Rate this Page
★ ★ ★ ★ ★

torch.onnx#

Created On: Jun 10, 2025 | Last Updated On: Jul 29, 2025

Overview#

Open Neural Network eXchange (ONNX) is an open standard format for representing machine learning models. The torch.onnx module captures the computation graph from a native PyTorch torch.nn.Module model and converts it into an ONNX graph.

The exported model can be consumed by any of the many runtimes that support ONNX, including Microsoft’s ONNX Runtime.

Next example shows how to export a simple model.

import torch

class MyModel(torch.nn.Module):
    def __init__(self):
        super(MyModel, self).__init__()
        self.conv1 = torch.nn.Conv2d(1, 128, 5)

    def forward(self, x):
        return torch.relu(self.conv1(x))

input_tensor = torch.rand((1, 1, 128, 128), dtype=torch.float32)

model = MyModel()

torch.onnx.export(
    model,                  # model to export
    (input_tensor,),        # inputs of the model,
    "my_model.onnx",        # filename of the ONNX model
    input_names=["input"],  # Rename inputs for the ONNX model
    dynamo=True             # True or False to select the exporter to use
)

torch.export-based ONNX Exporter#

The torch.export-based ONNX exporter is the newest exporter for PyTorch 2.6 and newer

torch.export engine is leveraged to produce a traced graph representing only the Tensor computation of the function in an Ahead-of-Time (AOT) fashion. The resulting traced graph (1) produces normalized operators in the functional ATen operator set (as well as any user-specified custom operators), (2) has eliminated all Python control flow and data structures (with certain exceptions), and (3) records the set of shape constraints needed to show that this normalization and control-flow elimination is sound for future inputs, before it is finally translated into an ONNX graph.

Learn more about the torch.export-based ONNX Exporter

Frequently Asked Questions#

Q: I have exported my LLM model, but its input size seems to be fixed?

The tracer records the shapes of the example inputs. If the model should accept inputs of dynamic shapes, set dynamic_shapes when calling torch.onnx.export().

Q: How to export models containing loops?

See torch.cond.

Contributing / Developing#

The ONNX exporter is a community project and we welcome contributions. We follow the PyTorch guidelines for contributions, but you might also be interested in reading our development wiki.

torch.onnx APIs#

Functions#

torch.onnx.export(model, args=(), f=None, *, kwargs=None, export_params=True, verbose=None, input_names=None, output_names=None, opset_version=None, dynamic_axes=None, keep_initializers_as_inputs=False, dynamo=False, external_data=True, dynamic_shapes=None, custom_translation_table=None, report=False, optimize=True, verify=False, profile=False, dump_exported_program=False, artifacts_dir='.', fallback=False, training=<TrainingMode.EVAL: 0>, operator_export_type=<OperatorExportTypes.ONNX: 0>, do_constant_folding=True, custom_opsets=None, export_modules_as_functions=False, autograd_inlining=True)[source]

Exports a model into ONNX format.

Setting dynamo=True enables the new ONNX export logic which is based on torch.export.ExportedProgram and a more modern set of translation logic. This is the recommended way to export models to ONNX.

When dynamo=True:

The exporter tries the following strategies to get an ExportedProgram for conversion to ONNX.

  1. If the model is already an ExportedProgram, it will be used as-is.

  2. Use torch.export.export() and set strict=False.

  3. Use torch.export.export() and set strict=True.

  4. Use draft_export which removes some soundness guarantees in data-dependent operations to allow export to proceed. You will get a warning if the exporter encounters any unsound data-dependent operation.

  5. Use torch.jit.trace() to trace the model then convert to ExportedProgram. This is the most unsound strategy but may be useful for converting TorchScript models to ONNX.

Parameters

  • model (torch.nn.Module | torch.export.ExportedProgram | torch.jit.ScriptModule | torch.jit.ScriptFunction) – The model to be exported.

  • args (tuple[Any, ...]) – Example positional inputs. Any non-Tensor arguments will be hard-coded into the exported model; any Tensor arguments will become inputs of the exported model, in the order they occur in the tuple.

  • f (str | os.PathLike | None) – Path to the output ONNX model file. E.g. “model.onnx”.

  • kwargs (dict[str, Any] | None) – Optional example keyword inputs.

  • export_params (bool) – If false, parameters (weights) will not be exported.

  • verbose (bool | None) – Whether to enable verbose logging.

  • input_names (Sequence[str] | None) – names to assign to the input nodes of the graph, in order.

  • output_names (Sequence[str] | None) – names to assign to the output nodes of the graph, in order.

  • opset_version (int | None) – The version of the default (ai.onnx) opset to target. Must be >= 7.

  • dynamic_axes (Mapping[str, Mapping[int, str]] | Mapping[str, Sequence[int]] | None) –

    By default the exported model will have the shapes of all input and output tensors set to exactly match those given in args. To specify axes of tensors as dynamic (i.e. known only at run-time), set dynamic_axes to a dict with schema:

    • KEY (str): an input or output name. Each name must also be provided in input_names or

      output_names.

    • VALUE (dict or list): If a dict, keys are axis indices and values are axis names. If a

      list, each element is an axis index.

    For example:

    class SumModule(torch.nn.Module):
    def forward(self, x):
        return torch.sum(x, dim=1)


torch.onnx.export(
    SumModule(),
    (torch.ones(2, 2),),
    "onnx.pb",
    input_names=["x"],
    output_names=["sum"],
)

    Produces:

    input {
  name: "x"
  ...
      shape {
        dim {
          dim_value: 2  # axis 0
        }
        dim {
          dim_value: 2  # axis 1
...
output {
  name: "sum"
  ...
      shape {
        dim {
          dim_value: 2  # axis 0
...

    While:

    torch.onnx.export(
    SumModule(),
    (torch.ones(2, 2),),
    "onnx.pb",
    input_names=["x"],
    output_names=["sum"],
    dynamic_axes={
        # dict value: manually named axes
        "x": {0: "my_custom_axis_name"},
        # list value: automatic names
        "sum": [0],
    },
)

    Produces:

    input {
  name: "x"
  ...
      shape {
        dim {
          dim_param: "my_custom_axis_name"  # axis 0
        }
        dim {
          dim_value: 2  # axis 1
...
output {
  name: "sum"
  ...
      shape {
        dim {
          dim_param: "sum_dynamic_axes_1"  # axis 0
...

  • keep_initializers_as_inputs (bool) –

    If True, all the initializers (typically corresponding to model weights) in the exported graph will also be added as inputs to the graph. If False, then initializers are not added as inputs to the graph, and only the user inputs are added as inputs.

    Set this to True if you intend to supply model weights at runtime. Set it to False if the weights are static to allow for better optimizations (e.g. constant folding) by backends/runtimes.

  • dynamo (bool) – Whether to export the model with torch.export ExportedProgram instead of TorchScript.

  • external_data (bool) – Whether to save the model weights as an external data file. This is required for models with large weights that exceed the ONNX file size limit (2GB). When False, the weights are saved in the ONNX file with the model architecture.

  • dynamic_shapes (dict[str, Any] | tuple[Any, ...] | list[Any] | None) – A dictionary or a tuple of dynamic shapes for the model inputs. Refer to torch.export.export() for more details. This is only used (and preferred) when dynamo is True. Note that dynamic_shapes is designed to be used when the model is exported with dynamo=True, while dynamic_axes is used when dynamo=False.

  • custom_translation_table (dict[Callable, Callable | Sequence[Callable]] | None) – A dictionary of custom decompositions for operators in the model. The dictionary should have the callable target in the fx Node as the key (e.g. torch.ops.aten.stft.default), and the value should be a function that builds that graph using ONNX Script. This option is only valid when dynamo is True.

  • report (bool) – Whether to generate a markdown report for the export process. This option is only valid when dynamo is True.

