Python APIs

Overview

Intel® Extension for TensorFlow* provides flexible Python APIs to configure settings for different types of application scenarios.

Prerequisite: import intel_extension_for_tensorflow as itex

Python APIs and Environment Variable Names

You can easily configure and tune Intel® Extension for TensorFlow* run models using Python APIs and environment variables. We recommend Python APIs.

Python APIs and preserved environment variable Names

Python APIs Default value Environment Variables Default value Definition
itex.ConfigProto OFF
ON
ON
OFF
OFF
ITEX_ONEDNN_GRAPH
ITEX_LAYOUT_OPT
ITEX_REMAPPER
ITEX_AUTO_MIXED_PRECISION
ITEX_SHARDING
0
1*
1
0
0
Set configuration options for specific backend type (CPU/GPU) and graph optimization.
*ITEX_LAYOUT_OPT default ON in Intel GPU (except Intel® Data Center GPU Max Series) and default OFF in Intel CPU by hardware attributes
itex.experimental_ops_override N/A N/A OFF Call this function to automatically override the operators with same name in TensorFlow by itex.ops.

Notes:

  1. The priority for setting values is as follows: Python APIs > Environment Variables > Default value.

  2. If pip install intel-extension-for-tensorflow[xpu], both GPU and CPU backends will be installed, the default backend will be selected by the platform device situation. If the platform with Intel GPU, the activate backend will be GPU, otherwise, CPU. If GPU backend was installed by pip install intel-extension-for-tensorflow[gpu], the backend will be GPU. If CPU backend was installed by pip install intel-extension-for-tensorflow[cpu], the backend is CPU.

Intel® Extension for TensorFlow* Config Protocol

itex.ConfigProto: ProtocolMessage for XPU configuration under different types of backends and optimization options.

enum class

enum class Descriptions
enum ITEXDataType {
DEFAULT_DATA_TYPE = 0;
FLOAT16 = 1;
BFLOAT16 = 2;
}
Datatype options of advanced auto mixed precision. You could set datatype for advanced auto mixed precision on CPUs or GPUs.
enum Toggle {
DEFAULT = 0;
ON = 1;
OFF = 2;
}
Configuration options for the graph optimizer. Unless otherwise noted, these configuration options do not apply to explicitly triggered optimization passes in the optimizers field.

Functions

itex.ConfigProto

Attribute Description
graph_options GraphOptions protocolMessage, graph optimization options.

itex.GPUOptions

Attribute Description
None N/A

itex.GraphOptions

Attribute Description
onednn_graph Toggle onednn_graph

Override the environment variable ITEX_ONEDNN_GRAPH. Set to enable or disable oneDNN graph(LLGA) optimization. The default value is OFF.

* If ON, will enable oneDNN graph in Intel® Extension for TensorFlow.
* If OFF, will disable oneDNN graph in Intel® Extension for TensorFlow
.
layout_opt Toggle layout_opt

Override the environment variable ITEX_LAYOUT_OPT. Set if oneDNN layout optimization is enabled to benefit from oneDNN block format.
Enable or disable the oneDNN layout. The default value is OFF.

* If ON, will enable oneDNN layout optimization.
* If OFF, will disable oneDNN layout optimization.
remapper Toggle remapper

Override the environment variable ITEX_REMAPPER. Set if remapper optimization is enabled to benefit from sub-graph fusion.
Enable or disable the remapper. The default value is ON.

* If ON, will enable remapper optimization.
* If OFF, will disable remapper optimization.
auto_mixed_precision Toggle auto_mixed_precision

Override the environment variable ITEX_AUTO_MIXED_PRECISION. Set if mixed precision is enabled to benefit from using both 16-bit and 32-bit floating-point types to accelerate modes.
Enable or disable the auto mixed precision. The default value is OFF.

* If ON, will enable auto mixed precision optimization.
* If OFF, will disable auto mixed precision optimization.
sharding Toggle sharding

Currently only supports Intel GPUs with multi-tiles. Override the environment variable ITEX_SHARDING. Set if XPUAutoShard is enabled to benefit from sharding input data/graph to maximize hardware usage.
Enable or disable the XPUAutoShard. The default value is OFF.

* If ON, will enable XPUAutoShard optimization.
* If OFF, will disable XPUAutoShard optimization.

