Implementation:Microsoft Onnxruntime Summary Ops

Field	Value
Implementation Name	Summary_Ops
Overview	TensorBoard summary operators for logging scalar, histogram, text, and merged training metrics during distributed training.
Type	API Doc
Language	C++
Domains	Distributed_Training, Training_Infrastructure
Source Repository	microsoft/onnxruntime
Last Updated	2026-02-10

Overview

TensorBoard summary operators for logging scalar, histogram, text, and merged training metrics during distributed training. These operators (SummaryScalarOp, SummaryHistogramOp, SummaryTextOp, SummaryMergeOp) are registered in the Microsoft domain and produce TensorBoard-compatible protobuf summary events.

API

// All operators implement the OpKernel::Compute interface
class SummaryScalarOp final : public OpKernel {
    Status Compute(OpKernelContext* context) const override;
};

class SummaryHistogramOp final : public OpKernel {
    Status Compute(OpKernelContext* context) const override;
};

class SummaryTextOp final : public OpKernel {
    Status Compute(OpKernelContext* context) const override;
};

class SummaryMergeOp final : public OpKernel {
    Status Compute(OpKernelContext* context) const override;
};

Source Code Reference

Repository: microsoft/onnxruntime
Header: orttraining/orttraining/training_ops/cpu/tensorboard/summary_op.h:L13-54
Implementation: orttraining/orttraining/training_ops/cpu/tensorboard/summary_op.cc:L1-284

Key Parameters

Operator	Input	Type Constraint	Description
SummaryScalarOp	tag (Tensor)	string	Tag name(s) identifying the scalar summary
SummaryScalarOp	input (Tensor)	float, double, bool	Scalar value(s) to log
SummaryHistogramOp	tag (Tensor)	string	Tag name identifying the histogram
SummaryHistogramOp	input (Tensor)	float, double	Tensor values for histogram distribution
SummaryTextOp	tag (Tensor)	string	Tag name identifying the text entry
SummaryTextOp	input (Tensor)	string	Text content to log
SummaryMergeOp	inputs (Tensor[])	string	Multiple serialized summary protobufs to merge

I/O Contract

Direction	Name	Type	Description
Input	tag	Tensor (string)	Identifying name/tag for the summary entry
Input	input	Tensor (varies)	Data to be serialized into the summary
Output	summary	Tensor (string)	Serialized TensorBoard summary protobuf bytes

Usage Examples

Configuring TensorBoard in TrainingRunner

// Summary ops are configured via TrainingRunner::Parameters
TrainingRunner::Parameters params;
params.log_dir = ORT_TSTR("/logs/tensorboard/");
params.summary_name = "training_summary";
params.scalar_names = {"total_loss", "learning_rate"};
params.histogram_names = {"layer1_weights", "layer2_gradients"};
params.norm_names = {"gradient_norm"};

TensorBoard Configuration in TrainingSession

// During Initialize(), TensorBoard is configured on the training session
if (params_.EnableTensorboard()) {
    TrainingSession::TrainingConfiguration::TensorboardConfiguration tb{};
    tb.summary_name = params_.summary_name;
    tb.scalar_node_names = params_.scalar_names;
    tb.histogram_node_names = params_.histogram_names;
    tb.norm_node_names = params_.norm_names;
    tb.dump_convergence_metrics = params_.dump_convergence_metrics;
    config.tensorboard_config = tb;
}

Viewing TensorBoard Output

# Launch TensorBoard to view logged training metrics
tensorboard --logdir /logs/tensorboard/

Operator Registration

The summary operators are registered in the Microsoft domain (kMSDomain) using the ONNX Runtime kernel registration macros:

ONNX_OPERATOR_KERNEL_EX(
    SummaryScalar, kMSDomain, 1, kCpuExecutionProvider,
    KernelDefBuilder()
        .TypeConstraint("T", {DataTypeImpl::GetTensorType<float>(),
                              DataTypeImpl::GetTensorType<double>(),
                              DataTypeImpl::GetTensorType<bool>()})
        .TypeConstraint("S", DataTypeImpl::GetTensorType<std::string>()),
    SummaryScalarOp);

ONNX_OPERATOR_KERNEL_EX(
    SummaryHistogram, kMSDomain, 1, kCpuExecutionProvider,
    KernelDefBuilder()
        .TypeConstraint("T", {DataTypeImpl::GetTensorType<float>(),
                              DataTypeImpl::GetTensorType<double>()})
        .TypeConstraint("S", DataTypeImpl::GetTensorType<std::string>()),
    SummaryHistogramOp);

Key Details

All summary operators run on the kCpuExecutionProvider regardless of the training device, since serialization is a CPU operation.
SummaryScalarOp maintains a tags_ vector for multiple scalar entries in a single op invocation.
SummaryHistogramOp and SummaryTextOp each have a single tag_ string.
The operators use TensorBoard's native protobuf format (tensorboard/compat/proto/summary.pb.h) for serialization.
SummaryMergeOp concatenates multiple summary protobufs into a single merged summary for efficient event writing.
TensorBoard logging only occurs on rank 0 (checked by EnableTensorboard() which verifies MPIContext::GetInstance().GetWorldRank() == 0).
The implementation is adapted from TensorFlow's summary ops (licensed under Apache 2.0) with modifications for ONNX Runtime.

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment