.. DO NOT EDIT.
.. THIS FILE WAS AUTOMATICALLY GENERATED BY SPHINX-GALLERY.
.. TO MAKE CHANGES, EDIT THE SOURCE PYTHON FILE:
.. "tutorials/export-to-executorch-tutorial.py"
.. LINE NUMBERS ARE GIVEN BELOW.
.. only:: html
.. note::
:class: sphx-glr-download-link-note
:ref:`Go to the end <sphx_glr_download_tutorials_export-to-executorch-tutorial.py>`
to download the full example code
.. rst-class:: sphx-glr-example-title
.. _sphx_glr_tutorials_export-to-executorch-tutorial.py:
Exporting to ExecuTorch Tutorial
================================
**Author:** `Angela Yi <https://github.com/angelayi>`__
.. GENERATED FROM PYTHON SOURCE LINES 16-29
ExecuTorch is a unified ML stack for lowering PyTorch models to edge devices.
It introduces improved entry points to perform model, device, and/or use-case
specific optizations such as backend delegation, user-defined compiler
transformations, default or user-defined memory planning, and more.
At a high level, the workflow looks as follows:
.. image:: ../executorch_stack.png
:width: 560
In this tutorial, we will cover the APIs in the "Program preparation" steps to
lower a PyTorch model to a format which can be loaded to device and run on the
ExecuTorch runtime.
.. GENERATED FROM PYTHON SOURCE LINES 31-50
Prerequisites
-------------
To run this tutorial, you’ll need to install ExecuTorch.
Set up a conda environment. To set up a conda environment in Google Colab::
!pip install -q condacolab
import condacolab
condacolab.install()
!conda create --name executorch python=3.10
!conda install -c conda-forge flatbuffers
Install ExecuTorch from source. If cloning is failing on Google Colab, make
sure Colab -> Setting -> Github -> Access Private Repo is checked::
!git clone https://{github_username}:{token}@github.com/pytorch/executorch.git
!cd executorch && bash ./install_requirements.sh
.. GENERATED FROM PYTHON SOURCE LINES 52-67
Exporting a Model
-----------------
Note: The Export APIs are still undergoing changes to align better with the
longer term state of export. Please refer to this
`issue <https://github.com/pytorch/executorch/issues/290>`__ for more details.
The first step of lowering to ExecuTorch is to export the given model (any
callable or ``torch.nn.Module``) to a graph representation. This is done via
the two-stage APIs, ``torch._export.capture_pre_autograd_graph``, and
``torch.export``.
Both APIs take in a model (any callable or ``torch.nn.Module``), a tuple of
positional arguments, optionally a dictionary of keywork arguments (not shown
in the example), and a list of constraints (covered later).
.. GENERATED FROM PYTHON SOURCE LINES 67-95
.. code-block:: default
import torch
from torch._export import capture_pre_autograd_graph
from torch.export import export, ExportedProgram
class M(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.conv = torch.nn.Conv2d(
in_channels=3, out_channels=16, kernel_size=3, padding=1
)
self.relu = torch.nn.ReLU()
def forward(self, x: torch.Tensor) -> torch.Tensor:
a = self.conv(x)
return self.relu(a)
example_args = (torch.randn(1, 3, 256, 256),)
pre_autograd_aten_dialect = capture_pre_autograd_graph(M(), example_args)
print("Pre-Autograd ATen Dialect Graph")
print(pre_autograd_aten_dialect)
aten_dialect: ExportedProgram = export(pre_autograd_aten_dialect, example_args)
print("ATen Dialect Graph")
print(aten_dialect)
.. GENERATED FROM PYTHON SOURCE LINES 96-121
The output of ``torch._export.capture_pre_autograd_graph`` is a fully
flattened graph (meaning the graph does not contain any module hierarchy,
except in the case of control flow operators). Furthermore, the captured graph
contains only ATen operators (~3000 ops) which are autograd safe, i.e. safe
for eager mode training.
The output of ``torch.export`` further compiles the graph to a lower and
cleaner representation. Specifically, it has the following:
- The graph is purely functional, meaning it does not contain operations with
side effects such as mutations or aliasing.
- The graph contains only a small defined
`Core ATen IR <https://pytorch.org/docs/stable/ir.html#core-aten-ir>`__
operator set (~180 ops), along with registered custom operators.
- The nodes in the graph contain metadata captured during tracing, such as a
stacktrace from user's code.
