Quantization

Quantization refers to techniques for performing neural network computations in lower precision than floating point. Quantization can reduce a model’s size and also improve a model’s inference latency and memory bandwidth requirement, because many hardware platforms offer high-performance implementations of quantized operations.

class coremltools.optimize.torch.quantization.ModuleLinearQuantizerConfig(weight_dtype: str | dtype = torch.qint8, weight_observer=ObserverType.moving_average_min_max, weight_per_channel: bool = True, activation_dtype: str | dtype = torch.quint8, activation_observer=ObserverType.moving_average_min_max, quantization_scheme=QuantizationScheme.symmetric, milestones: List[int] | None = None)[source]

Configuration class for specifying global and module level quantization options for linear quantization algorithm implemented in LinearQuantizer.

Linear quantization algorithm simulates the effects of quantization during training, by quantizing and dequantizing the weights and/or activations during the model’s forward pass. The forward and backward pass computations are conducted in float dtype, however, these float values follow the constraints imposed by int8 and quint8 dtypes. For more details, please refer to Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference.

For most applications, the only parameters that need to be set are quantization_scheme and milestones.

By default, quantization_scheme is set to QuantizationScheme.symmetric, which means all weights are quantized with zero point as zero, and activations are quantized with zero point as zero for non-negative activations and 128 for all other activations. The weights are quantized using torch.qint8 and activations are quantized using torch.quint8.

Linear quantization algorithm inserts observers for each weight/activation tensor. These observers collect statistics of these tensors’ values, for example, the minimum and maximum values they can take. These statistics are then used to compute the scale and zero point, which are in turn used for quantizing the weights/activations. By default, moving_average_min_max observer is used. For more details, please check MinMaxObserver.

The milestones parameter controls the flow of the quantization algorithm. The example below illustrates its usage in more detail:

model = define_model()

config = LinearQuantizerConfig(
    global_config=ModuleLinearQuantizerConfig(
        quantization_scheme="symmetric",
        milestones=[0, 100, 300, 200],
    )
)

quantizer = LinearQuantizer(model, config)

# prepare the model to insert FakeQuantize layers for QAT
model = quantizer.prepare()

# use quantizer in your PyTorch training loop
for inputs, labels in data:
    output = model(inputs)
    loss = loss_fn(output, labels)
    loss.backward()
    optimizer.step()
    quantizer.step()

# In this example, from step 0 onwards, observers will collect statistics
# of the values of weights/activations. However, between steps 0 and 100,
# effects of quantization will not be simulated. At step 100, quantization
# simulation will begin and at step 300, observer statistics collection will
# stop. A batch norm layer computes mean and variance of input batch for normalizing
# it during training, and collects running estimates of its computed mean and variance,
# which are then used for normalization during evaluation. At step 200, batch norm
# statistics collection is frozen, and the batch norm layers switch to evaluation
# mode, thus more closely simulating the inference numerics during training time.

Parameters:

weight_dtype (torch.dtype) – The dtype to use for quantizing the weights. The number of bits used for quantization is inferred from the dtype. When dtype is set to torch.float32, the weights corresponding to that layer are not quantized. Defaults to torch.int8 which corresponds to 8-bit quantization.
weight_observer (ObserverType) – Type of observer to use for quantizing weights. Defaults to moving_average_min_max.
weight_per_channel (bool) – When True, weights are quantized per channel; otherwise, per tensor.
activation_dtype (torch.dtype) – The dtype to use for quantizing the activations. When dtype is set to torch.float32, the activations corresponding to that layer are not quantized. Defaults to torch.quint8.
activation_observer (ObserverType) – Type of observer to use for quantizing activations. Allowed values are min_max and moving_average_min_max. Defaults to moving_average_min_max.
quantization_scheme – (QuantizationScheme): Type of quantization configuration to use. When this parameter is set to QuantizationScheme.symmetric, all weights are quantized with zero point as zero, and activations are quantized with zero point as zero for non-negative activations and 128 for all other activations. When it is set to QuantizationScheme.affine, zero point can be set anywhere in the range of values allowed for the quantized weight/activation. Defaults to QuantizationScheme.symmetric.
milestones (list of int) – A list of four integers indicating milestones to use during quantization. The first milestone corresponds to enabling observers, the second to enabling fake quantization simulation, the third to disabling observers, and the last to freezing batch norm statistics. Defaults to None, which means the step method of LinearQuantizer will be a no-op and all observers and quantization simulation will be turned on from the first step, batch norm layers always operate in training mode, and mean and varaince statistics collection is not frozen.

