Optimum documentation

Configuration

You are viewing main version, which requires installation from source. If you'd like regular pip install, checkout the latest stable version (v1.8.6).
Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

Configuration

The configuration classes are the way to specify how a task should be done. There are two tasks supported with the ONNX Runtime package:

  1. Optimization: Performed by the ORTOptimizer, this task can be tweaked using an OptimizationConfig.

  2. Quantization: Performed by the ORTQuantizer, quantization can be set using a QuantizationConfig. A calibration step is required in some cases (post training static quantization), which can be specified using a CalibrationConfig.

OptimizationConfig

class optimum.onnxruntime.OptimizationConfig

< >

( optimization_level: int = 1 optimize_for_gpu: bool = False fp16: bool = False optimize_with_onnxruntime_only: typing.Optional[bool] = None enable_transformers_specific_optimizations: bool = True disable_gelu: typing.Optional[bool] = None disable_gelu_fusion: bool = False disable_layer_norm: typing.Optional[bool] = None disable_layer_norm_fusion: bool = False disable_attention: typing.Optional[bool] = None disable_attention_fusion: bool = False disable_skip_layer_norm: typing.Optional[bool] = None disable_skip_layer_norm_fusion: bool = False disable_bias_skip_layer_norm: typing.Optional[bool] = None disable_bias_skip_layer_norm_fusion: bool = False disable_bias_gelu: typing.Optional[bool] = None disable_bias_gelu_fusion: bool = False disable_embed_layer_norm: bool = True disable_embed_layer_norm_fusion: bool = True enable_gelu_approximation: bool = False use_mask_index: bool = False no_attention_mask: bool = False disable_shape_inference: bool = False use_multi_head_attention: bool = False enable_gemm_fast_gelu_fusion: bool = False use_raw_attention_mask: bool = False disable_group_norm_fusion: bool = True disable_packed_kv: bool = True )

Parameters

  • optimization_level (int, defaults to 1) — Optimization level performed by ONNX Runtime of the loaded graph. Supported optimization level are 0, 1, 2 and 99.
    • 0: will disable all optimizations
    • 1: will enable basic optimizations
    • 2: will enable basic and extended optimizations, including complex node fusions applied to the nodes assigned to the CPU or CUDA execution provider, making the resulting optimized graph hardware dependent
    • 99: will enable all available optimizations including layout optimizations
  • optimize_for_gpu (bool, defaults to False) — Whether to optimize the model for GPU inference. The optimized graph might contain operators for GPU or CPU only when optimization_level > 1.
  • fp16 (bool, defaults to False) — Whether all weights and nodes should be converted from float32 to float16.
  • enable_transformers_specific_optimizations (bool, defaults to True) — Whether to only use transformers specific optimizations on top of ONNX Runtime general optimizations.
  • disable_gelu_fusion (bool, defaults to False) — Whether to disable the Gelu fusion.
  • disable_layer_norm_fusion (bool, defaults to False) — Whether to disable Layer Normalization fusion.
  • disable_attention_fusion (bool, defaults to False) — Whether to disable Attention fusion.
  • disable_skip_layer_norm_fusion (bool, defaults to False) — Whether to disable SkipLayerNormalization fusion.
  • disable_bias_skip_layer_norm_fusion (bool, defaults to False) — Whether to disable Add Bias and SkipLayerNormalization fusion.
  • disable_bias_gelu_fusion (bool, defaults to False) — Whether to disable Add Bias and Gelu / FastGelu fusion.
  • disable_embed_layer_norm_fusion (bool, defaults to True) — Whether to disable EmbedLayerNormalization fusion. The default value is set to True since this fusion is incompatible with ONNX Runtime quantization.
  • enable_gelu_approximation (bool, defaults to False) — Whether to enable Gelu / BiasGelu to FastGelu conversion. The default value is set to False since this approximation might slightly impact the model’s accuracy.
  • use_mask_index (bool, defaults to False) — Whether to use mask index instead of raw attention mask in the attention operator.
  • no_attention_mask (bool, defaults to False) — Whether to not use attention masks. Only works for bert model type.
  • disable_embed_layer_norm (bool, defaults to True) — Whether to disable EmbedLayerNormalization fusion. The default value is set to True since this fusion is incompatible with ONNX Runtime quantization
  • disable_shape_inference (bool, defaults to False) — Whether to disable symbolic shape inference. The default value is set to False but symbolic shape inference might cause issues sometimes.
  • use_multi_head_attention (bool, defaults to False) — Experimental argument. Use MultiHeadAttention instead of Attention operator, which has merged weights for Q/K/V projection, which might be faster in some cases since 3 MatMul is merged into one.” “Note that MultiHeadAttention might be slower than Attention when qkv are not packed. ”
  • enable_gemm_fast_gelu_fusion (bool, defaults to False) — Enable GemmfastGelu fusion.
  • use_raw_attention_mask (bool, defaults to False) — Use raw attention mask. Use this option if your input is not right-side padding. This might deactivate fused attention and get worse performance.
  • disable_group_norm_fusion (bool, defaults to True) — Do not fuse GroupNorm. Only works for model_type=unet.
  • disable_packed_kv (bool, defaults to True) — Do not use packed kv in cross attention. Only works for model_type=unet.

