Dense feed-forward (MLP) composite component. More...

Inheritance diagram for Mila::Dnn::MLP< TDeviceType, TPrecision >:

Public Types
using	ComponentPtr = typename CompositeComponentBase::ComponentPtr
using	CompositeComponentBase = CompositeComponent<TDeviceType, TPrecision>
using	GeluType = Gelu<TDeviceType, TPrecision>
using	LinearType = Linear<TDeviceType, TPrecision>
using	MR = typename DeviceTypeTraits<TDeviceType>::memory_resource
using	TensorType = Tensor<TPrecision, MR>
Public Types inherited from Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >
using	ComponentBase = Component<TDeviceType, TPrecision>
using	ComponentPtr = std::shared_ptr<Component<TDeviceType, TPrecision>>

Public Member Functions
	MLP (const std::string &name, const MLPConfig &config, std::optional< DeviceId > device_id=std::nullopt)
	Construct an MLP component.
	~MLP () override=default
	Default destructor.
TensorType &	backward (const TensorType &input, const TensorType &output_grad)
	Backward pass using captured forward intermediates.
TensorType &	decode (const TensorType &input) const
	Single-token inference convenience: fc1 -> gelu -> fc2 with no gradient capture.
TensorType &	forward (const TensorType &input)
	Forward pass.
MemoryStats	getMemoryStats () const override
	Return the current memory allocation breakdown for this component.
const ComponentType	getType () const override
	Get the component type identifier.
std::string	toString () const override
	Human-readable status and configuration summary.
void	zeroGradients () override
	Zero gradients for all child components.
Public Member Functions inherited from Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >
	CompositeComponent (CompositeComponent &&) noexcept=default
	CompositeComponent (const CompositeComponent &)=delete
	CompositeComponent (const std::string &name)
	Construct composite component with name.
CompositeComponent &	addComponent (ComponentPtr component)
	Add a pre-constructed child component (chainable).
size_t	childCount () const noexcept
	Get the number of direct children.
void	clearComponents ()
	Clear all child components.
ComponentPtr	findComponent (const std::string &path) const
	Resolve a dot-separated component path within this composite.
ComponentPtr	getComponent (const std::string &name) const
	Retrieve a direct child component by name.
const std::vector< ComponentPtr > &	getComponents () const
	Get all child components in insertion order.
DeviceId	getDeviceId () const override
	Get the compute device for this composite.
std::vector< ITensor * >	getGradients () const override
	Get all parameter gradients from all children.
std::vector< ITensor * >	getParameters () const override
	Get all parameters from all children.
bool	hasChildren () const noexcept
	Check if this composite has any children.
bool	hasComponent (const std::string &name) const
	Check if a named child component exists.
CompositeComponent &	operator= (CompositeComponent &&) noexcept=default
CompositeComponent &	operator= (const CompositeComponent &)=delete
dim_t	parameterCount () const override
	Count parameters across all children.
bool	removeComponent (const std::string &name)
void	saveFlatTensors (Serialization::SafeTensorsWriter &writer, const std::string &prefix, Serialization::TensorSavePass pass) const override
	Recurse into children, extending the flat dotted prefix.
void	synchronize () override
	Synchronize all child components.
ComponentPtr	tryFindComponent (const std::string &path) const
	Try to resolve a dot-separated component path within this composite.
Public Member Functions inherited from Mila::Dnn::Component< TDeviceType, TPrecision >
	Component (const std::string &name)
	Construct component with required name identifier.
virtual void	build (const BuildContext &context) final
	Build the component with the provided BuildContext (canonical overload).
const std::string	getName () const
	Get the component's name identifier.
virtual std::vector< std::string >	getParameterNames () const
	List all available parameter names for this component.
virtual MemoryStats	getRequiredMemory (const BuildContext &context) const
	Report what build( context ) would allocate, without allocating it.
TrainingMode	getTrainingMode () const noexcept
	The current runtime behavioral mode of this Component.
virtual bool	isBuilt () const final
	Returns true if build() has completed successfully.
virtual void	loadParameter (const std::string &, const Serialization::ITensorBlob &)
	Load a parameter from serialized tensor data.
void	setTrainingMode (TrainingMode mode)
	Set the runtime behavioral mode for this Component.

