
 //  (1) rewrites the indices of the `Arg` and `Retval` nodes placed
-//      on a particular device.  When a function is parittioned each
+//      on a particular device.  When a function is partitioned each
 //      partition, `subgraph`, get a subset of the arguments and
 //      return values. The `index` attributes of these _Arg and _Retval
 //      nodes reflect the indices of these parameters in the original

tensorflow/core/common_runtime/placer_test.cc

@@ -274,7 +274,7 @@ class PlacerTest : public ::testing::Test {
     RewriterConfig* rewriter_config = graph_opts->mutable_rewrite_options();
     rewriter_config->set_disable_meta_optimizer(true);
 
-    // Placing nested functions requires go through some PRE_PLACEMNT passes.
+    // Placing nested functions requires go through some PRE_PLACEMENT passes.
     // Currently, just the IsolateDeepOpsPass.
     GraphOptimizationPassOptions optimization_options;
     std::unique_ptr<Graph> graph_ptr(graph);
@@ -1493,7 +1493,7 @@ TEST_F(PlacerTest, TestUnknownAssignedDevice) {
 
 // Test that placement fails when an op with no registered kernels is
 // requested and no device is requested for the node
-TEST_F(PlacerTest, TestNoKernelsRegisteredWithNoRequstedDevice) {
+TEST_F(PlacerTest, TestNoKernelsRegisteredWithNoRequestedDevice) {
   Graph g(OpRegistry::Global());
   {  // Scope for temporary variables used to construct g.
     GraphDefBuilder b(GraphDefBuilder::kFailImmediately);
@@ -2286,7 +2286,7 @@ TEST_F(NestedPlacerTest, OutputTwoResources_UnassignedResource) {
    * the "second pass" as they are "sources". It assigns `r1` to GPU because it
    * is in the same group as `b`. It assigns `r2` to GPU because GPU has a
    * higher device preference. Finally, `a` is assigned to GPU because `r2` is
-   * on GPU - this test that the "second pass" heuristics respect colocaton
+   * on GPU - this test that the "second pass" heuristics respect colocation
    * groups (even when the consumer of the source, i.e. PCO is on a different
    * device).
    */
@@ -2494,7 +2494,7 @@ TEST_F(NestedPlacerTest, DuplicateInputResource_Conflict) {
    *                 r1:RESOURCE:GPU
    *
    * There is a conflict but Placer always overrides requested devices
-   * when they result in coflict due to resource edges. Which device
+   * when they result in conflict due to resource edges. Which device
    * is picked for a/r1/r2 is indeterministic.
    */
   FunctionDef func = test::function::Swap();

tensorflow/core/common_runtime/process_function_library_runtime.h

@@ -444,7 +444,7 @@ class ProcessFunctionLibraryRuntime {
   std::unordered_map<string, FunctionLibraryRuntime::Handle> table_
       GUARDED_BY(mu_);
 
-  // Function data for instantitated remote functions.
+  // Function data for instantiated remote functions.
   std::unordered_map<FunctionLibraryRuntime::Handle,
                      std::unique_ptr<FunctionData>>
       function_data_ GUARDED_BY(mu_);

tensorflow/core/debug/debug_graph_utils_test.cc

@@ -126,7 +126,7 @@ TEST_F(DebugGraphUtilsTest, TestValidDebugOpNameWithMoreThanOneAttributes) {
   ASSERT_EQ("100", attributes["first_n"]);
 }
 
-TEST_F(DebugGraphUtilsTest, TestValidDebugOpNameWithMoreDuplicatettributes) {
+TEST_F(DebugGraphUtilsTest, TestValidDebugOpNameWithMoreDuplicateAttributes) {
   string debug_op_name_proper;
   std::unordered_map<string, string> attributes;
   Status s = ParseDebugOpName(

tensorflow/core/distributed_runtime/eager/cluster_function_library_runtime.cc

@@ -162,7 +162,7 @@ void EagerClusterFunctionLibraryRuntime::Run(
     remote_op->add_inputs()->Swap(&(*args)[i]);
   }
   // The remote component function should use the same op_id as its parent
-  // multi-device function's in order to get the global unqiue op_id generated
+  // multi-device function's in order to get the global unique op_id generated
   // by the master context.
   remote_op->set_id(opts.op_id.value());
   remote_op->set_is_function(true);

tensorflow/core/distributed_runtime/master_session.cc

@@ -588,7 +588,7 @@ class RunManyGraphs {
   void ReportBadStatus(const Status& s) EXCLUSIVE_LOCKS_REQUIRED(mu_) {
     VLOG(1) << "Master received error status " << s;
     if (!cancel_issued_ && !StatusGroup::IsDerived(s)) {
-      // Only start cancelling other workers upon receiveing a non-derived
+      // Only start cancelling other workers upon receiving a non-derived
       // error
       cancel_issued_ = true;
 

tensorflow/core/distributed_runtime/rpc/grpc_master_service.cc

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
-// GrpcMasterService implements the RPC service MasterSerivce.
+// GrpcMasterService implements the RPC service MasterService.
 //
 // A GrpcMasterService maintains the state of live graph computation
 // sessions, each session orchestrates both local and remote devices

tensorflow/core/distributed_runtime/rpc/grpc_state.cc

@@ -24,8 +24,8 @@ const char* ToString(UntypedStreamingRPCState::Tag::TagType tag_type) {
       return "kCallStarted";
     case UntypedStreamingRPCState::Tag::TagType::kRequestWriteCompleted:
       return "kRequestWriteCompleted";
-    case UntypedStreamingRPCState::Tag::TagType::kResponseReadCommpleted:
-      return "kResponseReadCommpleted";
+    case UntypedStreamingRPCState::Tag::TagType::kResponseReadCompleted:
+      return "kResponseReadCompleted";
     case UntypedStreamingRPCState::Tag::TagType::kCallFinished:
       return "kCallFinished";
   }
@@ -43,7 +43,7 @@ void UntypedStreamingRPCState::Tag::OnCompleted(bool ok) {
     case TagType::kRequestWriteCompleted:
       streaming_state_->RequestWriteCompleted(ok);
       break;
-    case TagType::kResponseReadCommpleted:
+    case TagType::kResponseReadCompleted:
       streaming_state_->ResponseReadCompleted(ok);
       break;
     case TagType::kCallFinished:

tensorflow/core/distributed_runtime/rpc/grpc_state.h

@@ -221,7 +221,7 @@ class UntypedStreamingRPCState : public core::RefCounted {
     enum class TagType {
       kCallStarted,
       kRequestWriteCompleted,
-      kResponseReadCommpleted,
+      kResponseReadCompleted,
       kCallFinished,
     };
 
@@ -337,7 +337,7 @@ class ExchangeQueue {
 
   // Changes the state of the exchange that is current in kRequestWriteIssued
   // state to kRequestWriteCompleted state.
-  // REQUIRES: There is an exhange in kRequestWriteIssued state.
+  // REQUIRES: There is an exchange in kRequestWriteIssued state.
   void MarkRequestWriteCompleted();
 