More specifications about the result of ``torch.export`` can be found
`here <https://pytorch.org/docs/2.1/export.html>`__.
Since the result of ``torch.export`` is a graph containing the Core ATen
operators, we will call this the ``ATen Dialect``, and since
``torch._export.capture_pre_autograd_graph`` returns a graph containing the
set of ATen operators which are autograd safe, we will call it the
``Pre-Autograd ATen Dialect``.
.. GENERATED FROM PYTHON SOURCE LINES 123-129
Expressing Dynamism
^^^^^^^^^^^^^^^^^^^
By default, the exporting flow will trace the program assuming that all input
shapes are static, so if we run the program with inputs shapes that are
different than the ones we used while tracing, we will run into an error:
.. GENERATED FROM PYTHON SOURCE LINES 129-150
.. code-block:: default
import traceback as tb
def f(x, y):
return x + y
example_args = (torch.randn(3, 3), torch.randn(3, 3))
pre_autograd_aten_dialect = capture_pre_autograd_graph(f, example_args)
aten_dialect: ExportedProgram = export(f, example_args)
# Works correctly
print(aten_dialect(torch.ones(3, 3), torch.ones(3, 3)))
# Errors
try:
print(aten_dialect(torch.ones(3, 2), torch.ones(3, 2)))
except Exception:
tb.print_exc()
.. GENERATED FROM PYTHON SOURCE LINES 151-153
To express that some input shapes are dynamic, we can insert constraints to
the exporting flow. This is done through the ``dynamic_dim`` API:
.. GENERATED FROM PYTHON SOURCE LINES 153-179
.. code-block:: default
from torch.export import dynamic_dim
def f(x, y):
return x + y
example_args = (torch.randn(3, 3), torch.randn(3, 3))
constraints = [
# Input 0, dimension 1 is dynamic
dynamic_dim(example_args[0], 1),
# Input 0, dimension 1 must be greater than or equal to 1
1 <= dynamic_dim(example_args[0], 1),
# Input 0, dimension 1 must be less than or equal to 10
dynamic_dim(example_args[0], 1) <= 10,
# Input 1, dimension 1 is equal to input 0, dimension 1
dynamic_dim(example_args[1], 1) == dynamic_dim(example_args[0], 1),
]
pre_autograd_aten_dialect = capture_pre_autograd_graph(
f, example_args, constraints=constraints
)
aten_dialect: ExportedProgram = export(f, example_args, constraints=constraints)
print("ATen Dialect Graph")
print(aten_dialect)
.. GENERATED FROM PYTHON SOURCE LINES 180-191
Note that that the inputs ``arg0_1`` and ``arg1_1`` now have shapes (3, s0),
with ``s0`` being a symbol representing that this dimension can be a range
of values.
Additionally, we can see in the **Range constraints** that value of ``s0`` has
the range [1, 10], which was specified by our constraints. We also see in the
**Equality constraints**, the tuple ``(InputDim(input_name='arg1_1', dim=1),
InputDim(input_name='arg0_1', dim=1))```, meaning that input 0's dimension 1
is equal to input 1's dimension 1, which was also specified by our constraints.
Now let's try running the model with different shapes:
.. GENERATED FROM PYTHON SOURCE LINES 191-209
.. code-block:: default
# Works correctly
print(aten_dialect(torch.ones(3, 3), torch.ones(3, 3)))
print(aten_dialect(torch.ones(3, 2), torch.ones(3, 2)))
# Errors because it violates our constraint that input 0, dim 1 <= 10
try:
print(aten_dialect(torch.ones(3, 15), torch.ones(3, 15)))
except Exception:
tb.print_exc()
# Errors because it violates our constraint that input 0, dim 1 == input 1, dim 1
try:
print(aten_dialect(torch.ones(3, 3), torch.ones(3, 2)))
except Exception:
tb.print_exc()
.. GENERATED FROM PYTHON SOURCE LINES 210-220
Addressing Untraceable Code
^^^^^^^^^^^^^^^^^^^^^^^^^^^
As our goal is to capture the entire computational graph from a PyTorch
program, we might ultimately run into untraceable parts of programs. To
address these issues, the
`torch.export documentation <https://pytorch.org/docs/2.1/export.html#limitations-of-torch-export>`__,
or the
`torch.export tutorial <https://pytorch.org/tutorials/intermediate/torch_export_tutorial.html>`__
would be the best place to look.