OptimizationConfig is the configuration class handling all the ONNX Runtime optimization parameters. There are two stacks of optimizations:

  1. The ONNX Runtime general-purpose optimization tool: it can work on any ONNX model.
  2. The ONNX Runtime transformers optimization tool: it can only work on a subset of transformers models.

class optimum.onnxruntime.AutoOptimizationConfig

< >

( )

Factory to create common OptimizationConfig.

O1

< >

( for_gpu: bool = False **kwargs ) → OptimizationConfig

Parameters

  • for_gpu (bool, defaults to False) — Whether the model to optimize will run on GPU, some optimizations depends on the hardware the model will run on. Only needed for optimization_level > 1.
  • kwargs (Dict[str, Any]) — Arguments to provide to the ~OptimizationConfig constructor.

Returns

OptimizationConfig

The OptimizationConfig corresponding to the O1 optimization level.

Creates an O1 ~OptimizationConfig.

O2

< >

( for_gpu: bool = False **kwargs ) → OptimizationConfig

Parameters

  • for_gpu (bool, defaults to False) — Whether the model to optimize will run on GPU, some optimizations depends on the hardware the model will run on. Only needed for optimization_level > 1.
  • kwargs (Dict[str, Any]) — Arguments to provide to the ~OptimizationConfig constructor.

Returns

OptimizationConfig

The OptimizationConfig corresponding to the O2 optimization level.

Creates an O2 ~OptimizationConfig.

O3

< >

( for_gpu: bool = False **kwargs ) → OptimizationConfig

Parameters

  • for_gpu (bool, defaults to False) — Whether the model to optimize will run on GPU, some optimizations depends on the hardware the model will run on. Only needed for optimization_level > 1.
  • kwargs (Dict[str, Any]) — Arguments to provide to the ~OptimizationConfig constructor.

Returns

OptimizationConfig

The OptimizationConfig corresponding to the O3 optimization level.

Creates an O3 ~OptimizationConfig.

O4

< >

( for_gpu: bool = True **kwargs ) → OptimizationConfig

Parameters

  • for_gpu (bool, defaults to False) — Whether the model to optimize will run on GPU, some optimizations depends on the hardware the model will run on. Only needed for optimization_level > 1.
  • kwargs (Dict[str, Any]) — Arguments to provide to the ~OptimizationConfig constructor.

Returns

OptimizationConfig

The OptimizationConfig corresponding to the O4 optimization level.

Creates an O4 ~OptimizationConfig.

with_optimization_level

< >

( optimization_level: str for_gpu: bool = False **kwargs ) → OptimizationConfig

Parameters

  • optimization_level (str) — The optimization level, the following values are allowed:
    • O1: Basic general optimizations
    • O2: Basic and extended general optimizations, transformers-specific fusions.
    • O3: Same as O2 with Fast Gelu approximation.
    • O4: Same as O3 with mixed precision.
  • for_gpu (bool, defaults to False) — Whether the model to optimize will run on GPU, some optimizations depends on the hardware the model will run on. Only needed for optimization_level > 1.
  • kwargs (Dict[str, Any]) — Arguments to provide to the ~OptimizationConfig constructor.

Returns

OptimizationConfig

The OptimizationConfig corresponding to the requested optimization level.

Creates an ~OptimizationConfig with pre-defined arguments according to an optimization level.