   // Returns the exchange at the front of the queue.
@@ -536,7 +536,7 @@ class StreamingRPCState : public UntypedStreamingRPCState {
   void MarkDoneAndCompleteExchanges(Status status) EXCLUSIVE_LOCKS_REQUIRED(mu_)
       UNLOCK_FUNCTION(mu_) {
     call_state_ = State::kDone;
-    VLOG(2) << "Ending gRPC stremaing call on the client side due to "
+    VLOG(2) << "Ending gRPC streaming call on the client side due to "
             << status.ToString();
     // Swap the exchanges_ into a temporary ExchangeQueue so that we can
     // complete all exchanges without holding mu_ in case user callback
@@ -604,7 +604,7 @@ class StreamingRPCState : public UntypedStreamingRPCState {
   // Tags are immutable. No need to guard them.
   Tag call_started_tag_{this, Tag::TagType::kCallStarted};
   Tag request_write_completed_tag_{this, Tag::TagType::kRequestWriteCompleted};
-  Tag response_read_completed_tag_{this, Tag::TagType::kResponseReadCommpleted};
+  Tag response_read_completed_tag_{this, Tag::TagType::kResponseReadCompleted};
   Tag finished_tag_{this, Tag::TagType::kCallFinished};
 };
 

tensorflow/core/distributed_runtime/test_utils.h

@@ -62,7 +62,7 @@ class TestWorkerInterface : public WorkerInterface {
   }
 
   void RunGraphAsync(CallOptions* opts, RunGraphRequestWrapper* request,
-                     MutableRunGraphResponseWrapper* repsonse,
+                     MutableRunGraphResponseWrapper* response,
                      StatusCallback done) override {
     done(errors::Unimplemented("RunGraphAsync"));
   }

tensorflow/core/distributed_runtime/worker_interface.h

@@ -57,7 +57,7 @@ class WorkerInterface {
                                     StatusCallback done) = 0;
 
   virtual void RunGraphAsync(CallOptions* opts, RunGraphRequestWrapper* request,
-                             MutableRunGraphResponseWrapper* repsonse,
+                             MutableRunGraphResponseWrapper* response,
                              StatusCallback done) = 0;
 
   virtual void RunGraphAsync(CallOptions* opts, const RunGraphRequest* request,

tensorflow/core/framework/common_shape_fns.h

@@ -338,7 +338,7 @@ inline Status BroadcastBinaryOpShapeFn(InferenceContext* c) {
 // Shape function for random operations.
 Status RandomShape(shape_inference::InferenceContext* c);
 
-// Shape function for Slice opertaions.
+// Shape function for Slice operations.
 Status SliceShape(shape_inference::InferenceContext* c);
 
 // Validates the 3 component tensors of a sparse tensor have the proper

tensorflow/core/framework/dataset.h

@@ -493,7 +493,7 @@ class SerializationContext {
     // `input_list`.
     bool fail_if_unimplemented = true;
 
-    // Indicates whether (potentionally large) data tensors should be
+    // Indicates whether (potentially large) data tensors should be
     // serialized, or replaced with a placeholder returned in `input_list`. The
     // latter makes sense to do when performing data agnostic graph rewrites to
     // reduce the memory usage.
@@ -641,7 +641,7 @@ class IteratorBase {
     return errors::Unimplemented("RestoreInternal");
   }
 
-  // Returns the number of elements produced by this itertaor.
+  // Returns the number of elements produced by this iterator.
   int64 num_elements() const {
     if (node_) return node_->num_elements();
     return 0;

tensorflow/core/framework/dataset_stateful_op_whitelist.h

@@ -62,7 +62,7 @@ class WhitelistedStatefulOpRegistry {
 // Note that the state of the whitelisted ops inside functions will not be
 // saved during checkpointing, hence this should only be used if the op is
 // marked stateful for reasons like to avoid constant folding during graph
-// optimiztion but is not stateful.
+// optimization but is not stateful.
 // If possible, try to remove the stateful flag on the op first.
 // Example usage:
 //

tensorflow/core/framework/function.h

@@ -665,7 +665,7 @@ class FunctionLibraryRuntime {
   // "handle".
   //
   // If function execution succeeds, "done" is called with OK and
-  // "*rets" is filled with the function's return values. Otheriwse,
+  // "*rets" is filled with the function's return values. Otherwise,
   // "done" is called with an error status.
   //
   // Does not take ownership of "rets".

tensorflow/core/framework/model.cc

@@ -786,7 +786,7 @@ std::map<string, std::shared_ptr<Parameter>> Model::CollectTunableParameters(
 
 std::map<string, std::shared_ptr<Parameter>> Model::CollectEssentialParallelism(
     std::shared_ptr<Node> node) {
-  // Parallelism parameter is considered to be essential if the coressponding
+  // Parallelism parameter is considered to be essential if the corresponding
   // transformations's processing time is greater than essential rate times the
   // average transformation self processing time.
   constexpr double kEssentialRate = 0.3L;

tensorflow/core/framework/node_def_util.h

@@ -158,7 +158,7 @@ class AttrSlice {
   // account.
   //
   // TODO(irving): There is a bug in this routine inherited from its
-  // OptimizerCSE::EqualAttrs precedecessor.  The same tensor attr can be
+  // OptimizerCSE::EqualAttrs predecessor.  The same tensor attr can be
   // represented in more than one way as an AttrValue, since TensorProto is
   // not 1-1.  This bug will go away once I replace everything with NodeInfo,
   // which stores a Tensor object directly.  The Scratch object will also go

tensorflow/core/framework/run_handler.cc

@@ -205,7 +205,7 @@ class ThreadWorkSource {
       // thread pool waiting queues. Wake up threads from sub thread waiting
       // queues.
       // The waiting queues are defined at RunHandlerPool.
-      // Get the waiter_queue and coresponding mutex. Note, the thread work
+      // Get the waiter_queue and corresponding mutex. Note, the thread work
       // source may change afterwards if a new request comes or an old request
       // finishes.
       tf_shared_lock lock(run_handler_waiter_mu_);
@@ -522,7 +522,7 @@ class RunHandlerThreadPool {
 
   // Search tasks from Requets range searching_range_start to
   // searching_range_end. If there is no tasks in the search range and
-  // may_steal_blocking_work is true, then search from all reuqests.
+  // may_steal_blocking_work is true, then search from all requests.
   Task FindTask(
       int searching_range_start, int searching_range_end, int thread_id,
       int sub_thread_pool_id, int max_blocking_inflight,
@@ -596,7 +596,7 @@ Task RunHandlerThreadPool::FindTask(
   int current_index = thread_data_[thread_id].current_index;
   *task_from_blocking_queue = false;
 
-  // TODO(chaox): Chagne the search algorithm from round robin to random
+  // TODO(chaox): Change the search algorithm from round robin to random
   // walk.
   for (int i = 0; i < searching_range_end - searching_range_start; ++i) {
     if (current_index >= searching_range_end) {

tensorflow/core/framework/session_state.h

@@ -36,7 +36,7 @@ class SessionState {
   // Store a tensor in the session state.
   Status AddTensor(const string& handle, const Tensor& tensor);
 