.. GENERATED FROM PYTHON SOURCE LINES 222-237
Performing Quantization
-----------------------
To quantize a model, we can do so between the call to
``torch._export.capture_pre_autograd_graph`` and ``torch.export``, in the
``Pre-Autograd ATen Dialect``. This is because quantization must operate at a
level which is safe for eager mode training.
Compared to
`FX Graph Mode Quantization <https://pytorch.org/tutorials/prototype/fx_graph_mode_ptq_static.html>`__,
we will need to call two new APIs: ``prepare_pt2e`` and ``compare_pt2e``
instead of ``prepare_fx`` and ``convert_fx``. It differs in that
``prepare_pt2e`` takes a backend-specific ``Quantizer`` as an arugument, which
will annotate the nodes in the graph with information needed to quantize the
model properly for a specific backend.
.. GENERATED FROM PYTHON SOURCE LINES 237-258
.. code-block:: default
example_args = (torch.randn(1, 3, 256, 256),)
pre_autograd_aten_dialect = capture_pre_autograd_graph(M(), example_args)
print("Pre-Autograd ATen Dialect Graph")
print(pre_autograd_aten_dialect)
from torch.ao.quantization.quantize_pt2e import convert_pt2e, prepare_pt2e
from torch.ao.quantization.quantizer.xnnpack_quantizer import (
get_symmetric_quantization_config,
XNNPACKQuantizer,
)
quantizer = XNNPACKQuantizer().set_global(get_symmetric_quantization_config())
prepared_graph = prepare_pt2e(pre_autograd_aten_dialect, quantizer)
converted_graph = convert_pt2e(prepared_graph)
print("Quantized Graph")
aten_dialect: ExportedProgram = export(converted_graph, example_args)
print("ATen Dialect Graph")
print(aten_dialect)
.. GENERATED FROM PYTHON SOURCE LINES 259-262
More information on how to quantize a model, and how a backend can implement a
``Quantizer`` can be found
`here <https://pytorch.org/tutorials/prototype/pt2e_quant_ptq_static.html>`__.
.. GENERATED FROM PYTHON SOURCE LINES 264-279
Lowering to Edge Dialect
------------------------
After exporting and lowering the graph to the ``ATen Dialect``, the next step
is to lower to the ``Edge Dialect``, in which specializations that are useful
for edge devices but not necessary for general (server) will be applied.
Some of these specializations include:
- DType specialization
- Scalar to tensor conversion
- Converting all ops to the ``executorch.exir.dialects.edge`` namespace.
Note that this dialect is still backend (or target) agnostic.
The lowering is done through the ``to_edge`` API.
.. GENERATED FROM PYTHON SOURCE LINES 279-295
.. code-block:: default
from executorch.exir import EdgeProgramManager, to_edge
example_args = (torch.randn(1, 3, 256, 256),)
pre_autograd_aten_dialect = capture_pre_autograd_graph(M(), example_args)
print("Pre-Autograd ATen Dialect Graph")
print(pre_autograd_aten_dialect)
aten_dialect: ExportedProgram = export(pre_autograd_aten_dialect, example_args)
print("ATen Dialect Graph")
print(aten_dialect)
edge_program: EdgeProgramManager = to_edge(aten_dialect)
print("Edge Dialect Graph")
print(edge_program.exported_program())
.. GENERATED FROM PYTHON SOURCE LINES 296-300
``to_edge()`` returns an ``EdgeProgramManager`` object, which contains the
exported programs which will be placed on this device. This data structure
allows users to export multiple programs and combine them into one binary. If
there is only one program, it will by default be saved to the name "forward".
.. GENERATED FROM PYTHON SOURCE LINES 300-329
.. code-block:: default
def encode(x):
return torch.nn.functional.linear(x, torch.randn(5, 10))
def decode(x):
return torch.nn.functional.linear(x, torch.randn(10, 5))
encode_args = (torch.randn(1, 10),)
aten_encode: ExportedProgram = export(
capture_pre_autograd_graph(encode, encode_args),
encode_args,
)
decode_args = (torch.randn(1, 5),)
aten_decode: ExportedProgram = export(
capture_pre_autograd_graph(decode, decode_args),
decode_args,
)
edge_program: EdgeProgramManager = to_edge(
{"encode": aten_encode, "decode": aten_decode}
)
for method in edge_program.methods:
print(f"Edge Dialect graph of {method}")
print(edge_program.exported_program(method))
.. GENERATED FROM PYTHON SOURCE LINES 330-339
We can also run additional passes on the exported program through
the ``transform`` API. An in-depth documentation on how to write
transformations can be found
`here <../compiler-custom-compiler-passes.html>`__.