QuantizationConfig

class optimum.onnxruntime.QuantizationConfig

< >

( is_static: bool format: QuantFormat mode: QuantizationMode = <QuantizationMode.QLinearOps: 1> activations_dtype: QuantType = <QuantType.QUInt8: 1> activations_symmetric: bool = False weights_dtype: QuantType = <QuantType.QInt8: 0> weights_symmetric: bool = True per_channel: bool = False reduce_range: bool = False nodes_to_quantize: typing.List[str] = <factory> nodes_to_exclude: typing.List[str] = <factory> operators_to_quantize: typing.List[str] = <factory> qdq_add_pair_to_weight: bool = False qdq_dedicated_pair: bool = False qdq_op_type_per_channel_support_to_axis: typing.Dict[str, int] = <factory> )

Parameters

  • is_static (bool) — Whether to apply static quantization or dynamic quantization.
  • format (QuantFormat) — Targeted ONNX Runtime quantization representation format. For the Operator Oriented (QOperator) format, all the quantized operators have their own ONNX definitions. For the Tensor Oriented (QDQ) format, the model is quantized by inserting QuantizeLinear / DeQuantizeLinear operators.
  • mode (QuantizationMode, defaults to QuantizationMode.QLinearOps) — Targeted ONNX Runtime quantization mode, default is QLinearOps to match QDQ format. When targeting dynamic quantization mode, the default value is QuantizationMode.IntegerOps whereas the default value for static quantization mode is QuantizationMode.QLinearOps.
  • activations_dtype (QuantType, defaults to QuantType.QUInt8) — The quantization data types to use for the activations.
  • activations_symmetric (bool, defaults to False) — Whether to apply symmetric quantization on the activations.
  • weights_dtype (QuantType, defaults to QuantType.QInt8) — The quantization data types to use for the weights.
  • weights_symmetric (bool, defaults to True) — Whether to apply symmetric quantization on the weights.
  • per_channel (bool, defaults to False) — Whether we should quantize per-channel (also known as “per-row”). Enabling this can increase overall accuracy while making the quantized model heavier.
  • reduce_range (bool, defaults to False) — Whether to use reduce-range 7-bits integers instead of 8-bits integers.
  • nodes_to_quantize (list) — List of the nodes names to quantize.
  • nodes_to_exclude (list) — List of the nodes names to exclude when applying quantization.
  • operators_to_quantize (list, defaults to ["MatMul", "Add"]) — List of the operators types to quantize.
  • qdq_add_pair_to_weight (bool, defaults to False) — By default, floating-point weights are quantized and feed to solely inserted DeQuantizeLinear node. If set to True, the floating-point weights will remain and both QuantizeLinear / DeQuantizeLinear nodes will be inserted.
  • qdq_dedicated_pair (bool, defaults to False) — When inserting QDQ pair, multiple nodes can share a single QDQ pair as their inputs. If True, it will create an identical and dedicated QDQ pair for each node.
  • qdq_op_type_per_channel_support_to_axis (Dict[str, int]) — Set the channel axis for a specific operator type. Effective only when per channel quantization is supported and per_channel is set to True.

QuantizationConfig is the configuration class handling all the ONNX Runtime quantization parameters.

AutoQuantizationConfig

class optimum.onnxruntime.AutoQuantizationConfig

< >

( )

arm64

< >

( is_static: bool use_symmetric_activations: bool = False use_symmetric_weights: bool = True per_channel: bool = True nodes_to_quantize: typing.Optional[typing.List[str]] = None nodes_to_exclude: typing.Optional[typing.List[str]] = None operators_to_quantize: typing.List[str] = ['MatMul', 'Add'] )

Parameters

  • is_static (bool) — Boolean flag to indicate whether we target static or dynamic quantization.
  • use_symmetric_activations (bool, defaults to False) — Whether to use symmetric quantization for activations.
  • use_symmetric_weights (bool, defaults to True) — Whether to use symmetric quantization for weights.
  • per_channel (bool, defaults to True) — Whether we should quantize per-channel (also known as “per-row”). Enabling this can increase overall accuracy while making the quantized model heavier.
  • nodes_to_quantize (Optional[List[NodeName]], defaults to None) — Specific nodes to quantize. If None, all nodes being operators from operators_to_quantize will be quantized.
  • nodes_to_exclude (Optional[List[NodeName]], defaults to None) — Specific nodes to exclude from quantization.
  • operators_to_quantize (List[NodeName], defaults to ["MatMul", "Add"]) — Type of nodes to perform quantization on.