-  // Delete a tensdor from the session state.
+  // Delete a tensor from the session state.
   Status DeleteTensor(const string& handle);
 
   int64 GetNewId();

tensorflow/core/framework/stats_aggregator.h

@@ -36,7 +36,7 @@ namespace data {
 //
 // NOTE(mrry): `StatsAggregator` is a virtual interface because we anticipate
 // that many different implementations will have the same interface. For
-// example, we have diffferent implementations in "stats_aggregator_ops.cc" for
+// example, we have different implementations in "stats_aggregator_ops.cc" for
 // simple in-memory implementation that integrates with the pull-based summary
 // API, and for the push-based `SummaryWriterInterface`, and we may add
 // implementations that work well with other custom monitoring services.
@@ -50,7 +50,7 @@ class StatsAggregator {
                               gtl::ArraySlice<double> values,
                               int64 global_step) = 0;
 
-  // TODO(shivaniagarawal): consistency in double and float usage.
+  // TODO(shivaniagrawal): consistency in double and float usage.
   // Add the given `value` as Scalar with the given `name`.
   virtual void AddScalar(const string& name, float value,
                          int64 global_step) = 0;
@@ -67,7 +67,7 @@ class StatsAggregator {
                                 int64 val) = 0;
 };
 
-// A `StatsAggregatorResource` wraps a shareable `StatsAggregator` as a resource
+// A `StatsAggregatorResource` wraps a sharable `StatsAggregator` as a resource
 // in the TensorFlow resource manager.
 //
 // NOTE(mrry): This class is separate from `StatsAggregator` in order to

tensorflow/core/framework/variant_op_registry.h

@@ -332,7 +332,7 @@ Status BinaryOpVariants(OpKernelContext* ctx, VariantBinaryOp op,
                         const Variant& a, const Variant& b, Variant* out) {
   if (a.TypeId() != b.TypeId()) {
     return errors::Internal(
-        "BianryOpVariants: Variants a and b have different "
+        "BinaryOpVariants: Variants a and b have different "
         "type ids.  Type names: '",
         a.TypeName(), "' vs. '", b.TypeName(), "'");
   }

tensorflow/core/graph/gradients.cc

@@ -167,12 +167,12 @@ class SymbolicGradientBuilder {
 
   // A vector of output endpoints which represents backpropagated
   // gradients
-  typedef std::vector<NodeOut> BackpropedGradients;
+  typedef std::vector<NodeOut> BackproppedGradients;
 
   // backprops_ is a map from a node output to its accumulated
   // gradients.  When a node output has accumulated all its
   // gradients, we add a node which sums them up.
-  std::unordered_map<NodeOut, BackpropedGradients, NodeOutHash, NodeOutEq>
+  std::unordered_map<NodeOut, BackproppedGradients, NodeOutHash, NodeOutEq>
       backprops_;
 
   // pending[i] is count-down counter for i-th node's expected

tensorflow/core/graph/graph_constructor.cc

@@ -137,7 +137,7 @@ class GraphConstructor {
     // remove this.
     bool importing;
     // If true, validates that nodes being converted have all expected attrs
-    // set and no unknonw attrs set by calling ValidateNodeDef().
+    // set and no unknown attrs set by calling ValidateNodeDef().
     // `validate_nodes` is always true when `importing` is set.
     bool validate_nodes;
     bool validate_colocation_constraints;

tensorflow/core/graph/graph_constructor.h

@@ -44,7 +44,7 @@ struct GraphConstructorOptions {
   bool expect_device_spec = false;
 
   // If true, validates that nodes being converted have all expected attrs
-  // set and no unknonw attrs set by calling ValidateNodeDef().
+  // set and no unknown attrs set by calling ValidateNodeDef().
   // Setting validate_nodes without add_default_attributes, will fail if
   // the GraphDef does not have all required attributes set.
   bool validate_nodes = false;

tensorflow/core/graph/graph_constructor_test.cc

@@ -3196,7 +3196,7 @@ versions {
   EXPECT_EQ(17, refiner.graph_def_version());
 }
 
-TEST_F(GraphConstructorTest, ImportGraphDef_ValidateColationConstraints) {
+TEST_F(GraphConstructorTest, ImportGraphDef_ValidateColocationConstraints) {
   GraphDef def;
   ASSERT_TRUE(protobuf::TextFormat::ParseFromString(
       "node { name: 'A' op: 'TestInput' attr { key: '_class' value { list { "

tensorflow/core/graph/mkl_graph_util.h

@@ -74,7 +74,7 @@ int inline GetTensorMetaDataIndex(int n, int total_tensors) {
   return DataIndexToMetaDataIndex(tidx, total_tensors);
 }
 
-// check if the control between src and dst nodes alreay exists
+// check if the control between src and dst nodes already exists
 bool inline DoesControlEdgeExist(const Node* src, const Node* dst) {
   for (const Edge* edge : src->out_edges()) {
     if (edge->IsControlEdge() && edge->dst() == dst) {

tensorflow/core/graph/mkl_layout_pass.cc

@@ -3639,7 +3639,7 @@ const MklLayoutRewritePass::RewriteInfo*
 MklLayoutRewritePass::CheckForNodeRewrite(const Node* n) const {
   CHECK_NOTNULL(n);
 
-  // QuntizedOps may have attributes other than "T", so decoupled the check
+  // QuantizedOps may have attributes other than "T", so decoupled the check
   // with a function, CheckForQuantizedNodeRewrite(const Node*).
   const RewriteInfo* ri = CheckForQuantizedNodeRewrite(n);
   if (ri != nullptr) return ri;
@@ -3789,7 +3789,7 @@ MklLayoutRewritePass::CheckForNodeFusion(Node* a) const {
     // a.k.a. "a->b->c" matches "op1->op2->op3"
     //
 
-    // Stores the first unvisted outgoing edge of each matched node in "nodes".
+    // Stores the first unvisited outgoing edge of each matched node in "nodes".
     std::stack<EdgeSet::const_iterator> current_neighbor_stack;
     nodes.clear();
 

tensorflow/core/grappler/clusters/single_machine_test.cc

@@ -613,7 +613,7 @@ TEST_F(SingleMachineTest, PeakMemoryStatsNotEnabled) {
 
   TF_CHECK_OK(cluster_->Shutdown());
   cluster_.reset();
-  SingleMachine cluster(60 /* timout_s */, 3 /* num_cpu_cores */,
+  SingleMachine cluster(60 /* timeout_s */, 3 /* num_cpu_cores */,
                         0 /* num_gpus */);
 
   TF_CHECK_OK(cluster.Provision());

tensorflow/core/grappler/costs/graph_properties.cc

@@ -524,7 +524,7 @@ bool IsWhiteListedOpTypeForEvaluateNode(const string& op_type) {
           "TruncateDiv",
           "TruncateMod",
           "RealDiv",
-          // N-ary arithemtic ops
+          // N-ary arithmetic ops
           "AddN",
           // Others
           "StridedSlice",
@@ -1313,7 +1313,7 @@ class SymbolicShapeRefiner {
     return true;
   }
 