Note that since the graph is now in the Edge Dialect, all passes must also
result in a valid Edge Dialect graph (specifically one thing to point out is
that the operators are now in the ``executorch.exir.dialects.edge`` namespace,
rather than the ``torch.ops.aten`` namespace.
.. GENERATED FROM PYTHON SOURCE LINES 339-365
.. code-block:: default
example_args = (torch.randn(1, 3, 256, 256),)
pre_autograd_aten_dialect = capture_pre_autograd_graph(M(), example_args)
aten_dialect: ExportedProgram = export(pre_autograd_aten_dialect, example_args)
edge_program: EdgeProgramManager = to_edge(aten_dialect)
print("Edge Dialect Graph")
print(edge_program.exported_program())
from executorch.exir.dialects._ops import ops as exir_ops
from executorch.exir.pass_base import ExportPass
class ConvertReluToSigmoid(ExportPass):
def call_operator(self, op, args, kwargs, meta):
if op == exir_ops.edge.aten.relu.default:
return super().call_operator(
exir_ops.edge.aten.sigmoid.default, args, kwargs, meta
)
else:
return super().call_operator(op, args, kwargs, meta)
transformed_edge_program = edge_program.transform((ConvertReluToSigmoid(),))
print("Transformed Edge Dialect Graph")
print(transformed_edge_program.exported_program())
.. GENERATED FROM PYTHON SOURCE LINES 366-380
Delegating to a Backend
-----------------------
We can now delegate parts of the graph or the whole graph to a third-party
backend through the ``to_backend`` API. An in-depth documentation on the
specifics of backend delegation, including how to delegate to a backend and
how to implement a backend, can be found
`here <../compiler-delegate-and-partitioner.html>`__
There are three ways for using this API:
1. Lowering the whole module
2. Lowering the whole module and composing it in another module
3. Paritioning the module, and lowering parts of the module
.. GENERATED FROM PYTHON SOURCE LINES 382-388
Lowering the Whole Module
^^^^^^^^^^^^^^^^^^^^^^^^^
To lower an entire module, we can pass ``to_backend`` the backend name, the
module to be lowered, and a list of compile specs to help the backend with the
lowering process.
.. GENERATED FROM PYTHON SOURCE LINES 388-426
.. code-block:: default
class LowerableModule(torch.nn.Module):
def __init__(self):
super().__init__()
def forward(self, x):
return torch.sin(x)
# Export and lower the module to Edge Dialect
example_args = (torch.ones(1),)
pre_autograd_aten_dialect = capture_pre_autograd_graph(LowerableModule(), example_args)
aten_dialect: ExportedProgram = export(pre_autograd_aten_dialect, example_args)
edge_program: EdgeProgramManager = to_edge(aten_dialect)
to_be_lowered_module = edge_program.exported_program()
from executorch.exir.backend.backend_api import LoweredBackendModule, to_backend
# Import the backend
from executorch.exir.backend.test.backend_with_compiler_demo import ( # noqa
BackendWithCompilerDemo,
)
# Lower the module
lowered_module: LoweredBackendModule = to_backend(
"BackendWithCompilerDemo", to_be_lowered_module, []
)
print(lowered_module)
print(lowered_module.backend_id)
print(lowered_module.processed_bytes)
print(lowered_module.original_module)
# Serialize and save it to a file
save_path = "delegate.pte"
with open(save_path, "wb") as f:
f.write(lowered_module.buffer())
.. GENERATED FROM PYTHON SOURCE LINES 427-435
In this call, ``to_backend`` will return a ``LoweredBackendModule``. Some
important attributes of the ``LoweredBackendModule`` are:
- ``backend_id``: The name of the backend this lowered module will run on in
the runtime
- ``processed_bytes``: a binary blob which will tell the backend how to run
this program in the runtime
- ``original_module``: the original exported module
.. GENERATED FROM PYTHON SOURCE LINES 437-442
Compose the Lowered Module into Another Module
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
In cases where we want to reuse this lowered module in multiple programs, we
can compose this lowered module with another module.