Creates a QuantizationConfig fit for ARM64.

avx2

< >

( is_static: bool use_symmetric_activations: bool = False use_symmetric_weights: bool = True per_channel: bool = True reduce_range: bool = False nodes_to_quantize: typing.Optional[typing.List[str]] = None nodes_to_exclude: typing.Optional[typing.List[str]] = None operators_to_quantize: typing.List[str] = ['MatMul', 'Add'] )

Parameters

  • is_static (bool) — Boolean flag to indicate whether we target static or dynamic quantization.
  • use_symmetric_activations (bool, defaults to False) — Whether to use symmetric quantization for activations.
  • use_symmetric_weights (bool, defaults to True) — Whether to use symmetric quantization for weights.
  • per_channel (bool, defaults to True) — Whether we should quantize per-channel (also known as “per-row”). Enabling this can increase overall accuracy while making the quantized model heavier.
  • reduce_range (bool, defaults to False) — Indicate whether to use 8-bits integers (False) or reduce-range 7-bits integers (True). As a baseline, it is always recommended testing with full range (reduce_range = False) and then, if accuracy drop is significant, to try with reduced range (reduce_range = True). Intel’s CPUs using AVX512 (non VNNI) can suffer from saturation issue when invoking the VPMADDUBSW instruction. To counter this, one should use 7-bits rather than 8-bits integers.
  • nodes_to_quantize (Optional[List[NodeName]], defaults to None) — Specific nodes to quantize. If None, all nodes being operators from operators_to_quantize will be quantized.
  • nodes_to_exclude (Optional[List[NodeName]], defaults to None) — Specific nodes to exclude from quantization.
  • operators_to_quantize (List[NodeName], defaults to ["MatMul", "Add"]) — Type of nodes to perform quantization on.

Creates a QuantizationConfig fit for CPU with AVX2 instruction set.

avx512

< >

( is_static: bool use_symmetric_activations: bool = False use_symmetric_weights: bool = True per_channel: bool = True reduce_range: bool = False nodes_to_quantize: typing.Optional[typing.List[str]] = None nodes_to_exclude: typing.Optional[typing.List[str]] = None operators_to_quantize: typing.List[str] = ['MatMul', 'Add'] )

Parameters

  • is_static (bool) — Boolean flag to indicate whether we target static or dynamic quantization.
  • use_symmetric_activations (bool, defaults to False) — Whether to use symmetric quantization for activations.
  • use_symmetric_weights (bool, defaults to True) — Whether to use symmetric quantization for weights.
  • per_channel (bool, defaults to True) — Whether we should quantize per-channel (also known as “per-row”). Enabling this can increase overall accuracy while making the quantized model heavier.
  • reduce_range (bool, defaults to False) — Indicate whether to use 8-bits integers (False) or reduce-range 7-bits integers (True). As a baseline, it is always recommended testing with full range (reduce_range = False) and then, if accuracy drop is significant, to try with reduced range (reduce_range = True). Intel’s CPUs using AVX512 (non VNNI) can suffer from saturation issue when invoking the VPMADDUBSW instruction. To counter this, one should use 7-bits rather than 8-bits integers.
  • nodes_to_quantize (Optional[List[NodeName]], defaults to None) — Specific nodes to quantize. If None, all nodes being operators from operators_to_quantize will be quantized.
  • nodes_to_exclude (Optional[List[NodeName]], defaults to None) — Specific nodes to exclude from quantization.
  • operators_to_quantize (List[NodeName], defaults to ["MatMul", "Add"]) — Type of nodes to perform quantization on.

Creates a QuantizationConfig fit for CPU with AVX512 instruction set.

avx512_vnni

< >

( is_static: bool use_symmetric_activations: bool = False use_symmetric_weights: bool = True per_channel: bool = True nodes_to_quantize: typing.Optional[typing.List[str]] = None nodes_to_exclude: typing.Optional[typing.List[str]] = None operators_to_quantize: typing.List[str] = ['MatMul', 'Add'] )

Parameters

  • is_static (bool) — Boolean flag to indicate whether we target static or dynamic quantization.
  • use_symmetric_activations (bool, defaults to False) — Whether to use symmetric quantization for activations.
  • use_symmetric_weights (bool, defaults to True) — Whether to use symmetric quantization for weights.
  • per_channel (bool, defaults to True) — Whether we should quantize per-channel (also known as “per-row”). Enabling this can increase overall accuracy while making the quantized model heavier.
  • nodes_to_quantize (Optional[List[NodeName]], defaults to None) — Specific nodes to quantize. If None, all nodes being operators from operators_to_quantize will be quantized.
  • nodes_to_exclude (Optional[List[NodeName]], defaults to None) — Specific nodes to exclude from quantization.
  • operators_to_quantize (List[NodeName], defaults to ["MatMul", "Add"]) — Type of nodes to perform quantization on.