-  // Create input tensors from the NodeConext.
+  // Create input tensors from the NodeContext.
   void CreateInputTensors(NodeContext* c,
                           std::vector<Tensor>* input_tensor_vector,
                           TensorVector* inputs) {
@@ -1362,7 +1362,7 @@ class SymbolicShapeRefiner {
 
     // Input to EvaluateNode()
     TensorVector inputs;
-    // Container for temporaily created tensor object.
+    // Container for temporarily created tensor object.
     std::vector<Tensor> input_tensor_vector(ic->num_inputs());
     CreateInputTensors(c, &input_tensor_vector, &inputs);
 

tensorflow/core/grappler/costs/graph_properties_test.cc

@@ -879,7 +879,7 @@ TEST_F(GraphPropertiesTest, InferRestoreOpShape) {
   EXPECT_EQ(DT_FLOAT, restorev2_prop.dtype());
   EXPECT_EQ("float: [128,256]", PropToString(restorev2_prop));
 
-  // Check input shapes of assign op are propagted correctly.
+  // Check input shapes of assign op are propagated correctly.
   const auto input_props = properties.GetInputProperties("init_restore");
   ASSERT_EQ(2, input_props.size());
   const OpInfo::TensorProperties& input_prop = input_props[1];
@@ -995,7 +995,7 @@ TEST_F(GraphPropertiesTest, SkippingValueInferenceForLargeTensors) {
   // When using aggressive_shape_inference, we run EvaluateNode() for
   // whitelisted ops and small input / output tensors. For instance, Fill op is
   // evaluated and produces output tensor value if output tensor size is smal
-  // (currently, fewer than 17 elements); otherwise we don't run EvalauteNode().
+  // (currently, fewer than 17 elements); otherwise we don't run EvaluateNode().
   // This is to avoid wasting time and memory for producing huge tensors (e.g.,
   // initializing a large table using Fill.
   {

tensorflow/core/grappler/costs/op_level_cost_estimator.cc

@@ -868,7 +868,7 @@ int64 OpLevelCostEstimator::CountMatMulOperations(const OpInfo& op_info,
     LOG(ERROR) << "Incompatible Matrix dimensions";
     return ops;
   } else {
-    // One of k_dim and k_dim_b might be 1 (mininum dimension size).
+    // One of k_dim and k_dim_b might be 1 (minimum dimension size).
     k_dim = std::max(k_dim, k_dim_b);
   }
 

tensorflow/core/grappler/costs/op_level_cost_estimator.h

@@ -200,7 +200,7 @@ class OpLevelCostEstimator {
   typedef std::function<Costs(const OpContext& op_context)> CostImpl;
   std::map<string, CostImpl> device_cost_impl_;
   // If true, assume compute and memory overlap; hence, the op cost is max of
-  // compute_time and memory_time, insteaf of sum of those two.
+  // compute_time and memory_time, instead of sum of those two.
   bool compute_memory_overlap_;
   std::set<string> persistent_ops_;
 

tensorflow/core/grappler/costs/op_level_cost_estimator_test.cc

@@ -521,7 +521,7 @@ class OpLevelCostEstimatorTest : public ::testing::Test {
     estimator_.compute_memory_overlap_ = value;
   }
 
-  void ValidateOpDimensionsFromImputs(const int n, const int h, const int w,
+  void ValidateOpDimensionsFromInputs(const int n, const int h, const int w,
                                       const int c, const int kx, const int ky,
                                       const int sx, const int sy,
                                       const string& data_format,
@@ -571,12 +571,12 @@ class OpLevelCostEstimatorTest : public ::testing::Test {
 TEST_F(OpLevelCostEstimatorTest, TestPersistentOpCosts) {
   OpContext op_context;
   SetCpuDevice(&op_context.op_info);
-  std::unordered_set<string> persisent_ops = {
+  std::unordered_set<string> persistent_ops = {
       "Const",       "Variable",       "VariableV2", "AutoReloadVariable",
       "VarHandleOp", "ReadVariableOp",
   };
-  // Minmum cost for all persistent ops.
-  for (const auto& op : persisent_ops) {
+  // Minimum cost for all persistent ops.
+  for (const auto& op : persistent_ops) {
     op_context.op_info.set_op(op);
     auto cost = estimator_.PredictCosts(op_context);
     EXPECT_EQ(Costs::Duration(0), cost.memory_time);
@@ -1128,10 +1128,10 @@ TEST_F(OpLevelCostEstimatorTest, OpDimensionsFromInputs) {
   for (const auto& p : paddings) {
     for (const auto& f : formats) {
       // n, h, w, c, kx, ky, sx, sy, data_format, padding.
-      ValidateOpDimensionsFromImputs(10, 20, 20, 100, 3, 3, 2, 2, f, p);
-      ValidateOpDimensionsFromImputs(10, 20, 20, 100, 1, 1, 3, 3, f, p);
-      ValidateOpDimensionsFromImputs(10, 200, 200, 100, 5, 5, 3, 3, f, p);
-      ValidateOpDimensionsFromImputs(10, 14, 14, 3840, 3, 3, 2, 2, f, p);
+      ValidateOpDimensionsFromInputs(10, 20, 20, 100, 3, 3, 2, 2, f, p);
+      ValidateOpDimensionsFromInputs(10, 20, 20, 100, 1, 1, 3, 3, f, p);
+      ValidateOpDimensionsFromInputs(10, 200, 200, 100, 5, 5, 3, 3, f, p);
+      ValidateOpDimensionsFromInputs(10, 14, 14, 3840, 3, 3, 2, 2, f, p);
     }
   }
 }

tensorflow/core/grappler/costs/virtual_scheduler.cc

@@ -262,7 +262,7 @@ const NodeDef* CompositeNodeManager::GetCurrNode() {
   // FirstReady for _Send and _Recv (separately),
   // Globally (among the LIFO-selected ops from each device and _Send and
   // _Recv) FirstReady,
-  // Priorty order: _Send, _Recv, and then the rest, if time_ready is equal.
+  // Priority order: _Send, _Recv, and then the rest, if time_ready is equal.
   std::vector<std::pair<const NodeDef*, Costs::Duration>> candidates;
   for (auto& ops_lifo : ops_lifo_map_) {
     if (!ops_lifo.second.Empty()) {
@@ -294,7 +294,7 @@ const NodeDef* CompositeNodeManager::GetCurrNode() {
             // Both are normal ops; use node name as tie breaker.
             return a.first->name().compare(b.first->name()) < 0;
           } else {
-            // Priortize by op type: _Send, _Recv, and normap ops.
+            // Prioritize by op type: _Send, _Recv, and normap ops.
             return a_score > b_score;
           }
         } else {
@@ -989,7 +989,7 @@ Costs VirtualScheduler::Summary() const {
     std::map<string, int64> op_to_memory;
     // First profile only persistent memory usage.
     int64 persistent_memory_usage = 0;
-    std::set<string> persisent_ops;
+    std::set<string> persistent_ops;
     for (const auto& node_port : state.persistent_nodes) {
       const auto* node = node_port.first;
       const auto port = node_port.second;
@@ -997,7 +997,7 @@ Costs VirtualScheduler::Summary() const {
           CalculateOutputSize(node_map_.at(node).output_properties, port);
       persistent_memory_usage += output_size;
       op_to_memory[node->op()] += output_size;
-      persisent_ops.insert(node->op());
+      persistent_ops.insert(node->op());
     }
     int64 max_memory_usage = persistent_memory_usage + state.max_memory_usage;
     critical_path_costs.estimated_max_memory_per_device[name] =
@@ -1076,7 +1076,7 @@ Costs VirtualScheduler::Summary() const {
                        memory_cost, intermediate_memory_cost)
                 << " (" << HumanReadableNumBytes(op_mem_usage) << " ["
                 << mem_usage_percent << "%] "
-                << (persisent_ops.count(op) > 0 ? ": persistent op)" : ")");
+                << (persistent_ops.count(op) > 0 ? ": persistent op)" : ")");
       }
     }
 