  • optimize (bool) – Whether to optimize the exported model. This option is only valid when dynamo is True. Default is True.

  • verify (bool) – Whether to verify the exported model using ONNX Runtime. This option is only valid when dynamo is True.

  • profile (bool) – Whether to profile the export process. This option is only valid when dynamo is True.

  • dump_exported_program (bool) – Whether to dump the torch.export.ExportedProgram to a file. This is useful for debugging the exporter. This option is only valid when dynamo is True.

  • artifacts_dir (str | os.PathLike) – The directory to save the debugging artifacts like the report and the serialized exported program. This option is only valid when dynamo is True.

  • fallback (bool) – Whether to fallback to the TorchScript exporter if the dynamo exporter fails. This option is only valid when dynamo is True. When fallback is enabled, It is recommended to set dynamic_axes even when dynamic_shapes is provided.

  • training (_C_onnx.TrainingMode) – Deprecated option. Instead, set the training mode of the model before exporting.

  • operator_export_type (_C_onnx.OperatorExportTypes) – Deprecated option. Only ONNX is supported.

  • do_constant_folding (bool) – Deprecated option.

  • custom_opsets (Mapping[str, int] | None) –

    Deprecated. A dictionary:

    • KEY (str): opset domain name

    • VALUE (int): opset version

    If a custom opset is referenced by model but not mentioned in this dictionary, the opset version is set to 1. Only custom opset domain name and version should be indicated through this argument.

  • export_modules_as_functions (bool | Collection[type[torch.nn.Module]]) –

    Deprecated option.

    Flag to enable exporting all nn.Module forward calls as local functions in ONNX. Or a set to indicate the particular types of modules to export as local functions in ONNX. This feature requires opset_version >= 15, otherwise the export will fail. This is because opset_version < 15 implies IR version < 8, which means no local function support. Module variables will be exported as function attributes. There are two categories of function attributes.

    1. Annotated attributes: class variables that have type annotations via PEP 526-style will be exported as attributes. Annotated attributes are not used inside the subgraph of ONNX local function because they are not created by PyTorch JIT tracing, but they may be used by consumers to determine whether or not to replace the function with a particular fused kernel.

    2. Inferred attributes: variables that are used by operators inside the module. Attribute names will have prefix “inferred::”. This is to differentiate from predefined attributes retrieved from python module annotations. Inferred attributes are used inside the subgraph of ONNX local function.

    • False (default): export nn.Module forward calls as fine grained nodes.

    • True: export all nn.Module forward calls as local function nodes.

    • Set of type of nn.Module: export nn.Module forward calls as local function nodes,

      only if the type of the nn.Module is found in the set.

  • autograd_inlining (bool) – Deprecated. Flag used to control whether to inline autograd functions. Refer to pytorch/pytorch#74765 for more details.

Returns

torch.onnx.ONNXProgram if dynamo is True, otherwise None.

Return type

ONNXProgram | None

Changed in version 2.6: training is now deprecated. Instead, set the training mode of the model before exporting. operator_export_type is now deprecated. Only ONNX is supported. do_constant_folding is now deprecated. It is always enabled. export_modules_as_functions is now deprecated. autograd_inlining is now deprecated.

Changed in version 2.7: optimize is now True by default.

torch.onnx.is_in_onnx_export()[source]

Returns whether it is in the middle of ONNX export.

Return type

bool

torch.onnx.enable_fake_mode()[source]

Enable fake mode for the duration of the context.

Internally it instantiates a torch._subclasses.fake_tensor.FakeTensorMode context manager that converts user input and model parameters into torch._subclasses.fake_tensor.FakeTensor.

A torch._subclasses.fake_tensor.FakeTensor is a torch.Tensor with the ability to run PyTorch code without having to actually do computation through tensors allocated on a meta device. Because there is no actual data being allocated on the device, this API allows for initializing and exporting large models without the actual memory footprint needed for executing it.

It is highly recommended to initialize the model in fake mode when exporting models that are too large to fit into memory.