Creates a QuantizationConfig fit for CPU with AVX512-VNNI instruction set.

When targeting Intel AVX512-VNNI CPU underlying execution engine leverage the CPU instruction VPDPBUSD to compute \i32 += i8(w) * u8(x)\ within a single instruction.

AVX512-VNNI (AVX512 Vector Neural Network Instruction) is an x86 extension Instruction set and is a part of the AVX-512 ISA.

AVX512 VNNI is designed to accelerate convolutional neural network for INT8 inference.

tensorrt

< >

( per_channel: bool = True nodes_to_quantize: typing.Optional[typing.List[str]] = None nodes_to_exclude: typing.Optional[typing.List[str]] = None operators_to_quantize: typing.List[str] = ['MatMul', 'Add'] )

Parameters

  • per_channel (bool, defaults to True) — Whether we should quantize per-channel (also known as “per-row”). Enabling this can increase overall accuracy while making the quantized model heavier.
  • nodes_to_quantize (Optional[List[NodeName]], defaults to None) — Specific nodes to quantize. If None, all nodes being operators from operators_to_quantize will be quantized.
  • nodes_to_exclude (Optional[List[NodeName]], defaults to None) — Specific nodes to exclude from quantization.
  • operators_to_quantize (List[NodeName], defaults to ["MatMul", "Add"]) — Type of nodes to perform quantization on.

Creates a QuantizationConfig fit for TensorRT static quantization, targetting NVIDIA GPUs.

CalibrationConfig

class optimum.onnxruntime.CalibrationConfig

< >

( dataset_name: str dataset_config_name: str dataset_split: str dataset_num_samples: int method: CalibrationMethod num_bins: typing.Optional[int] = None num_quantized_bins: typing.Optional[int] = None percentile: typing.Optional[float] = None moving_average: typing.Optional[bool] = None averaging_constant: typing.Optional[float] = None )

Parameters

  • dataset_name (str) — The name of the calibration dataset.
  • dataset_config_name (str) — The name of the calibration dataset configuration.
  • dataset_split (str) — Which split of the dataset is used to perform the calibration step.
  • dataset_num_samples (int) — The number of samples composing the calibration dataset.
  • method (CalibrationMethod) — The method chosen to calculate the activations quantization parameters using the calibration dataset.
  • num_bins (Optional[int], defaults to None) — The number of bins to use when creating the histogram when performing the calibration step using the Percentile or Entropy method.
  • num_quantized_bins (Optional[int], defaults to None) — The number of quantized bins to use when performing the calibration step using the Entropy method.
  • percentile (Optional[float], defaults to None) — The percentile to use when computing the activations quantization ranges when performing the calibration step using the Percentile method.
  • moving_average (Optional[bool], defaults to None) — Whether to compute the moving average of the minimum and maximum values when performing the calibration step using the MinMax method.
  • averaging_constant (Optional[float], defaults to None) — The constant smoothing factor to use when computing the moving average of the minimum and maximum values. Effective only when the MinMax calibration method is selected and moving_average is set to True.

CalibrationConfig is the configuration class handling all the ONNX Runtime parameters related to the calibration step of static quantization.

ORTConfig

class optimum.onnxruntime.ORTConfig

< >

( opset: typing.Optional[int] = None use_external_data_format: bool = False one_external_file: bool = True optimization: typing.Optional[optimum.onnxruntime.configuration.OptimizationConfig] = None quantization: typing.Optional[optimum.onnxruntime.configuration.QuantizationConfig] = None **kwargs )

Parameters

  • opset (Optional[int], defaults to None) — ONNX opset version to export the model with.
  • use_external_data_format (bool, defaults to False) — Allow exporting model >= than 2Gb.
  • one_external_file (bool, defaults to True) — When use_external_data_format=True, whether to save all tensors to one external file. If false, save each tensor to a file named with the tensor name. (Can not be set to False for the quantization)
  • optimization (Optional[OptimizationConfig], defaults to None) — Specify a configuration to optimize ONNX Runtime model
  • quantization (Optional[QuantizationConfig], defaults to None) — Specify a configuration to quantize ONNX Runtime model

ORTConfig is the configuration class handling all the ONNX Runtime parameters related to the ONNX IR model export, optimization and quantization parameters.