tensorflow/core/grappler/costs/virtual_scheduler_test.cc

@@ -415,7 +415,7 @@ TEST_F(ReadyNodeManagerTest, RemoveSingleNodeCompositeNodeManager) {
   EXPECT_TRUE(manager.Empty());
 }
 
-TEST_F(ReadyNodeManagerTest, GetAndRemoveMultipleComopsiteNodeManager) {
+TEST_F(ReadyNodeManagerTest, GetAndRemoveMultipleCompositeNodeManager) {
   CompositeNodeManager manager;
   TF_EXPECT_OK(manager.Init(&node_states_));
   manager.AddNode(&node1_);
@@ -445,7 +445,7 @@ TEST_F(ReadyNodeManagerTest, GetAndRemoveMultipleComopsiteNodeManager) {
   EXPECT_TRUE(manager.Empty());
 }
 
-TEST_F(ReadyNodeManagerTest, MultiDeviceSendRecvComopsiteNodeManager) {
+TEST_F(ReadyNodeManagerTest, MultiDeviceSendRecvCompositeNodeManager) {
   CompositeNodeManager manager;
   TF_EXPECT_OK(manager.Init(&node_states_));
   // Additional nodes on kCPU1.
@@ -2966,7 +2966,7 @@ TEST_F(VirtualSchedulerTest, GraphWihtOnlyRecv) {
 }
 
 TEST_F(VirtualSchedulerTest, AddMergeSwitch) {
-  // Override scheduler_ with CompositeNodeNamager.
+  // Override scheduler_ with CompositeNodeManager.
   scheduler_ = absl::make_unique<TestVirtualScheduler>(
       /*use_static_shapes=*/true,
       /*use_aggressive_shape_inference=*/true, &composite_node_manager_,

tensorflow/core/grappler/graph_analyzer/sig_node.cc

@@ -89,7 +89,7 @@ void SigNode::ComputeTopoHash0() {
   topo_hash_.clear();
   last_hashed_nodes_ = next_hashed_nodes_ = node_mask_;
 
-  // TODO(babkin): include the attrbutes too, as an option.
+  // TODO(babkin): include the attributes too, as an option.
   size_t hval = std::hash<string>()(opcode());
 
   // Getting the topology of the links in to the hash early should get more

tensorflow/core/grappler/mutable_graph_view.h

@@ -195,7 +195,7 @@ class MutableGraphView : public internal::GraphViewInternal<GraphDef, NodeDef> {
 
   // Removes control dependency `fanin_node_name` from the target node named
   // `node_name`. If the node or fanin do not exist in the graph, nothing will
-  // be modified in the graph. To remove regular fanins, use RemoveRegualrFanin.
+  // be modified in the graph. To remove regular fanins, use RemoveRegularFanin.
   //
   // If the fanin being removed doesn't exist in the node's inputs, this will
   // not result in an error and the node will not be modified.

tensorflow/core/grappler/mutable_graph_view_test.cc

@@ -1947,7 +1947,7 @@ TEST(MutableGraphViewTest, SwapRegularFaninsByPorts) {
                                /*to_port=*/2, /*success=*/true, error_msg,
                                {"a", "b:2", "b:2", "^c", "^d"});
 
-  // Swaping fanins at out of bounds ports.
+  // Swapping fanins at out of bounds ports.
   // Node with no regular fanins and no controls.
   error_msg =
       "MutableGraphView::SwapRegularFaninsByPorts(node_name='foo_5', "

tensorflow/core/grappler/optimizers/arithmetic_optimizer.cc

@@ -1044,7 +1044,7 @@ class MinimizeBroadcasts : public ArithmeticNodesGroupOptimizerStage {
       // short tree, pushing to the front will create a tall tree. We prefer to
       // get a wide tree, it minimizes the potential number of temporary tensors
       // required to keep in memory, though sometimes we can go up to prevent
-      // propagating a brodcast from leaves to the root. Example:
+      // propagating a broadcast from leaves to the root. Example:
       //
       // inputs: [s, s, s, M] (s - scalar, M - matrix)
       // @* - op with broadcast
@@ -1466,7 +1466,7 @@ class HoistCWiseUnaryChainsStage : public ArithmeticOptimizerStage {
       }
       if (NumControlOutputs(*node, *ctx().node_map) > 0) {
         // TODO(ezhulenev): Unary ops after Split might have a control path to
-        // the Split node, and we currently do not propertly handle cycles.
+        // the Split node, and we currently do not properly handle cycles.
         return false;
       }
       return num_split > 1 && !IsAlreadyOptimized(*node);
@@ -1509,7 +1509,7 @@ class HoistCWiseUnaryChainsStage : public ArithmeticOptimizerStage {
     // Follow the chains starting at each concat input or split output as long
     // as all the following conditions hold:
     //   1. The ops in all chains are the same.
-    //   2. The ops are unary elemenwise op.
+    //   2. The ops are unary elementwise op.
     //   3. The op output has only a single consumer (concat only).
     ChainLinkSet cur_tails;
     TF_RETURN_IF_ERROR(InitializeChains(root_node, &cur_tails));

tensorflow/core/grappler/optimizers/constant_folding.cc

@@ -1781,7 +1781,7 @@ void ConstantFolding::ReplaceOperationWithSnapshot(
     int input_to_forward, const GraphProperties& properties, NodeDef* node,
     GraphDef* graph) {
   // If the graph contains no ops that mutate their inputs, we can
-  // use Identity insted of Snapshot.
+  // use Identity instead of Snapshot.
   if (!graph_contains_assign_or_inplace_op_) {
     ReplaceOperationWithIdentity(input_to_forward, properties, node, graph);
     return;
@@ -3138,7 +3138,7 @@ bool ConstantFolding::ConstantPushDown(GraphProperties* properties,
       ctx.left_child_is_const ? node->input(1) : node->input(0);
   VLOG(1) << "input_c = " << input_c << "\ninput_x = " << input_x;
 
-  // Now we have identified the nodes to swap, updare the nodemap accordingly.
+  // Now we have identified the nodes to swap, update the nodemap accordingly.
   node_map_->UpdateInput(node->name(), input_c, input_x);
   node_map_->AddOutput(input_c, ctx.op_child->name());
   if (input_x != input_y) {

tensorflow/core/grappler/optimizers/constant_folding_test.cc

@@ -2658,7 +2658,7 @@ TEST_F(ConstantFoldingTest, PaddingWithZeroSize) {
   test::ExpectTensorEqual<int>(tensors_expected[0], tensors[0]);
 }
 