.. GENERATED FROM PYTHON SOURCE LINES 442-478
.. code-block:: default
class NotLowerableModule(torch.nn.Module):
def __init__(self, bias):
super().__init__()
self.bias = bias
def forward(self, a, b):
return torch.add(torch.add(a, b), self.bias)
class ComposedModule(torch.nn.Module):
def __init__(self):
super().__init__()
self.non_lowerable = NotLowerableModule(torch.ones(1) * 0.3)
self.lowerable = lowered_module
def forward(self, x):
a = self.lowerable(x)
b = self.lowerable(a)
ret = self.non_lowerable(a, b)
return a, b, ret
example_args = (torch.ones(1),)
pre_autograd_aten_dialect = capture_pre_autograd_graph(ComposedModule(), example_args)
aten_dialect: ExportedProgram = export(pre_autograd_aten_dialect, example_args)
edge_program: EdgeProgramManager = to_edge(aten_dialect)
exported_program = edge_program.exported_program()
print("Edge Dialect graph")
print(exported_program)
print("Lowered Module within the graph")
print(exported_program.graph_module.lowered_module_0.backend_id)
print(exported_program.graph_module.lowered_module_0.processed_bytes)
print(exported_program.graph_module.lowered_module_0.original_module)
.. GENERATED FROM PYTHON SOURCE LINES 479-483
Notice that there is now a ``torch.ops.executorch_call_delegate`` node in the
graph, which is calling ``lowered_module_0``. Additionally, the contents of
``lowered_module_0`` are the same as the ``lowered_module`` we created
previously.
.. GENERATED FROM PYTHON SOURCE LINES 485-493
Partition and Lower Parts of a Module
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
A separate lowering flow is to pass ``to_backend`` the module that we want to
lower, and a backend-specific partitioner. ``to_backend`` will use the
backend-specific partitioner to tag nodes in the module which are lowerable,
partition those nodes into subgraphs, and then create a
``LoweredBackendModule`` for each of those subgraphs.
.. GENERATED FROM PYTHON SOURCE LINES 493-520
.. code-block:: default
def f(a, x, b):
y = torch.mm(a, x)
z = y + b
a = z - a
y = torch.mm(a, x)
z = y + b
return z
example_args = (torch.randn(2, 2), torch.randn(2, 2), torch.randn(2, 2))
pre_autograd_aten_dialect = capture_pre_autograd_graph(f, example_args)
aten_dialect: ExportedProgram = export(pre_autograd_aten_dialect, example_args)
edge_program: EdgeProgramManager = to_edge(aten_dialect)
exported_program = edge_program.exported_program()
print("Edge Dialect graph")
print(exported_program)
from executorch.exir.backend.test.op_partitioner_demo import AddMulPartitionerDemo
delegated_program = to_backend(exported_program, AddMulPartitionerDemo)
print("Delegated program")
print(delegated_program)
print(delegated_program.graph_module.lowered_module_0.original_module)
print(delegated_program.graph_module.lowered_module_1.original_module)
.. GENERATED FROM PYTHON SOURCE LINES 521-527
Notice that there are now 2 ``torch.ops.executorch_call_delegate`` nodes in the
graph, one containing the operations `add, mul` and the other containing the
operations `mul, add`.
Alternatively, a more cohesive API to lower parts of a module is to directly
call ``to_backend`` on it:
.. GENERATED FROM PYTHON SOURCE LINES 527-548
.. code-block:: default
def f(a, x, b):
y = torch.mm(a, x)
z = y + b
a = z - a
y = torch.mm(a, x)
z = y + b
return z
example_args = (torch.randn(2, 2), torch.randn(2, 2), torch.randn(2, 2))
pre_autograd_aten_dialect = capture_pre_autograd_graph(f, example_args)
aten_dialect: ExportedProgram = export(pre_autograd_aten_dialect, example_args)
edge_program: EdgeProgramManager = to_edge(aten_dialect)
exported_program = edge_program.exported_program()
delegated_program = edge_program.to_backend(AddMulPartitionerDemo)
print("Delegated program")
print(delegated_program.exported_program())
.. GENERATED FROM PYTHON SOURCE LINES 549-561
Running User-Defined Passes and Memory Planning
-----------------------------------------------
As a final step of lowering, we can use the ``to_executorch()`` API to pass in
backend-specific passes, such as replacing sets of operators with a custom
backend operator, and a memory planning pass, to tell the runtime how to
allocate memory ahead of time when running the program.
