TransformerBlock implements a standard transformer encoder block. More...

Inheritance diagram for Mila::Dnn::TransformerBlock< TDeviceType, TDataType >:

[legend]

Collaboration diagram for Mila::Dnn::TransformerBlock< TDeviceType, TDataType >:

[legend]

Public Types
using	CompositeModuleBase = CompositeModule< TDeviceType, TDataType >
	Alias for base module type.

using	MR = std::conditional_t< TDeviceType==DeviceType::Cuda, CudaMemoryResource, CpuMemoryResource >
	Memory resource type used for tensors, selected based on device type.

Public Types inherited from Mila::Dnn::CompositeModule< TDeviceType, TDataType >
using	ModuleBase = Module< TDeviceType, TDataType, TDataType >
	Base class type for the module.

using	MR = std::conditional_t< TDeviceType==DeviceType::Cuda, CudaMemoryResource, HostMemoryResource >
	Memory resource type based on device type.

Public Types inherited from Mila::Dnn::Module< TDeviceType, TInput, TOutput >
using	MR = std::conditional_t< TDeviceType==DeviceType::Cuda, CudaMemoryResource, CpuMemoryResource >

Public Member Functions
	TransformerBlock (const std::string &device_name, const TransformerBlockConfig &config)
	Constructs a new TransformerBlock module with a device name.

	TransformerBlock (std::shared_ptr< DeviceContext > device_context, const TransformerBlockConfig &config)
	Constructs a new TransformerBlock module with a provided device context.

void	forward (const Tensor< TDataType, MR > &input, Tensor< TDataType, MR > &output)
	Performs the forward pass of the TransformerBlock.

void	load (ModelArchive &archive) override
	Deserializes the module state from a ZIP archive.

size_t	parameterCount () const override
	Gets the number of trainable parameters in this module.

void	save (ModelArchive &archive) const override
	Serializes the module state to a ZIP archive.

std::string	toString () const override
	Generates a string representation of this module's configuration.

Public Member Functions inherited from Mila::Dnn::CompositeModule< TDeviceType, TDataType >
	CompositeModule ()
	Default constructor.

	CompositeModule (const std::string &device_name, const ComponentConfig &config)
	Constructor with device name.

	CompositeModule (std::shared_ptr< DeviceContext > context, const ComponentConfig &config)
	Constructor with device context.

virtual	~CompositeModule ()=default
	Virtual destructor.

CompositeModule &	addModule (const std::string &name, std::shared_ptr< Module< TDeviceType, TDataType, TDataType > > module)
	Add a named child module to this module.

CompositeModule &	addModule (std::shared_ptr< Module< TDeviceType, TDataType, TDataType > > module)
	Add an unnamed child module to this module.

std::shared_ptr< Module< TDeviceType, TDataType, TDataType > >	getModule (const std::string &name) const
	Get a specific sub-module by name.

const std::vector< std::shared_ptr< Module< TDeviceType, TDataType, TDataType > > > &	getModules () const
	Get all sub-modules contained in this module.

const std::unordered_map< std::string, std::shared_ptr< Module< TDeviceType, TDataType, TDataType > > > &	getNamedModules () const
	Get all named sub-modules contained in this module.

bool	hasModule (const std::string &name) const
	Check if a sub-module with the given name exists.

bool	removeModule (const std::string &name)
	Remove a sub-module by name.

bool	replaceModule (const std::string &name, std::shared_ptr< Module< TDeviceType, TDataType, TDataType > > module)
	Replace an existing sub-module with a new one.

void	setTraining (bool is_training) override
	Set the training mode for this module and all its sub-modules.

Public Member Functions inherited from Mila::Dnn::Module< TDeviceType, TInput, TOutput >
	Module (const std::string &device_name, const ComponentConfig &config)
	Constructor with device name.

	Module (std::shared_ptr< DeviceContext > context, const ComponentConfig &config)
	Constructor with a specific device context.

virtual	~Module ()=default
	Virtual destructor for proper cleanup in derived classes.

std::shared_ptr< Compute::DeviceContext >	getDeviceContext () const
	Get the device context for this module.

Compute::DeviceType	getDeviceType () const
	Get the device type of the current device context.

std::string	getName () const
	Get the name of the module.

const auto &	getParameterTensors () const
	Get the parameter tensors of this module.

const ComputePrecision::Policy &	getPrecision () const

const auto &	getStateTensors () const
	Get the state tensors of this module.

bool	isTraining () const
	Check if the module is in training mode.

Private Member Functions
void	initializeModules ()
	Initializes the sub-modules and output tensors for the transformer block.

Private Attributes
std::shared_ptr< MultiHeadAttention< TDeviceType, TDataType > >	attn_block_ { nullptr }
	Multi-head self-attention block including projections.