-TEST_F(ConstantFoldingTest, SqueezeWithAllDimesionsGreaterThanOne) {
+TEST_F(ConstantFoldingTest, SqueezeWithAllDimensionsGreaterThanOne) {
   tensorflow::Scope scope = tensorflow::Scope::NewRootScope();
 
   auto in1 = ops::Variable(scope.WithOpName("in1"), {2, 3}, DT_INT32);
@@ -3379,7 +3379,7 @@ TEST_F(ConstantFoldingTest, SwitchIdenticalInputs) {
   EXPECT_EQ(1, tensors.size());
   test::ExpectTensorEqual<bool>(tensors_expected[0], tensors[0]);
 
-  // Evalute id_false when input tensor is false.
+  // Evaluate id_false when input tensor is false.
   x_t.flat<bool>()(0) = false;
   tensors_expected = EvaluateNodes(item.graph, {"id_false"}, {{"x", x_t}});
   EXPECT_EQ(1, tensors_expected.size());
@@ -3689,8 +3689,8 @@ TEST_F(ConstantFoldingTest, TrivialPack) {
   EXPECT_EQ(tensors_expected[1].shape(), tensors[1].shape());
 }
 
-// The test does not evalute the optimized and original graphs to check if their
-// outputs are the same. See b/78233179.
+// The test does not evalaute the optimized and original graphs to check if
+// their outputs are the same. See b/78233179.
 TEST_F(ConstantFoldingTest, Enter) {
   GrapplerItem item;
   AttrValue frame_name;

tensorflow/core/grappler/optimizers/data/latency_all_edges_test.cc

@@ -40,7 +40,7 @@ TEST(LatencyAllEdgesTest, AddLatenciesAfterTensorMapPrefetch) {
   NodeDef map_node = NDef("map_node", "MapDataset",
                           {"from_tensor_node", "captured_input_node"},
                           {{"f", {}},
-                           {"Targumemts", {}},
+                           {"Targuments", {}},
                            {"output_shapes", {}},
                            {"output_types", {}}});
   NodeDef buffer_size_node = NDef("buffer_size_node", "Const", {},

tensorflow/core/grappler/optimizers/function_optimizer.cc

@@ -1177,7 +1177,7 @@ Status InlineFunctionCalls(const GrapplerItem& item,
                            GraphDef* output_graph) {
   bool is_aggressive = opt_level == RewriterConfig::AGGRESSIVE;
   VLOG(2) << "Inline function calls: grappler_item_id=" << item.id
-          << " (aggessive_mode=" << is_aggressive << ")";
+          << " (aggressive_mode=" << is_aggressive << ")";
 
   FunctionLibraryDefinition flib_def =
       FunctionLibraryDefinition(OpRegistry::Global(), item.graph.library());

tensorflow/core/grappler/optimizers/layout_optimizer.cc

@@ -735,7 +735,7 @@ class NodeProcessor : public GraphProcessor {
     if (IsConstant(*param_node)) {
       TF_RETURN_IF_ERROR(UpdateAttrValueOfInput(param_index, permute));
     } else {
-      AddDataFormatTranformToParamInput(op, param_index, dtype);
+      AddDataFormatTransformToParamInput(op, param_index, dtype);
     }
     return Status::OK();
   }
@@ -1038,8 +1038,8 @@ class NodeProcessor : public GraphProcessor {
     return added_node;
   }
 
-  void AddDataFormatTranformToParamInput(const string& op, int input_pos,
-                                         DataType dtype) {
+  void AddDataFormatTransformToParamInput(const string& op, int input_pos,
+                                          DataType dtype) {
     string suffix = (op == "DataFormatVecPermute") ? kVecPermuteNHWCToNCHW
                                                    : kDimMapNHWCToNCHW;
     string name = LayoutOptimizerNode(

tensorflow/core/grappler/optimizers/memory_optimizer.cc

@@ -1147,7 +1147,7 @@ bool SwappingPass(RewriterConfig::MemOptType optimization_level,
     // Use heuristics to figure out what needs to be swapped;
     IdentifySwappingCandidates(cluster, item, skip_list, &nodes_to_swap);
   }
-  // Look for manual annotatations in the graph.
+  // Look for manual annotations in the graph.
   for (auto& node : *item->graph.mutable_node()) {
     if (node.attr().count("_swap_to_host") != 0) {
       SwapInfo& swap_info = nodes_to_swap[&node];

tensorflow/core/grappler/optimizers/memory_optimizer_test.cc

@@ -520,7 +520,7 @@ TEST_F(RelaxAllocatorConstraintsTest, SameDeviceType) {
                                   {128, 128}, DT_FLOAT);
   Output assign = ops::Assign(s.WithOpName("assign").WithDevice("/cpu:0"),
                               variable, constant);
-  // Assign and Exp run on different devies, but do not straddle a CPU:GPU
+  // Assign and Exp run on different devices, but do not straddle a CPU:GPU
   // boundary, so we can we do not need to enforce allocation in pinned memory.
   Output exp = ops::Exp(s.WithOpName("exp").WithDevice("/cpu:1"), assign);
 

tensorflow/core/grappler/optimizers/meta_optimizer.cc

@@ -737,7 +737,7 @@ Status MetaOptimizer::Optimize(Cluster* cluster, const GrapplerItem& item,
         // TPU metadata and Grappler passes could prune that away. Grappler
         // passes could also cause issues around shape inference. Since the
         // desired and existing behavior is to not optimize TPU functions with
-        // Grappler, this check preserves that. The only execption is
+        // Grappler, this check preserves that. The only exception is
         // implementation selector what is required to swap in some TPU specific
         // lowering code and is verified the work correctly on TPUs.
         ImplementationSelector implementation_selector;

tensorflow/core/grappler/optimizers/meta_optimizer.h

@@ -69,7 +69,7 @@ class MetaOptimizer : public GraphOptimizer {
   const RewriterConfig::CustomGraphOptimizer* GetCustomGraphOptimizerConfig(
       const string& name) const;
 
-  // Initialiaze active verifiers from the RewriterConfig toggles.
+  // Initialize active verifiers from the RewriterConfig toggles.
   void InitializeVerifiers(
       std::vector<std::unique_ptr<GraphVerifier>>* inter_optimizer_verifiers,
       std::vector<std::unique_ptr<GraphVerifier>>* post_optimization_verifiers)

tensorflow/core/grappler/optimizers/model_pruner.cc

@@ -299,7 +299,7 @@ Status RewriteIdentityNAndInputsOutputs(
   // Rewrite IdentityN node and associated inputs and outputs. For inputs and
   // outputs that don't lead to a terminal node, a new Identity node is created
   // and those inputs and outputs are rewritten to use the new Identity node as
-  // their outputs and inputs respectively. For the remaining nodes, the ouputs
+  // their outputs and inputs respectively. For the remaining nodes, the outputs
   // have their inputs updated with the adjusted port, from the IdentityN node
   // having less inputs.
   struct NodeOutputUpdate {

tensorflow/core/grappler/optimizers/model_pruner_test.cc

@@ -406,7 +406,7 @@ TEST_F(ModelPrunerTest, PruningForwardsCtrlDependencies) {
 }
 */
 