as_dict() → Dict[str, Any]: Returns the config as a dictionary.

classmethod from_dict(config_dict)[source]

Create class from a dictionary of string keys and values.

Parameters:: data_dict (dict of str and values) – A nested dictionary of strings and values.

classmethod from_yaml(yml: IO | str) → DictableDataClass

Create class from a yaml stream.

Parameters:: yml – An IO stream containing yaml or a str path to the yaml file.

class coremltools.optimize.torch.quantization.LinearQuantizerConfig(global_config: ModuleLinearQuantizerConfig | None = None, module_type_configs: ModuleTypeConfigType = NOTHING, module_name_configs: Dict[str, ModuleLinearQuantizerConfig | None] = NOTHING, non_traceable_module_names: List[str] = [], preserved_attributes: List[str] = NOTHING)[source]

Configuration class for specifying how different submodules of a model are quantized by LinearQuantizer.

In order to disable quantizing a layer or an operation, module_type_config or module_name_config corresponding to that operation can be set to None.

For example:

# The following config will enable weight only quantization for all layers:
config = LinearQuantizerConfig.from_dict(
    {
        "global_config": {
            "activation_dtype": "float32",
        }
    }
)

# The following config will disable quantization for all linear layers and
# set quantization mode to weight only quantization for convolution layers:
config = LinearQuantizerConfig.from_dict(
    {
        "module_type_configs": {
            "Linear": None,
            "Conv2d": {
                "activation_dtype": "float32",
            },
        }
    }
)

# The following config will disable quantization for layers named conv1 and conv2:
config = LinearQuantizerConfig.from_dict(
    {
        "module_name_configs": {
            "conv1": None,
            "conv2": None,
        }
    }
)

# If model has some methods and attributes which are not used in the forward
# pass, but are needed to be preserved after quantization is added, they can
# be preserved on the quantized model by passing them in preserved_attributes
# parameter

model = MyModel()
model.key_1 = value_1
model.key_2 = value_2

config = LinearQuantizerConfig.from_dict({"preserved_attributes": ["key_1", "key_2"]})

Parameters:

global_config (ModuleLinearQuantizerConfig) – Config to be applied globally to all supported modules. Missing values are chosen from the default config.
module_type_configs (dict of str to ModuleLinearQuantizerConfig) – Module type level configs applied to a specific module class, such as torch.nn.Linear. The keys can be either strings or module classes.
module_name_configs (dict of str to ModuleLinearQuantizerConfig) – Module level configs applied to specific modules. The name of the module must be a fully qualified name that can be used to fetch it from the top level module using the module.get_submodule(target) method.
non_traceable_module_names (list of str) – Names of modules which cannot be traced using torch.fx.
preserved_attributes (list of str) – Names of attributes of the model which should be preserved on the prepared and finalized models, even if they are not used in the model’s forward pass.

Note

The quantization_scheme parameter must be the same across all configs.

as_dict() → Dict[str, Any]: Returns the config as a dictionary.

classmethod from_dict(config_dict: Dict[str, Any]) → LinearQuantizerConfig[source]

Create class from a dictionary of string keys and values.

Parameters:: config_dict (dict of str and values) – A nested dictionary of strings and values.

classmethod from_yaml(yml: IO | str) → DictableDataClass

Create class from a yaml stream.