Tensor< TDataType, MR >	attn_output_
	Output tensor from attention block.

TransformerBlockConfig	config_
	Configuration for the TransformerBlock module.

std::shared_ptr< Dropout< TDeviceType, TDataType > >	dropout_ { nullptr }
	Optional dropout module.

std::shared_ptr< LayerNorm< TDeviceType, TDataType > >	ln_1_ { nullptr }
	First layer normalization module.

Tensor< TDataType, MR >	ln_1_output_
	Output tensor from first layer normalization.

std::shared_ptr< LayerNorm< TDeviceType, TDataType > >	ln_2_ { nullptr }
	Second layer normalization module.

Tensor< TDataType, MR >	ln_2_output_
	Output tensor from second layer normalization.

std::shared_ptr< MLP< TDeviceType, TDataType > >	mlp_ { nullptr }
	Feed-forward network (MLP).

Tensor< TDataType, MR >	mlp_output_
	Output tensor from MLP.

Tensor< TDataType, MR >	res_1_output_
	Output tensor from first residual connection.

Tensor< TDataType, MR >	res_2_output_
	Output tensor from second residual connection.

Additional Inherited Members
Protected Member Functions inherited from Mila::Dnn::Module< TDeviceType, TInput, TOutput >
const std::string	parametersToString () const
	Helper method to convert parameters to string representation.

const std::string	stateToString () const
	Helper method to convert state tensors to string representation.

Protected Attributes inherited from Mila::Dnn::Module< TDeviceType, TInput, TOutput >
std::unordered_map< std::string, std::shared_ptr< Tensor< TOutput, MR > > >	parameter_map_ = {}
	Map of parameter names to parameter tensors.

std::unordered_map< std::string, std::shared_ptr< Tensor< TOutput, MR > > >	state_map_ = {}
	Map of state names to state tensors.

Detailed Description

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>
requires ValidFloatTensorType<TDataType>
class Mila::Dnn::TransformerBlock< TDeviceType, TDataType >

TransformerBlock implements a standard transformer encoder block.

The transformer block consists of:

Multi-head self-attention mechanism with residual connection
Feed-forward network (MLP) with residual connection
Layer normalization before or after each sub-block (configurable)

This is the fundamental building block of transformer architectures like BERT and GPT. The implementation supports both pre-LN (more stable) and post-LN (original) architectures, configurable dropout rates, and other hyperparameters.

Template Parameters

TDeviceType	The device type (CPU or CUDA) on which to perform computations.
TDataType	The data type used for tensor elements throughout the network.

Member Typedef Documentation

◆ CompositeModuleBase

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

using Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::CompositeModuleBase = CompositeModule<TDeviceType, TDataType>

export

Alias for base module type.

◆ MR

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

using Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::MR = std::conditional_t<TDeviceType == DeviceType::Cuda, CudaMemoryResource, CpuMemoryResource>

export

Memory resource type used for tensors, selected based on device type.

Constructor & Destructor Documentation

◆ TransformerBlock() [1/2]

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::TransformerBlock	(	const std::string &	device_name,
		const TransformerBlockConfig &	config
	)

inlineexplicitexport

Constructs a new TransformerBlock module with a device name.

Creates a new DeviceContext internally using the provided device name. This constructor is useful for creating standalone modules without pre-existing device contexts.

Parameters

device_name	The name of the device to use (e.g., "CPU", "CUDA:0").
config	Configuration parameters for the TransformerBlock module.

Exceptions

std::invalid_argument	If the device name is invalid or the configuration is invalid
std::runtime_error	If device type doesn't match template parameter TDeviceType

Here is the call graph for this function:

◆ TransformerBlock() [2/2]

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::TransformerBlock	(	std::shared_ptr< DeviceContext >	device_context,
		const TransformerBlockConfig &	config
	)

inlineexplicitexport

Constructs a new TransformerBlock module with a provided device context.

Uses a pre-existing DeviceContext instance. This constructor is useful when integrating the module into a larger network that shares device contexts across modules.

Parameters

device_context	The device context to use for this module.
config	Configuration parameters for the TransformerBlock module.

Exceptions

std::invalid_argument	If device_context is null or configuration is invalid
std::runtime_error	If device context type doesn't match template parameter TDeviceType

Here is the call graph for this function:

Member Function Documentation

◆ forward()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

void Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::forward	(	const Tensor< TDataType, MR > &	input,
		Tensor< TDataType, MR > &	output
	)

inlineexport

Performs the forward pass of the TransformerBlock.

The forward pass follows either pre-LN or post-LN architecture based on configuration:

Pre-LN (default):

Layer normalization 1
Self-attention
Residual connection
Layer normalization 2
Feed-forward network
Residual connection

Post-LN:

Self-attention
Residual connection
Layer normalization 1
Feed-forward network
Residual connection
Layer normalization 2

Parameters

input	The input tensor to be processed.
output	The output tensor where the results will be stored.