-TEST_F(ModelPrunerTest, PruningPerservesFetch) {
+TEST_F(ModelPrunerTest, PruningPreservesFetch) {
   // Build a simple graph with a few trivially prunable ops.
   GrapplerItem item;
   {
@@ -445,7 +445,7 @@ TEST_F(ModelPrunerTest, PruningPerservesFetch) {
   test::ExpectTensorEqual<float>(actual_tensors[0], expected_tensors[0]);
 }
 
-TEST_F(ModelPrunerTest, PruningPerservesCrossDeviceIdentity) {
+TEST_F(ModelPrunerTest, PruningPreservesCrossDeviceIdentity) {
   GrapplerItem item;
   {
     tensorflow::Scope s = tensorflow::Scope::NewRootScope();

tensorflow/core/grappler/optimizers/pin_to_host_optimizer.cc

@@ -119,7 +119,7 @@ Status IsNodeOutputPortHostFriendly(const GraphView& graph,
   }
 
   // These nodes may be optimized away downstream (even if pinned to Host), we
-  // should (recusively) check their source.
+  // should (recursively) check their source.
   if (IsIdentity(node) || IsIdentityNSingleInput(node)) {
     for (const auto& fanin : graph.GetFanins(node, false)) {
       bool fanin_candidate = false;

tensorflow/core/grappler/optimizers/scoped_allocator_optimizer.cc

@@ -389,7 +389,7 @@ class UnaryElementwiseRewriter : public ScopedAllocatorOptimizer::Rewriter {
     for (const InputDesc& nd : inputs) {
       if (op_set.find(nd.from_node_def->name()) != op_set.end()) {
         if (nd.output_slot != tensorflow::Graph::kControlSlot) {
-          return errors::Internal("Data edge exists bewtween ",
+          return errors::Internal("Data edge exists between ",
                                   nd.from_node_def->name(),
                                   " and another "
                                   "node in the set");

tensorflow/core/grappler/optimizers/scoped_allocator_optimizer_test.cc

@@ -250,7 +250,7 @@ class ScopedAllocatorOptimizerTest : public ::testing::Test {
   }
 
   // Invokes ScopedAllocatorOptimizer on `graph_def`, then executes it and
-  // returns the outputs specifed by `output_names` in `outputs`.
+  // returns the outputs specified by `output_names` in `outputs`.
   void ExecuteGraph(const GraphDef& graph_def,
                     const std::vector<string>& output_names,
                     std::vector<Tensor>* outputs) {

tensorflow/core/grappler/utils/graph_view.cc

@@ -982,7 +982,7 @@ void MutableGraphView::FixRenamedFanouts(
   // leftover fanouts, mark their respective fanin fanout_index_ to
   // internal::kMissingIndex as an indicator so when it comes to updating or
   // removing fanins inplace, nodes with the same index don't get affected and
-  // other fanouts are accidently removed.
+  // other fanouts are accidentally removed.
   for (auto& renamed_fanout : renamed_fanouts) {
     for (auto& regular_fanouts :
          renamed_fanout.second.regular_fanouts_by_port_) {

tensorflow/core/grappler/utils/grappler_test.h

@@ -55,13 +55,13 @@ class GrapplerTest : public ::testing::Test {
   // with the same inputs and attributes. Nodes can be in different order.
   //
   // NOTE: This function uses EXPECT/ASSERT macros to check node properties
-  // equality, and adds all failuires to the current test.
+  // equality, and adds all failures to the current test.
   void CompareGraphs(GraphDef want, GraphDef got) const;
 
   // Checks if two nodes have the same name, op, inputs and attributes.
   //
   // NOTE: This function uses EXPECT/ASSERT macros to check node properties
-  // equality, and adds all failuires to the current test.
+  // equality, and adds all failures to the current test.
   void CompareNodes(const NodeDef& want, const NodeDef& got) const;
 
   // Checks if two functions are equal. Both functions must have the same set of

tensorflow/core/kernels/adjust_contrast_op.cc

@@ -273,7 +273,7 @@ class AdjustContrastOpv2<CPUDevice, float> : public AdjustContrastOpV2Base {
       //
       // The algorithm itself can handle size that is not power-of-two. Note
       // that in each round we sum up elements that are contiguous. So we can
-      // use their flattened structure to gain vectorinization efficiency.
+      // use their flattened structure to gain vectorization efficiency.
       do {
         int64 right_size = remaining_size / 2;
         int64 left_size = remaining_size - right_size;
@@ -324,7 +324,7 @@ class AdjustContrastOpv2<CPUDevice, float> : public AdjustContrastOpV2Base {
     // Similar to the reduction case, a straightforward implementation of this
     // does not utilize vectorization well because of the small channel size.
     // This algorithm repeatedly increases the area to be copied, and leads to
-    // much better vectorinizations in the copy.
+    // much better vectorizations in the copy.
     for (int64 i = 0; i < batch; i++) {
       // Copy over the inputs into outputs in this batch. Effectively:
       // outputs(i, :, k) = inputs(i, k). An example of how this algorithm
@@ -354,7 +354,7 @@ class AdjustContrastOpv2<CPUDevice, float> : public AdjustContrastOpV2Base {
       int64 copied = 1;
       while (copied < image_size) {
         // Repeatedly increases the number of elements to copy so they have
-        // better vectorinizations. However, the source of the copy has to be
+        // better vectorizations. However, the source of the copy has to be
         // not too large to stay in the cache.
         const int64 kMaxToCopy = 1024;
         int64 to_copy = std::min({copied, image_size - copied, kMaxToCopy});

tensorflow/core/kernels/batching_util/adaptive_shared_batch_scheduler.h

@@ -409,7 +409,7 @@ template <typename TaskType>
 void AdaptiveSharedBatchScheduler<TaskType>::MaybeScheduleNextBatch() {
   if (batches_.empty() || in_flight_batches_ >= in_flight_batches_limit_)
     return;
-  // Non-integer limit handled probabilistially.
+  // Non-integer limit handled probabilistically.
   if (in_flight_batches_limit_ - in_flight_batches_ < 1 &&
       rand_double_(rand_engine_) >
           in_flight_batches_limit_ - in_flight_batches_) {

tensorflow/core/kernels/batching_util/shared_batch_scheduler.h

@@ -234,7 +234,7 @@ class Queue {
   using SchedulableBatchCallback = std::function<void()>;
   Queue(const typename SharedBatchScheduler<TaskType>::QueueOptions& options,
         Env* env, ProcessBatchCallback process_batch_callback,
-        SchedulableBatchCallback schdulable_batch_callback);
+        SchedulableBatchCallback schedulable_batch_callback);
 