Parameters:: yml – An IO stream containing yaml or a str path to the yaml file.

set_global(global_config: ModuleOptimizationConfig | None) → OptimizationConfig: Set the global config.

set_module_name(module_name: str, opt_config: ModuleOptimizationConfig | None) → OptimizationConfig: Set the module level optimization config for a given module instance. If the module level optimization config for an existing module was already set, the new config will override the old one.

set_module_type(object_type: Callable | str, opt_config: ModuleOptimizationConfig | None) → OptimizationConfig: Set the module level optimization config for a given module type. If the module level optimization config for an existing module type was already set, the new config will override the old one.

class coremltools.optimize.torch.quantization.LinearQuantizer(model: Module, config: LinearQuantizerConfig | None = None)[source]

Perform quantization aware training (QAT) of models. This algorithm simulates the effects of quantization during training, by quantizing and dequantizing the weights and/or activations during the model’s forward pass. The forward and backward pass computations are conducted in float dtype, however, these float values follow the constraints imposed by int8 and quint8 dtypes. Thus, this algorithm adjusts the model’s weights while closely simulating the numerics which get executed during quantized inference, allowing model’s weights to adjust to quantization constraints.

For more details, please refer to Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference.

Example

import torch.nn as nn
from coremltools.optimize.torch.quantization import (
    LinearQuantizer,
    LinearQuantizerConfig,
)

model = nn.Sequential(
    OrderedDict(
        {
            "conv": nn.Conv2d(1, 20, (3, 3)),
            "relu1": nn.ReLU(),
            "conv2": nn.Conv2d(20, 20, (3, 3)),
            "relu2": nn.ReLU(),
        }
    )
)

loss_fn = define_loss()

# initialize the quantizer
config = LinearQuantizerConfig.from_dict(
    {
        "global_config": {
            "quantization_scheme": "symmetric",
            "milestones": [0, 100, 400, 400],
        }
    }
)

quantizer = LinearQuantizer(model, config)

# prepare the model to insert FakeQuantize layers for QAT
model = quantizer.prepare()

# use quantizer in your PyTorch training loop
for inputs, labels in data:
    output = model(inputs)
    loss = loss_fn(output, labels)
    loss.backward()
    optimizer.step()
    quantizer.step()

# convert operations to their quanitzed counterparts using parameters learnt via QAT
model = quantizer.finalize(inplace=True)

Parameters:

model (torch.nn.Module) – Module to be trained.
config (_LinearQuantizerConfig) – Config that specifies how different submodules in the model will be quantized. Default config is used when passed as None.

finalize(model: Module | None = None, inplace: bool = False) → Module[source]

Prepares the model for export.

Parameters:

model (_torch.nn.Module) – Model to be finalized.
inplace (bool) – If True, model transformations are carried out in-place and the original module is mutated; otherwise, a copy of the model is mutated and returned.

Note

Once the model is finalized with in_place = True, it may not be runnable on the GPU.

prepare(example_inputs: Tuple[Any, ...], inplace: bool = False) → Module[source]

Prepares the model for quantization aware training by inserting torch.ao.quantization.FakeQuantize layers in the model in appropriate places.

Parameters:

example_inputs (Tuple[Any, ...]) – Example inputs for forward function of the model, tuple of positional args (keyword args can be passed as positional args as well)
inplace (bool) – If True, model transformations are carried out in-place and the original module is mutated, otherwise a copy of the model is mutated and returned.

Note

This method uses prepare_qat_fx method to insert quantization layers and the returned model is a torch.fx.GraphModule. Some models, like those with dynamic control flow, may not be trace-able into a torch.fx.GraphModule. Please follow directions in Limitations of Symbolic Tracing to update your model first before using LinearQuantizer algorithm.

report() → _Report[source]: Returns a dictionary with important statistics related to current state of quantization. Each key in the dictionary corresponds to a module name, and the value is a dictionary containing the statistics such as scale, zero point, number of parameters, and so on.

step()[source]

Steps through the milestones defined for this quantizer.

The first milestone corresponds to enabling observers, the second to enabling fake quantization simulation, the third to disabling observers, and the last to freezing batch norm statistics.

Note

If milestones argument is set as None, this method is a no-op.