Note

This function does not support torch.onnx.export(…, dynamo=True, optimize=True). Please call ONNXProgram.optimize() outside of the function after the model is exported.

Example:

# xdoctest: +REQUIRES(env:TORCH_DOCTEST_ONNX)
>>> import torch
>>> class MyModel(torch.nn.Module):  # Model with a parameter
...     def __init__(self) -> None:
...         super().__init__()
...         self.weight = torch.nn.Parameter(torch.tensor(42.0))
...     def forward(self, x):
...         return self.weight + x
>>> with torch.onnx.enable_fake_mode():
...     # When initialized in fake mode, the model's parameters are fake tensors
...     # They do not take up memory so we can initialize large models
...     my_nn_module = MyModel()
...     arg1 = torch.randn(2, 2, 2)
>>> onnx_program = torch.onnx.export(my_nn_module, (arg1,), dynamo=True, optimize=False)
>>> # Saving model WITHOUT initializers (only the architecture)
>>> onnx_program.save(
...     "my_model_without_initializers.onnx",
...     include_initializers=False,
...     keep_initializers_as_inputs=True,
... )
>>> # Saving model WITH initializers after applying concrete weights
>>> onnx_program.apply_weights({"weight": torch.tensor(42.0)})
>>> onnx_program.save("my_model_with_initializers.onnx")

Warning

This API is experimental and is NOT backward-compatible.

Classes#

class torch.onnx.ONNXProgram(model, exported_program)

A class to represent an ONNX program that is callable with torch tensors.

Variables

  • model – The ONNX model as an ONNX IR model object.

  • exported_program – The exported program that produced the ONNX model.

class torch.onnx.OnnxExporterError

Errors raised by the ONNX exporter. This is the base class for all exporter errors.

Deprecated APIs#

Deprecated since version 2.6: These functions are deprecated and will be removed in a future version.

torch.onnx.register_custom_op_symbolic(symbolic_name, symbolic_fn, opset_version)[source]#

Registers a symbolic function for a custom operator.

When the user registers symbolic for custom/contrib ops, it is highly recommended to add shape inference for that operator via setType API, otherwise the exported graph may have incorrect shape inference in some extreme cases. An example of setType is test_aten_embedding_2 in test_operators.py.

See “Custom Operators” in the module documentation for an example usage.

Parameters

  • symbolic_name (str) – The name of the custom operator in “<domain>::<op>” format.

  • symbolic_fn (Callable) – A function that takes in the ONNX graph and the input arguments to the current operator, and returns new operator nodes to add to the graph.

  • opset_version (int) – The ONNX opset version in which to register.

torch.onnx.unregister_custom_op_symbolic(symbolic_name, opset_version)[source]#

Unregisters symbolic_name.

See “Custom Operators” in the module documentation for an example usage.

Parameters

  • symbolic_name (str) – The name of the custom operator in “<domain>::<op>” format.

  • opset_version (int) – The ONNX opset version in which to unregister.

torch.onnx.select_model_mode_for_export(model, mode)[source]#

A context manager to temporarily set the training mode of model to mode, resetting it when we exit the with-block.

Deprecated since version 2.7: Please set training mode before exporting the model.

Parameters

  • model – Same type and meaning as model arg to export().

  • mode (TrainingMode) – Same type and meaning as training arg to export().

class torch.onnx.JitScalarType(value)#

Scalar types defined in torch.

Use JitScalarType to convert from torch and JIT scalar types to ONNX scalar types.

Examples

>>> JitScalarType.from_value(torch.ones(1, 2)).onnx_type()
TensorProtoDataType.FLOAT

>>> JitScalarType.from_value(torch_c_value_with_type_float).onnx_type()
TensorProtoDataType.FLOAT

>>> JitScalarType.from_dtype(torch.get_default_dtype).onnx_type()
TensorProtoDataType.FLOAT
On this page
Edit on GitHub
Show Source
PyTorch Libraries