Examples:

I. Setting the options while creating the config protocol object

# TensorFlow and Intel® Extension for TensorFlow*
import tensorflow as tf
import intel_extension_for_tensorflow as itex

graph_opts=itex.GraphOptions(onednn_graph=itex.ON)
config=itex.ConfigProto(graph_options=graph_opts)
print(config)

Then the log will output the information like below.

graph_options {
  onednn_graph: ON
}

II. Setting the options after creating the config protocol object

# TensorFlow and Intel® Extension for TensorFlow*
import tensorflow as tf
import intel_extension_for_tensorflow as itex

config=itex.ConfigProto()

config.graph_options.onednn_graph=itex.ON
config.graph_options.layout_opt=itex.OFF

print(config)

Then the log will output the information like below.

graph_options {
  onednn_graph: ON
  layout_opt: OFF
}

itex.AutoMixedPrecisionOptions

ProtocolMessage for auto mixed precision optimization options.

Refer to Advanced Auto Mixed Precision.

itex.ShardingConfig

ProtocolMessage for XPUAutoShard optimization options. Currently only supports Intel GPUs with multi-tiles.

Refer to XPUAutoShard on GPU.

itex.DebugOptions

ProtocolMessage for debug options.

Python APIs Environment Variables Definition
auto_mixed_precision_log_path ITEX_AUTO_MIXED_PRECISION_LOG_PATH Save auto mixed precision "pre-optimization" and "post-optimization" graph to log path.
xpu_force_sync ITEX_SYNC_EXEC Run the graph with sync mode. The default value is OFF. If ON, the whole model will be run with sync mode, which will hurt performance.

itex.set_config

Set Config Protocol. Note that the protocol is a global value, so this API is not thread safe.

itex.set_config(config)

| Args | Description | | ———————–| ————————————————————————| | config | ConfigProto object|

Raises Description
ValueError If argument validation fails.

itex.get_config

Get Config Protocol.

itex.get_config()
Raises Description
Returns Return the current config.

Example:

import intel_extension_for_tensorflow as itex

graph_opts=itex.GraphOptions(onednn_graph=itex.ON)
config=itex.ConfigProto(graph_options=graph_opts)
itex.set_config(config)
print(itex.get_config())

Then the log will output the information like below:

graph_options {
  onednn_graph: ON
}

itex operators

itex.ops: Public API for extended XPU ops(operations) for itex.ops namespace.

For details, refer to Customized Operators.

itex ops override

itex.experimental_ops_override: Public API to override TensorFlow specific operators with same name by Customized Operators in itex.ops namespace.

For details, refer to Intel® Extension for TensorFlow* ops override.

Example:

import intel_extension_for_tensorflow as itex
import tensorflow as tf
itex.experimental_ops_override()
print(tf.nn.gelu == itex.ops.gelu)

Then it will output the result “True”.

itex graph

itex.graph: Public API for extended ITEX graph optimization operations.

N/A

itex version

itex.version: Public API for itex.version namespace.

Other Members Description
VERSION The release version. For example, 0.3.0
GIT_VERSION The git version. For example, v0.3.0-7112d33
ONEDNN_CPU_GIT_VERSION The oneDNN git version of CPU. For example, v2.5.2-a930253
ONEDNN_GPU_GIT_VERSION The oneDNN git version of GPU. For example, v2.5.2-a930253
COMPILER_VERSION The compiler version. For example, gcc-8.2.1 20180905, dpcpp-2022.1.0.122
TF_COMPATIBLE_VERSION The compatible TensorFlow versions. For example, tensorflow >= 2.5.0, < 2.7.0, !=2.5.3, ~=2.6

Example:

import tensorflow as tf
import intel_extension_for_tensorflow as itex

print(itex.__version__)
print(itex.version.VERSION)
print(itex.version.GIT_VERSION)
if hasattr(itex.version, "ONEDNN_CPU_GIT_VERSION"):
  print(itex.version.ONEDNN_CPU_GIT_VERSION)    # For CPU or XPU
if hasattr(itex.version, "ONEDNN_GPU_GIT_VERSION"):
  print(itex.version.ONEDNN_GPU_GIT_VERSION)    # For GPU or XPU
print(itex.version.COMPILER_VERSION)
print(itex.version.TF_COMPATIBLE_VERSION)