A default memory planning pass is provided, but we can also choose a
backend-specific memory planning pass if it exists. More information on
writing a custom memory planning pass can be found
`here <../compiler-memory-planning.html>`__
.. GENERATED FROM PYTHON SOURCE LINES 561-579
.. code-block:: default
from executorch.exir import ExecutorchBackendConfig, ExecutorchProgramManager
from executorch.exir.passes import MemoryPlanningPass
executorch_program: ExecutorchProgramManager = edge_program.to_executorch(
ExecutorchBackendConfig(
passes=[], # User-defined passes
memory_planning_pass=MemoryPlanningPass(
"greedy"
), # Default memory planning pass
)
)
print("ExecuTorch Dialect")
print(executorch_program.exported_program())
import executorch.exir as exir
.. GENERATED FROM PYTHON SOURCE LINES 580-596
Notice that in the graph we now see operators like ``torch.ops.aten.sub.out``
and ``torch.ops.aten.div.out`` rather than ``torch.ops.aten.sub.Tensor`` and
``torch.ops.aten.div.Tensor``.
This is because between running the backend passes and memory planning passes,
to prepare the graph for memory planning, an out-variant pass is run on
the graph to convert all of the operators to their out variants. Instead of
allocating returned tensors in the kernel implementations, an operator's
``out`` variant will take in a prealloacated tensor to its out kwarg, and
store the result there, making it easier for memory planners to do tensor
lifetime analysis.
We also insert ``alloc`` nodes into the graph containing calls to a special
``executorch.exir.memory.alloc`` operator. This tells us how much memory is
needed to allocate each tensor output by the out-variant operator.
.. GENERATED FROM PYTHON SOURCE LINES 598-605
Saving to a File
----------------
Finally, we can save the ExecuTorch Program to a file and load it to a device
to be run.
Here is an example for an entire end-to-end workflow:
.. GENERATED FROM PYTHON SOURCE LINES 605-638
.. code-block:: default
import torch
from torch._export import capture_pre_autograd_graph
from torch.export import export, ExportedProgram
class M(torch.nn.Module):
def __init__(self):
super().__init__()
self.param = torch.nn.Parameter(torch.rand(3, 4))
self.linear = torch.nn.Linear(4, 5)
def forward(self, x):
return self.linear(x + self.param).clamp(min=0.0, max=1.0)
example_args = (torch.randn(3, 4),)
pre_autograd_aten_dialect = capture_pre_autograd_graph(M(), example_args)
# Optionally do quantization:
# pre_autograd_aten_dialect = convert_pt2e(prepare_pt2e(pre_autograd_aten_dialect, CustomBackendQuantizer))
aten_dialect: ExportedProgram = export(pre_autograd_aten_dialect, example_args)
edge_program: exir.EdgeProgramManager = exir.to_edge(aten_dialect)
# Optionally do delegation:
# edge_program = edge_program.to_backend(CustomBackendPartitioner)
executorch_program: exir.ExecutorchProgramManager = edge_program.to_executorch(
ExecutorchBackendConfig(
passes=[], # User-defined passes
)
)
with open("model.pte", "wb") as file:
file.write(executorch_program.buffer)
.. GENERATED FROM PYTHON SOURCE LINES 639-654
Conclusion
----------
In this tutorial, we went over the APIs and steps required to lower a PyTorch
program to a file that can be run on the ExecuTorch runtime.
Links Mentioned
^^^^^^^^^^^^^^^
- `torch.export Documentation <https://pytorch.org/docs/2.1/export.html>`__
- `Quantization Tutorial <https://pytorch.org/tutorials/prototype/pt2e_quant_ptq_static.html>`__
- `IR Spec <../ir-exir.html>`__
- `Writing Compiler Passes + Partitioner Documentation <../compiler-custom-compiler-passes.html>`__
- `Backend Delegation Documentation <../compiler-delegate-and-partitioner.html>`__
- `Memory Planning Documentation <../compiler-memory-planning.html>`__
.. rst-class:: sphx-glr-timing
**Total running time of the script:** (0 minutes 0.000 seconds)
.. _sphx_glr_download_tutorials_export-to-executorch-tutorial.py:
.. only:: html
.. container:: sphx-glr-footer sphx-glr-footer-example
.. container:: sphx-glr-download sphx-glr-download-python
:download:`Download Python source code: export-to-executorch-tutorial.py <export-to-executorch-tutorial.py>`
.. container:: sphx-glr-download sphx-glr-download-jupyter
:download:`Download Jupyter notebook: export-to-executorch-tutorial.ipynb <export-to-executorch-tutorial.ipynb>`
.. only:: html
.. rst-class:: sphx-glr-signature
`Gallery generated by Sphinx-Gallery <https://sphinx-gallery.github.io>`_