Note

In order to not use a particular milestone, its value can be set as float('inf').

class coremltools.optimize.torch.quantization.ObserverType(value)[source]: An enum indicating the type of observer. Allowed options are moving_average_min_max and mix_max.

class coremltools.optimize.torch.quantization.QuantizationScheme(value)[source]: An enum indicating the type of quantization to be performed. Allowed options are symmetric and affine.

class coremltools.optimize.torch.quantization.ModulePostTrainingQuantizerConfig(weight_dtype: str | dtype = 'int8', granularity='per_channel', quantization_scheme=QuantizationScheme.symmetric, block_size=None)[source]

Configuration class for specifying global and module level quantizer options for PostTrainingQuantizer algorithm.

Parameters:

weight_dtype (torch.dtype) – The dtype to use for quantizing the weights. The number of bits used for quantization is inferred from the dtype. When dtype is set to torch.float32, the weights corresponding to that layer are not quantized. Defaults to torch.int8 which corresponds to 8-bit quantization.
granularity (QuantizationGranularity) – Specifies the granularity at which quantization parameters will be computed. Can be one of per_channel, per_tensor or per_block. When using per_block, block_size argument must be specified. Defaults to per_channel.
quantization_scheme (QuantizationScheme) – Type of quantization configuration to use. When this parameter is set to QuantizationScheme.symmetric, all weights are quantized with zero point as zero. When it is set to QuantizationScheme.affine, zero point can be set anywhere in the range of values allowed for the quantized weight. Defaults to QuantizationScheme.symmetric.
block_size (tuple of int or int) – When block_size is specified, block_size number of values will share the same quantization parameters of scale (and zero point if applicable) across the input-channel axis. A tuple of integers can be provided for arbitrary sized blockwise quantization. See below for more details on different possible configurations. Defaults to None.

This class supports few different configurations to structure the quantization:

1. Per-channel quantization: This is the default configuration where granularity is per_channel and block_size is None. In this configuration, quantization parameters are computed for each output channel.

2. Per-tensor quantization: In this configuration, quantization paramaters are computed for the tensor as a whole. That is, all values in the tensor will share a single scale (and a single zero point if applicable). The granularity argument is set to per_tensor.

3. Per-block quantization: This is used to structure the tensor for block-wise quantization. For this configuration, the granularity is set to per_block and the block_size argument has to be specified. The block_size argument can either be:

int: In this case, each row along the output-channel axis will have its own quantization parameters (similar to per_channel).
Additionally, block_size number of values will share the same quantization parameters, along the input-channel axis. For example, for a weight matrix of shape (10, 10), if we provide block_size = 2, the shape of the quantization parameters would be (10, 5).

tuple: For more advanced configuration, users can provide an arbitrary N-D shaped block to share the quantization parameters.
This is specified in the form of a tuple where each value corresponds to the block size for the respective axis of the weight matrix. The length of the provided tuple should be at most the number of dimensions of the weight matrix.

class coremltools.optimize.torch.quantization.PostTrainingQuantizer(model: Module, config: PostTrainingQuantizerConfig | None = None)[source]

Perform post-training quantization on a torch model. After quantization, weights of all submodules selected for quantization contain full precision values obtained by quantizing and dequantizing the original weights which captures the error induced by quantization.

Note

After quantization, the weight values stored will still remain in full precision, therefore the PyTorch model size will not be reduced. To see the reduction in model size, please convert the model using coremltools.convert(...), which will produce a MIL model containing the compressed weights.

