PyTorch Deployment via “torch.compile”#

The torch.compile feature enables you to use OpenVINO for PyTorch-native applications. It speeds up PyTorch code by JIT-compiling it into optimized kernels. By default, Torch code runs in eager-mode, but with the use of torch.compile it goes through the following steps:

  1. Graph acquisition - the model is rewritten as blocks of subgraphs that are either:

    • compiled by TorchDynamo and “flattened”,

    • falling back to the eager-mode, due to unsupported Python constructs (like control-flow code).

  2. Graph lowering - all PyTorch operations are decomposed into their constituent kernels specific to the chosen backend.

  3. Graph compilation - the kernels call their corresponding low-level device-specific operations.

How to Use#

To use torch.compile, you need to define the openvino backend in your PyTorch application. This way Torch FX subgraphs will be directly converted to OpenVINO representation without any additional PyTorch-based tracing/scripting. This approach works only for the package distributed via pip, as it is now configured with torch_dynamo_backends entrypoint.

...
model = torch.compile(model, backend='openvino')
...

For OpenVINO installed via channels other than pip, such as conda, and versions older than 2024.1, an additional import statement is needed:

import openvino.torch

...
model = torch.compile(model, backend='openvino')
...

Execution diagram:

torch.compile execution diagram

Options#

It is possible to use additional arguments for torch.compile to set the backend device, enable model caching, set the cache directory etc. You can use a dictionary of the available options:

  • device - enables selecting a specific hardware device to run the application. By default, the OpenVINO backend for torch.compile runs PyTorch applications on CPU. If you set this variable to GPU.0, for example, the application will use the integrated graphics processor instead.

  • aot_autograd - enables a graph capture needed for dynamic shapes or to finetune a model. For models with dynamic shapes, it is recommended to set this option to True. By default, aot_autograd is set to False.

  • model_caching - enables saving the optimized model files to a hard drive, after the first application run. This makes them available for the following application executions, reducing the first-inference latency. By default, this variable is set to False. Set it to True to enable caching.

  • cache_dir - enables defining a custom directory for the model files (if model_caching is set to True). By default, the OpenVINO IR is saved in the cache sub-directory, created in the application’s root directory.

  • decompositions - enables defining additional operator decompositions. By default, this is an empty list. For example, to add a decomposition for an operator my_op, add 'decompositions': [torch.ops.aten.my_op.default] to the options.

  • disabled_ops - enables specifying operators that can be disabled from openvino execution and make it fall back to native PyTorch runtime. For example, to disable an operator my_op from OpenVINO execution, add 'disabled_ops': [torch.ops.aten.my_op.default] to the options. By default, this is an empty list.

  • config - enables passing any OpenVINO configuration option as a dictionary to this variable. For details on the various options, refer to the OpenVINO Advanced Features.

See the example below for details:

model = torch.compile(model, backend="openvino", options = {"device" : "CPU", "model_caching" : True, "cache_dir": "./model_cache"})

You can also set OpenVINO specific configuration options by adding them as a dictionary under config key in options:

opts = {"device" : "CPU", "config" : {"PERFORMANCE_HINT" : "LATENCY"}}
model = torch.compile(model, backend="openvino", options=opts)

Windows support#

PyTorch supports torch.compile officially on Windows from version 2.3.0 onwards.

For PyTorch versions below 2.3.0, the torch.compile feature is not supported on Windows officially. However, it can be accessed by running the following instructions:

  1. Install the PyTorch nightly wheel file - 2.1.0.dev20230713 ,

  2. Update the file at <python_env_root>/Lib/site-packages/torch/_dynamo/eval_frames.py

  3. Find the function called check_if_dynamo_supported():

    def check_if_dynamo_supported():
    if sys.platform == "win32":
        raise RuntimeError("Windows not yet supported for torch.compile")
    if sys.version_info >= (3, 11):
        raise RuntimeError("Python 3.11+ not yet supported for torch.compile")

  4. Put in comments the first two lines in this function, so it looks like this:

    def check_if_dynamo_supported():
 #if sys.platform == "win32":
 #    raise RuntimeError("Windows not yet supported for torch.compile")
 if sys.version_info >= (3, 11):
     `raise RuntimeError("Python 3.11+ not yet supported for torch.compile")

Support for PyTorch 2 export quantization (Preview)#

PyTorch 2 export quantization is supported by OpenVINO backend in torch.compile. To be able to access this feature, follow the steps provided in PyTorch 2 Export Post Training Quantization with X86 Backend through Inductor and update the provided sample as explained below.

  1. If you are using PyTorch version 2.3.0 or later, disable constant folding in quantization to be able to benefit from the optimization in the OpenVINO backend. This can be done by passing fold_quantize=False parameter into the convert_pt2e function. To do so, change this line:

    converted_model = convert_pt2e(prepared_model)

    to the following:

    converted_model = convert_pt2e(prepared_model, fold_quantize=False)

  2. Set torch.compile backend as OpenVINO and execute the model.

    Update this line below:

    optimized_model = torch.compile(converted_model)

    As below:

    optimized_model = torch.compile(converted_model, backend="openvino")

TorchServe Integration#

TorchServe is a performant, flexible, and easy to use tool for serving PyTorch models in production. For more information on the details of TorchServe, you can refer to TorchServe github repository.. With OpenVINO torch.compile integration into TorchServe you can serve PyTorch models in production and accelerate them with OpenVINO on various Intel hardware. Detailed instructions on how to use OpenVINO with TorchServe are available in TorchServe examples.

Support for Automatic1111 Stable Diffusion WebUI#

Automatic1111 Stable Diffusion WebUI is an open-source repository that hosts a browser-based interface for the Stable Diffusion based image generation. It allows users to create realistic and creative images from text prompts. Stable Diffusion WebUI is supported on Intel CPUs, Intel integrated GPUs, and Intel discrete GPUs by leveraging OpenVINO torch.compile capability. Detailed instructions are available in Stable Diffusion WebUI repository.

Architecture#

The torch.compile feature is part of PyTorch 2.0, and is based on:

  • TorchDynamo - a Python-level JIT that hooks into the frame evaluation API in CPython, (PEP 523) to dynamically modify Python bytecode right before it is executed (PyTorch operators that cannot be extracted to FX graph are executed in the native Python environment). It maintains the eager-mode capabilities using Guards to ensure the generated graphs are valid.

  • AOTAutograd - generates the backward graph corresponding to the forward graph captured by TorchDynamo.

  • PrimTorch - decomposes complicated PyTorch operations into simpler and more elementary ops.

  • TorchInductor - a deep learning compiler that generates fast code for multiple accelerators and backends.

When the PyTorch module is wrapped with torch.compile, TorchDynamo traces the module and rewrites Python bytecode to extract sequences of PyTorch operations into an FX Graph, which can be optimized by the OpenVINO backend. The Torch FX graphs are first converted to inlined FX graphs and the graph partitioning module traverses inlined FX graph to identify operators supported by OpenVINO.

All the supported operators are clustered into OpenVINO submodules, converted to the OpenVINO graph using OpenVINO’s PyTorch decoder, and executed in an optimized manner using OpenVINO runtime. All unsupported operators fall back to the native PyTorch runtime on CPU. If the subgraph fails during OpenVINO conversion, the subgraph falls back to PyTorch’s default inductor backend.

Additional Resources#