   // Illegal to destruct unless the queue is empty.
   ~Queue();

tensorflow/core/kernels/boosted_trees/quantile_ops.cc

@@ -104,7 +104,7 @@ class BoostedTreesCreateQuantileStreamResourceOp : public OpKernel {
   void Compute(OpKernelContext* context) override {
     // Only create one, if one does not exist already. Report status for all
     // other exceptions. If one already exists, it unrefs the new one.
-    // An epsilon value of zero could cause perfoamance issues and is therefore,
+    // An epsilon value of zero could cause performance issues and is therefore,
     // disallowed.
     const Tensor* epsilon_t;
     OP_REQUIRES_OK(context, context->input(kEpsilonName, &epsilon_t));

tensorflow/core/kernels/boosted_trees/quantiles/weighted_quantiles_stream.h

@@ -58,7 +58,7 @@ namespace quantiles {
 // Compute: O(n * log(1/eps * log(eps * n))).
 // Memory: O(1/eps * log^2(eps * n)) <- for one worker streaming through the
 //                                      entire dataset.
-// An epsilon value of zero would make the algorithm extremely inefficent and
+// An epsilon value of zero would make the algorithm extremely inefficient and
 // therefore, is disallowed.
 template <typename ValueType, typename WeightType,
           typename CompareFn = std::less<ValueType>>
@@ -72,7 +72,7 @@ class WeightedQuantilesStream {
   explicit WeightedQuantilesStream(double eps, int64 max_elements)
       : eps_(eps), buffer_(1LL, 2LL), finalized_(false) {
     // See the class documentation. An epsilon value of zero could cause
-    // perfoamance issues.
+    // performance issues.
     QCHECK(eps > 0) << "An epsilon value of zero is not allowed.";
     std::tie(max_levels_, block_size_) = GetQuantileSpecs(eps, max_elements);
     buffer_ = Buffer(block_size_, max_elements);

tensorflow/core/kernels/boosted_trees/resources.cc

@@ -340,7 +340,7 @@ boosted_trees::Node* BoostedTreesEnsembleResource::AddLeafNodes(
         node->mutable_leaf());
   }
   node->mutable_metadata()->set_gain(candidate.gain);
-  // TODO(npononareva): this is LAYER-BY-LAYER boosting; add WHOLE-TREE.
+  // TODO(nponomareva): this is LAYER-BY-LAYER boosting; add WHOLE-TREE.
   if (logits_dimension == 1) {
     const float prev_logit_value = node->metadata().original_leaf().scalar();
     left_node->mutable_leaf()->set_scalar(prev_logit_value +

tensorflow/core/kernels/check_numerics_op_gpu.cu.cc

@@ -55,7 +55,7 @@ __global__ void CheckNumericsKernel(const T* __restrict__ data, int size,
 }
 
 // V2 of CheckNumericsKernel for GPU.
-// Unlike CheckNumericsKernel (V1), this kernel disinguishes -Inf and +Inf.
+// Unlike CheckNumericsKernel (V1), this kernel distinguishes -Inf and +Inf.
 // The 3 elements of `abnormal_detected` are used to signify NaN, -Inf and +Inf,
 // respectively.
 template <typename T>

tensorflow/core/kernels/collective_nccl_broadcaster.cc

@@ -68,7 +68,7 @@ void NcclBroadcaster::Run(StatusCallback done) {
   {
     // When all devices at this worker have called `SignalMultiNodeReady`, the
     // `NcclManager` will enqueue the NCCL kernel on the NCCL stream.  Thus the
-    // implementation of `UnblockDepdendencies` keeps track of the number of
+    // implementation of `UnblockDependencies` keeps track of the number of
     // devices that have launched.
     profiler::TraceMe activity("Schedule", profiler::TraceMeLevel::kInfo);
     col_ctx_->col_exec->UnblockDependencies(*col_params_);

tensorflow/core/kernels/conv_ops_fused_impl.h

@@ -299,7 +299,7 @@ inline int64 ConvolveScratchSize() {
   return convolve_scratch_size;
 }
 
-// Finds the best convolutiun algorithm for the given ConvLaunch (cuda
+// Finds the best convolution algorithm for the given ConvLaunch (cuda
 // convolution on the stream) and parameters, by running all possible
 // algorithms and measuring execution time.
 // TODO(ezhulenev): Move it to conv_ops_gpu.h and share with conv_ops.cc.

tensorflow/core/kernels/cuda_sparse.h

@@ -286,7 +286,7 @@ class GpuSparse {
   //
   // **NOTE** Matrices B and C are expected to be in column-major
   // order; to make them consistent with TensorFlow they
-  // must be transposed (or the matmul op's pre/post-procesisng must take this
+  // must be transposed (or the matmul op's pre/post-processing must take this
   // into account).
   //
   // **NOTE** This is an in-place operation for data in C.

tensorflow/core/kernels/cwise_op_add_1.cc

@@ -47,7 +47,7 @@ REGISTER_KERNEL_BUILDER(Name("AddV2")
 #if TENSORFLOW_USE_SYCL
 #define REGISTER_KERNEL(type)                          \
   REGISTER(BinaryOp, SYCL, "Add", functor::add, type); \
-  REEGISTER(BinaryOp, SYCL, "AddV2", functor::add, type);
+  REGISTER(BinaryOp, SYCL, "AddV2", functor::add, type);
 
 TF_CALL_SYCL_NUMBER_TYPES(REGISTER_KERNEL);
 

tensorflow/core/kernels/cwise_op_select.cc

@@ -348,8 +348,8 @@ struct SelectScalarHandler {
   }
 };
 
-// Specilization for CPU device. Forward input to output depending on the `cond`
-// value.
+// Specialization for CPU device. Forward input to output depending on the
+// `cond` value.
 // TODO(sjhwang): Consider specializing for GPUDevice as well by using
 // GPUDevice::memcpyDeviceToHost() to fetch bool value.
 template <typename T>

tensorflow/core/kernels/cwise_ops_gpu_gradients.cu.h

@@ -49,7 +49,7 @@ struct SimpleBinaryFunctor<GPUDevice, Functor> {
 };
 
 // Macros to explicitly instantiate kernels on GPU for multiple types
-// (T0, T1, etc.) for SimpleBiaryFunctor (e.g., functor::tanh_grad).
+// (T0, T1, etc.) for SimpleBinaryFunctor (e.g., functor::tanh_grad).
 #define DEFINE_SIMPLE_BINARY1(F, T) \
   template struct SimpleBinaryFunctor<GPUDevice, F<T> >
 #define DEFINE_SIMPLE_BINARY2(F, T0, T1) \

tensorflow/core/kernels/data/captured_function.cc

@@ -910,7 +910,7 @@ Status CapturedFunction::IsMultiDevice(IteratorContext* ctx,
     const FunctionDef* fdef;
     TF_RETURN_IF_ERROR(LookupFunction(*metadata_->lib_def(), name, &fdef));
     for (const auto& node : fdef->node_def()) {
-      // Check if the op has a kernel availabe for the current device.
+      // Check if the op has a kernel available for the current device.
       if (!KernelDefAvailable(current_device_type, node)) {
         *is_multi_device = true;
         return Status::OK();