Example:

import torch.nn as nn
from coremltools.optimize.torch.quantization import (
    PostTrainingQuantizerConfig,
    PostTrainingQuantizer,
)

model = nn.Sequential(
    OrderedDict(
        {
            "conv": nn.Conv2d(1, 20, (3, 3)),
            "relu1": nn.ReLU(),
            "conv2": nn.Conv2d(20, 20, (3, 3)),
            "relu2": nn.ReLU(),
        }
    )
)

# initialize the quantizer
config = PostTrainingquantizerConfig.from_dict(
    {
        "global_config": {
            "weight_dtype": "int8",
        },
    }
)

ptq = PostTrainingQuantizer(model, config)
quantized_model = ptq.compress()

Parameters:

model (torch.nn.Module) – Module to be compressed.
config (PostTrainingQuantizerConfig) – Config that specifies how different submodules in the model will be quantized.

class coremltools.optimize.torch.quantization.PostTrainingQuantizerConfig(global_config: ModulePostTrainingQuantizerConfig | None = None, module_type_configs: ModuleTypeConfigType = NOTHING, module_name_configs: Dict[str, ModulePostTrainingQuantizerConfig | None] = NOTHING)[source]

Configuration class for specifying how different submodules of a model should be post-training quantized by PostTrainingQuantizer.

Parameters:

global_config (ModulePostTrainingQuantizerConfig) – Config to be applied globally to all supported modules.
module_type_configs (dict of str to ModulePostTrainingQuantizerConfig) – Module type configs applied to a specific module class, such as torch.nn.Linear. The keys can be either strings or module classes.
module_name_configs (dict of str to ModulePostTrainingQuantizerConfig) – Module name configs applied to specific modules. This can be a dictionary with module names pointing to their corresponding :py:class:`ModulePostTrainingQuantizerConfig`s

class coremltools.optimize.torch.layerwise_compression.LayerwiseCompressorConfig(layers: List[Module | str] | ModuleList | None = None, global_config: LayerwiseCompressionAlgorithmConfig | None = None, module_type_configs: ModuleTypeConfigType = NOTHING, module_name_configs: Dict[str, LayerwiseCompressionAlgorithmConfig | None] = NOTHING, input_cacher: str = 'default', calibration_nsamples: int = 128)[source]

Configuration class for specifying how different submodules of a model are compressed by LayerwiseCompressor.

Only sequential models are supported.

Parameters:

layers (list of torch.nn.Module or str) – List of layers which should be compressed. The layer names can also be specified as a regex. The layers listed should be immediate child modules of the parent container torch.nn.Sequential model and they should be contiguous, i.e., output of layer n should be input of layer n+1.
global_config (ModuleGPTQConfig or ModuleSparseGPTConfig) – Config to be applied globally to all supported modules. Missing values are chosen from the default config.
module_type_configs (dict of str to ModuleGPTQConfig or ModuleSparseGPTConfig) – Module type configs applied to a specific module class, such as torch.nn.Linear. The keys can be either strings or module classes.
module_name_configs (dict of str to ModuleGPTQConfig or ModuleSparseGPTConfig) – Module level configs applied to specific modules. The name of the module must either be a regex or a fully qualified name that can be used to fetch it from the top level module using the module.get_submodule(target) method.
input_cacher (str or FirstLayerInputCacher) – Cacher object which caches inputs which are fed to the first layer which is set up for compression
calibration_nsamples (int) – Number of samples to be used for calibration.

as_dict() → Dict[str, Any]: Returns the config as a dictionary.

classmethod from_dict(config_dict: Dict[str, Any]) → LayerwiseCompressorConfig[source]

Create class from a dictionary of string keys and values.

Parameters:: config_dict (dict of str and values) – A nested dictionary of strings and values.

classmethod from_yaml(yml: IO | str) → DictableDataClass

Create class from a yaml stream.

Parameters:: yml – An IO stream containing yaml or a str path to the yaml file.

class coremltools.optimize.torch.layerwise_compression.LayerwiseCompressor(model: Module, config: LayerwiseCompressorConfig)[source]

A post training compression algorithm which compresses a sequential model layer by layer. The implementation supports two variations of this algorithm:

At a high level, it compresses weights of a model layer by layer, by minimizing the L2 norm of the difference between the original activations and activations obtained on compressing the weights of a layer. The activations are computed using a few samples of training data.

Only sequential models are supported, where output of one layer feeds into the input of the next layer.

For HuggingFace models, disable use_cache config. This is used to speed up decoding but, to generalize forward pass for LayerwiseCompressor algorithms across all model types we need to disable this behavior.