Protected Member Functions
void	onBuilding (const BuildContext &context) override
	Build-time callback invoked by the CompositeComponent framework.
void	onTrainingModeChanging (TrainingMode training_mode) override
	Propagate training-mode changes to child components.
Protected Member Functions inherited from Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >
template<typename TComponent>
std::shared_ptr< TComponent >	getComponentAs (const std::string &name) const
	Retrieve a typed child component by name.
void	load_ (ModelArchive &archive, SerializationMode mode) override
	Restore children from their nested scopes, mirroring save_().
void	onExecutionContextSet () override
	Hook invoked after ExecutionContext is set.
virtual void	optimize ()
	Virtual hook for graph optimization after construction.
void	requireSerializableParameters () const override
	No-op override: a composite names no parameters of its own.
void	save_ (ModelArchive &archive, SerializationMode mode) const override
Protected Member Functions inherited from Mila::Dnn::Component< TDeviceType, TPrecision >
IExecutionContext *	getExecutionContext () const
	Get the shared execution context.
bool	hasExecutionContext () const noexcept
	Check if execution context has been set.
template<TensorDataType TParameterPrecision, typename TMemoryResource>
void	loadParameterFromBlob (const std::string &param_name, const Serialization::ITensorBlob &blob, Tensor< TParameterPrecision, TMemoryResource > &target, const shape_t &expected_shape)
	Load a tensor blob into a parameter tensor with validation.
template<TensorDataType TParameterPrecision, typename TMemoryResource>
void	saveParameterToArchive (ModelArchive &archive, const std::string &parameter_name, const Tensor< TParameterPrecision, TMemoryResource > &parameter) const
	Write one parameter tensor into the archive under "tensors/<name>".
template<TensorDataType TParameterPrecision, typename TMemoryResource>
void	saveParameterToWriter (Serialization::SafeTensorsWriter &writer, const std::string &flat_name, const Tensor< TParameterPrecision, TMemoryResource > &parameter, Serialization::TensorSavePass pass) const
	Drive one parameter through one pass of a flat safetensors save.
void	setExecutionContext (IExecutionContext *context)
	Set the execution context for this component.

Additional Inherited Members
Static Public Member Functions inherited from Mila::Dnn::Component< TDeviceType, TPrecision >
static constexpr DeviceType	getDeviceType ()
	Compile-time device type for this component instance.
static constexpr TensorDataType	getPrecision () noexcept
	Compile-time tensor precision for this component instance.
Protected Types inherited from Mila::Dnn::Component< TDeviceType, TPrecision >
using	HostStagingMemoryResource
	Host memory a device-resident parameter stages through.
Protected Attributes inherited from Mila::Dnn::Component< TDeviceType, TPrecision >
BuildContext	build_context_ { shape_t{ 1 }, RuntimeMode::Training }
	The BuildContext stored at build time.

Detailed Description

template<DeviceType TDeviceType, TensorDataType TPrecision>
requires PrecisionSupportedOnDevice<TPrecision, TDeviceType>
class Mila::Dnn::MLP< TDeviceType, TPrecision >

Dense feed-forward (MLP) composite component.

Device-templated composite implementing the GPT-2 dense FFN: Input -> Linear(in_features, hidden_size) -> GELU -> Linear(hidden_size, in_features) -> Output

The component is honest about the single activation it supports: GELU. The gated FFN family (SwiGLU and friends) is a separate component (GatedMLP) with a different 2H -> H shape contract; it is not expressible here and is not a runtime option. A future generalized elementwise Activation component will replace the fixed Gelu child, at which point the activation function becomes a compile-time parameter (see Specifications/FfnAndMoE.md). Until then MLP is the dense GELU FFN, full stop.

The component composes child components (two Linear projections and a GELU) and delegates forward/backward calls to them. Child components own their intermediate tensors; MLP stores non-owning pointers to those tensors after forward() to chain backward().

Threading: call sites must ensure that forward/backward/zeroGradients are invoked in a thread-safe manner relative to one another; this class does not provide internal synchronization.

Template Parameters

TDeviceType	Device type for execution (CPU, CUDA, ...).
TPrecision	Tensor data precision. Must be supported on the device.

Constructor & Destructor Documentation

◆ MLP()

template<DeviceType TDeviceType, TensorDataType TPrecision>

Mila::Dnn::MLP< TDeviceType, TPrecision >::MLP	(	const std::string &	name,
		const MLPConfig &	config,
		std::optional< DeviceId >	device_id = std::nullopt )

inlineexplicit

Construct an MLP component.

The constructor validates the provided config, constructs the internal child component graph (fc1 -> gelu -> fc2), and optionally creates and assigns an execution context when device_id is provided.

Parameters

name	Component name used to name child subcomponents.
config	MLP configuration (input features, hidden size, bias).
device_id	Optional device identifier; when present the MLP creates an owned execution context bound to that device and sets it on the component. If the provided device_id type does not match the template TDeviceType, an exception is thrown.

Exceptions

std::invalid_argument	if config is invalid (via config.validate()).
std::invalid_argument	if device_id is present but has a mismatched device type.

◆ ~MLP()

template<DeviceType TDeviceType, TensorDataType TPrecision>

Mila::Dnn::MLP< TDeviceType, TPrecision >::~MLP ( )

overridedefault

Default destructor.

Child components are stored as shared_ptr and will be destroyed automatically.

Member Function Documentation

◆ backward()

template<DeviceType TDeviceType, TensorDataType TPrecision>

TensorType & Mila::Dnn::MLP< TDeviceType, TPrecision >::backward	(	const TensorType &	input,
		const TensorType &	output_grad )

inline

Backward pass using captured forward intermediates.

Uses the child-owned tensors captured by the most recent forward() invocation to chain backward calls without recomputing forward:

fc2_->backward(captured_activation_output, output_grad)
gelu_->backward(...)
fc1_->backward(input, ...)

The method clears the cached forward pointers before returning to avoid accidental reuse.