Here is the call graph for this function:

◆ initializeModules()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

void Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::initializeModules ( )

inlineexportprivate

Initializes the sub-modules and output tensors for the transformer block.

Creates and configures all components of the transformer block according to the configuration, including layer norm, attention, and feed-forward network.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ load()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

void Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::load ( ModelArchive & archive )

inlineoverrideexportvirtual

Deserializes the module state from a ZIP archive.

Loads the state of all sub-modules from the provided ZIP archive.

Parameters

zip	The ZIP archive to load the module state from.

Reimplemented from Mila::Dnn::CompositeModule< TDeviceType, TDataType >.

Here is the call graph for this function:

◆ parameterCount()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

size_t Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::parameterCount ( ) const

inlineoverrideexportvirtual

Gets the number of trainable parameters in this module.

Counts the total number of parameters in all sub-modules.

Returns: size_t The total number of parameters.

Reimplemented from Mila::Dnn::CompositeModule< TDeviceType, TDataType >.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ save()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

void Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::save ( ModelArchive & archive ) const

inlineoverrideexportvirtual

Serializes the module state to a ZIP archive.

Saves the state of all sub-modules to the provided ZIP archive.

Parameters

zip	The ZIP archive to save the module state to.

Reimplemented from Mila::Dnn::CompositeModule< TDeviceType, TDataType >.

Here is the call graph for this function:

◆ toString()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

std::string Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::toString ( ) const

inlineoverrideexportvirtual

Generates a string representation of this module's configuration.

Returns: std::string A formatted string with module information

Reimplemented from Mila::Dnn::CompositeModule< TDeviceType, TDataType >.

Here is the call graph for this function:

Member Data Documentation

◆ attn_block_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

std::shared_ptr<MultiHeadAttention<TDeviceType, TDataType> > Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::attn_block_ { nullptr }

exportprivate

Multi-head self-attention block including projections.

◆ attn_output_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

Tensor<TDataType, MR> Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::attn_output_

exportprivate

Output tensor from attention block.

◆ config_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

TransformerBlockConfig Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::config_

exportprivate

Configuration for the TransformerBlock module.

◆ dropout_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

std::shared_ptr<Dropout<TDeviceType, TDataType> > Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::dropout_ { nullptr }

exportprivate

Optional dropout module.

◆ ln_1_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

std::shared_ptr<LayerNorm<TDeviceType, TDataType> > Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::ln_1_ { nullptr }

exportprivate

First layer normalization module.

In pre-LN architecture, applied before attention. In post-LN architecture, applied after attention and residual connection.

◆ ln_1_output_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

Tensor<TDataType, MR> Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::ln_1_output_

exportprivate

Output tensor from first layer normalization.

◆ ln_2_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

std::shared_ptr<LayerNorm<TDeviceType, TDataType> > Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::ln_2_ { nullptr }

exportprivate

Second layer normalization module.

In pre-LN architecture, applied before MLP. In post-LN architecture, applied after MLP and residual connection.

◆ ln_2_output_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

Tensor<TDataType, MR> Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::ln_2_output_

exportprivate

Output tensor from second layer normalization.

◆ mlp_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

std::shared_ptr<MLP<TDeviceType, TDataType> > Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::mlp_ { nullptr }

exportprivate

Feed-forward network (MLP).

◆ mlp_output_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

Tensor<TDataType, MR> Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::mlp_output_

exportprivate

Output tensor from MLP.

◆ res_1_output_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

Tensor<TDataType, MR> Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::res_1_output_

exportprivate

Output tensor from first residual connection.

◆ res_2_output_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TDataType = float>

Tensor<TDataType, MR> Mila::Dnn::TransformerBlock< TDeviceType, TDataType >::res_2_output_

exportprivate

Output tensor from second residual connection.

The documentation for this class was generated from the following file:

/home/runner/work/Mila/Mila/Mila/Src/Dnn/Modules/Blocks/TransformerBlock.ixx

Public Types

Public Member Functions

Private Member Functions

Private Attributes

Additional Inherited Members

Detailed Description

Member Typedef Documentation

◆ CompositeModuleBase

◆ MR

Constructor & Destructor Documentation

◆ TransformerBlock() [1/2]

◆ TransformerBlock() [2/2]

Member Function Documentation

◆ forward()

◆ initializeModules()

◆ load()

◆ parameterCount()

◆ save()

◆ toString()

Member Data Documentation

◆ attn_block_

◆ attn_output_

◆ config_

◆ dropout_

◆ ln_1_

◆ ln_1_output_

◆ ln_2_

◆ ln_2_output_

◆ mlp_

◆ mlp_output_

◆ res_1_output_

◆ res_2_output_