Example

import torch.nn as nn
from coremltools.optimize.torch.layerwise_compression import (
    LayerwiseCompressor,
    LayerwiseCompressorConfig,
)

model = nn.Sequential(
    OrderedDict(
        {
            "conv": nn.Conv2d(1, 20, (3, 3)),
            "relu1": nn.ReLU(),
            "conv2": nn.Conv2d(20, 20, (3, 3)),
            "relu2": nn.ReLU(),
        }
    )
)

dataloder = load_calibration_data()

# initialize the quantizer
config = LayerwiseCompressorConfig.from_dict(
    {
        "global_config": {
            "algorithm": "gptq",
            "weight_dtype": "int4",
        },
        "input_cacher": "default",
        "calibration_nsamples": 16,
    }
)

compressor = LayerwiseCompressor(model, config)

compressed_model = compressor.compress(dataloader)

Parameters:

model (torch.nn.Module) – Module to be compressed.
config (LayerwiseCompressorConfig) – Config that specifies how different submodules in the model will be compressed.

compress(dataloader: Iterable, device: str, inplace: bool = False) → Module[source]

Compresses model using samples from dataloader.

Parameters:

dataloader (Iterable) – An iterable where each element is an input to the model to be compressed.
device (str) – Device string for device to run compression on.
inplace (bool) – If True, model transformations are carried out in-place and the original module is mutated, otherwise a copy of the model is mutated and returned. Defaults to False.

GPTQ

class coremltools.optimize.torch.layerwise_compression.algorithms.ModuleGPTQConfig(weight_dtype: str | dtype = 'uint8', granularity='per_channel', quantization_scheme='symmetric', block_size: int | None = None, enable_normal_float: bool = False, hessian_dampening: float = 0.01, use_activation_order_heuristic: bool = False, processing_group_size: int = 128, algorithm: str = 'gptq')[source]

Bases: LayerwiseCompressionAlgorithmConfig

Configuration class for specifying global and module level compression options for GPTQ algorithm.

Parameters:

weight_dtype (torch.dtype) – The dtype to use for quantizing the weights. The number of bits used for quantization is inferred from the dtype. When dtype is set to torch.float32, the weights corresponding to that layer are not quantized. Defaults to torch.uint8 which corresponds to 8-bit quantization.
granularity (QuantizationGranularity) – Specifies the granularity at which quantization parameters will be computed. Can be one of per_channel, per_tensor or per_block. When using per_block, block_size argument must be specified. Defaults to per_channel.
quantization_scheme –

(QuantizationScheme): Type of
quantization configuration to use. When this parameter is set to QuantizationScheme.symmetric, all weights are quantized with zero point as zero. When it is set to QuantizationScheme.affine, zero point can be set anywhere in the range of values allowed for the quantized weight. Defaults to QuantizationScheme.symmetric.

block_size (int): When block_size is specified, block_size
number of values will share the same quantization parameters of scale (and zero point if applicable) across the input-channel axis. Defaults to None.
enable_normal_float (bool) – When True, normal float format is used for quantization. It’s only supported when weight_dtype is equal to int3 and int4. Defaults to False.
hessian_dampening – (float): Dampening factor added to the diagonal of the Hessian used by GPTQ algorithm. Defaults to 0.01.
use_activation_order_heuristic (bool) – When True, columns of weight are sorted in descending order of values of Hessian diagonal elements. Defaults to True.
processing_group_size (int) – The weights are updated in blocks of size processing_group_size. Defaults to 128.
.. note – Currently blocking is limited to only the input-channel axis for GPTQ.

class coremltools.optimize.torch.layerwise_compression.algorithms.GPTQ(layer: Module, config: ModuleGPTQConfig)[source]

Bases: OBSCompressionAlgorithm

A post training compression algorithm based on the paper GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers.

Parameters:

layer (torch.nn.Module) – Module to be compressed.
config (ModuleGPTQConfig) – Config specifying hyper-parameters for the GPTQ algorithm.