Preconditions:

Component must be built.
forward() must have been called previously to populate internal forward caches.

Parameters

input	The original input tensor passed to forward(); required by fc1_->backward.
output_grad	Gradient tensor w.r.t. the MLP output.

Returns: Reference to the input-gradient tensor (owned by the fc1 child).

Exceptions

std::runtime_error if the component is not built or if forward() was not called.

◆ decode()

template<DeviceType TDeviceType, TensorDataType TPrecision>

TensorType & Mila::Dnn::MLP< TDeviceType, TPrecision >::decode ( const TensorType & input ) const

inline

Single-token inference convenience: fc1 -> gelu -> fc2 with no gradient capture.

Relies on single-stream ordering for inter-op dependencies; the caller synchronizes before reading results on the host.

◆ forward()

template<DeviceType TDeviceType, TensorDataType TPrecision>

TensorType & Mila::Dnn::MLP< TDeviceType, TPrecision >::forward ( const TensorType & input )

inline

Forward pass.

Chains child component forward calls:

fc1_->forward(input)
gelu_->forward(...)
fc2_->forward(...)

The function stores non-owning pointers to child-owned intermediate tensors produced during the forward call; these pointers are used by backward() to chain gradients.

Preconditions:

Component must be built (onBuilding called).
Input tensor must be bound to the same device/context as the component.

Parameters

input Input tensor bound to this component's device/context.

Returns: Reference to the output tensor produced by the final Linear child (owned by that child).

Exceptions

std::runtime_error if the component is not built prior to calling forward.

◆ getMemoryStats()

template<DeviceType TDeviceType, TensorDataType TPrecision>

MemoryStats Mila::Dnn::MLP< TDeviceType, TPrecision >::getMemoryStats ( ) const

inlineoverridevirtual

Return the current memory allocation breakdown for this component.

Reflects allocations at the moment of the call. The returned stats naturally track the component lifecycle:

After construction – nothing; construction allocates none After build( Inference ) – parameters + T=1 state buffers After build( Training ) – parameters + T=full state buffers After setTrainingMode( Train ) – parameters + state + gradients

For CompositeComponent and Network, the returned stats are the recursive aggregate of all child components.

May be called at any time – no lifecycle preconditions.

Returns: MemoryStats reflecting current allocations.

Implements Mila::Dnn::Component< TDeviceType, TPrecision >.

◆ getType()

template<DeviceType TDeviceType, TensorDataType TPrecision>

const ComponentType Mila::Dnn::MLP< TDeviceType, TPrecision >::getType ( ) const

inlineoverridevirtual

Get the component type identifier.

Used for serialization and runtime type identification.

Returns: Component type enum value.

Implements Mila::Dnn::Component< TDeviceType, TPrecision >.

◆ onBuilding()

template<DeviceType TDeviceType, TensorDataType TPrecision>

void Mila::Dnn::MLP< TDeviceType, TPrecision >::onBuilding ( const BuildContext & context )

inlineoverrideprotectedvirtual

Build-time callback invoked by the CompositeComponent framework.

Validates the input shape, computes the hidden shape, and builds each child with the appropriate shape: fc1 receives the input shape, the GELU and fc2 receive the hidden shape. After building, cached forward pointers are cleared.

Parameters

context Build context; inputShape().back() must equal config_.getInputFeatures().

Exceptions

std::invalid_argument if the input shape rank < 1 or last dimension mismatches config.

Reimplemented from Mila::Dnn::Component< TDeviceType, TPrecision >.

◆ onTrainingModeChanging()

template<DeviceType TDeviceType, TensorDataType TPrecision>

void Mila::Dnn::MLP< TDeviceType, TPrecision >::onTrainingModeChanging ( TrainingMode training_mode )

inlineoverrideprotectedvirtual

Propagate training-mode changes to child components.

Parameters

training_mode New training mode.

Reimplemented from Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >.

◆ toString()

template<DeviceType TDeviceType, TensorDataType TPrecision>

std::string Mila::Dnn::MLP< TDeviceType, TPrecision >::toString ( ) const

inlineoverridevirtual

Human-readable status and configuration summary.

Returns: String containing component introspection information suitable for logging.

Reimplemented from Mila::Dnn::CompositeComponent< TDeviceType, TPrecision >.

◆ zeroGradients()

template<DeviceType TDeviceType, TensorDataType TPrecision>

void Mila::Dnn::MLP< TDeviceType, TPrecision >::zeroGradients ( )

inlineoverridevirtual

Zero gradients for all child components.

Reimplemented from Mila::Dnn::Component< TDeviceType, TPrecision >.

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

Mila/Src/Dnn/Components/FFN/MLP/MLP.ixx

Public Types

Public Member Functions

Protected Member Functions

Additional Inherited Members

Detailed Description

Constructor & Destructor Documentation

◆ MLP()

◆ ~MLP()

Member Function Documentation

◆ backward()

◆ decode()

◆ forward()

◆ getMemoryStats()

◆ getType()

◆ onBuilding()

◆ onTrainingModeChanging()

◆ toString()

◆ zeroGradients()