Root composite network container. More...

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

Public Types
using	ComponentPtr = typename CompositeBase::ComponentPtr
using	CompositeBase = CompositeComponent<TDeviceType, TPrecision>
using	MR = typename DeviceTypeTraits<TDeviceType>::memory_resource
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
	Network (const std::string &name)
	Construct network (context managed by derived class).
template<typename TOptimizer, typename TConfig>
std::shared_ptr< TOptimizer >	createOptimizer (const TConfig &config)
	Create and configure an optimizer for this network's parameters.
DeviceId	getDeviceId () const noexcept override
	Get the compute device for this composite.
IExecutionContext *	getExecutionContext () const
	Public access to the network's shared execution context.
const ComponentType	getType () const override
	Get the component type identifier.
void	load (ModelArchive &archive, SerializationMode mode)
	Restore this network's parameters from an archive.
void	save (ModelArchive &archive, SerializationMode mode) const
	Save network to archive.
void	synchronize () override
	Synchronize all child components.
std::string	toString () const override
	Generate a human-readable description.
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.
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.
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).
virtual MemoryStats	getMemoryStats () const =0
	Return the current memory allocation breakdown for this component.
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.
virtual void	zeroGradients ()
	Clear all model-owned gradients for this component.

Protected Member Functions
virtual void	load_ (ModelArchive &archive, SerializationMode mode) override
	Hook for concrete classes to validate type-specific state on load.
virtual void	save_ (ModelArchive &archive, SerializationMode mode) const override=0
	Hook for concrete classes to save type-specific state.
void	verifyArchitectureCompatibility (const PretrainedMetadata &metadata)
	Verify that imported model is compatible with network architecture.
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	onExecutionContextSet () override
	Hook invoked after ExecutionContext is set.
void	onTrainingModeChanging (TrainingMode training_mode) override
	Hook invoked when training mode is about to change.
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.
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.
virtual void	onBuilding (const BuildContext &)
	Hook invoked by build() to allocate component buffers.
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>
class Mila::Dnn::Network< TDeviceType, TPrecision >

Root composite network container.

Network is a specialized CompositeComponent that represents a complete neural network model and serves as the top-level entry point. It provides high-level serialization semantics while delegating lifecycle management to concrete subclasses.

Ownership Model:

Concrete subclasses own ExecutionContext (flexibility for construction ordering)
Network receives non-owning context pointer via setExecutionContext()
All child components share the Network's ExecutionContext (efficient resource pooling)

Construction Pattern (Concrete Subclass):

class MnistClassifier : public Network<DeviceType::Cpu, TensorDataType::FP32>
{
public:
    explicit MnistClassifier(const std::string& name, int64_t batch_size, DeviceId device_id)
        : Network(name),
          owned_context_(createExecutionContext(device_id)),
          batch_size_(batch_size)
    {
        // 1. Create component graph (context-independent)
        createGraph();
        
        // 2. Propagate context to self and children
        this->setExecutionContext(owned_context_.get());
    }
 
private:
    std::unique_ptr<IExecutionContext> owned_context_;  // Concrete class owns context
};

Serialization Contract:

Base class (Network): Saves component graph topology and generic metadata
Concrete class: MUST override save_() to write type identifier and configuration
Concrete class: MUST provide static Load() factory method for deserialization

Deserialization Pattern (Concrete Subclass):

// REQUIRED: Static factory method for type-safe deserialization
static std::unique_ptr<MnistClassifier> Load(ModelArchive& archive, DeviceId device_id)
{
    // 1. Read concrete-specific metadata
    json meta = archive.readJson("network/classifier_meta.json");
    std::string name = meta.at("name");
    int64_t batch_size = meta.at("batch_size");
    
    // 2. Construct via normal constructor path
    auto classifier = std::make_unique<MnistClassifier>(name, batch_size, device_id);
    
    // 3. Build with saved input shape
    shape_t input_shape = meta.at("input_shape");
    classifier->build(input_shape);
    
    // 4. Load component weights
    // (Base class handles graph traversal; weights loaded into already-built components)
    
    return classifier;
}

Design Rationale:

Concrete classes control infrastructure (context) lifecycle
Network base class focuses on container semantics and serialization
Clear initialization order: create context ? pass to base ? build graph
Enables future flexibility (custom contexts, multi-device, etc.)
Type-safe deserialization via concrete class Load() methods

Constructor & Destructor Documentation

◆ Network()

template<DeviceType TDeviceType, TensorDataType TPrecision>

Mila::Dnn::Network< TDeviceType, TPrecision >::Network ( const std::string & name )

inlineexplicit

Construct network (context managed by derived class).

Base constructor for concrete network classes. Derived classes are responsible for:

Creating and owning ExecutionContext
Building the component graph via createGraph()
Calling setExecutionContext() to propagate context to children

Parameters

name	Network name for identification and serialization

Exceptions

std::invalid_argument if name is not a valid identifier

Member Function Documentation

◆ createOptimizer()

template<DeviceType TDeviceType, TensorDataType TPrecision>

template<typename TOptimizer, typename TConfig>

std::shared_ptr< TOptimizer > Mila::Dnn::Network< TDeviceType, TPrecision >::createOptimizer ( const TConfig & config )

inline

Create and configure an optimizer for this network's parameters.

Factory method that creates an optimizer, enables training mode on the network, and registers all network parameters and gradients in a single atomic operation.

Lifecycle:

Enables training mode (allocates gradients for all components)
Creates optimizer using network's ExecutionContext
Registers all network parameters and gradients
Returns ready-to-use optimizer

Usage Pattern:

// Build network
mnist_net->build(input_shape);
 
// Create optimizer in one step
auto optimizer = mnist_net->createOptimizer<AdamWOptimizer<DeviceType::Cuda, TensorDataType::FP32>>(
    AdamWConfig()
        .withLearningRate(0.001f)
        .withWeightDecay(0.01f)
);
 
// Optimizer is ready to use
optimizer->step();

Template Parameters

TOptimizer	Optimizer type (e.g., AdamWOptimizer, SGD)
TConfig	Optimizer configuration type

Parameters

config Optimizer configuration

Returns: Shared pointer to configured and ready-to-use optimizer

Exceptions

std::runtime_error	if network is not built
std::runtime_error	if parameter/gradient count mismatch

Note: This method automatically calls setTraining(true), so explicit training mode activation is not required.

◆ getDeviceId()

template<DeviceType TDeviceType, TensorDataType TPrecision>

DeviceId Mila::Dnn::Network< TDeviceType, TPrecision >::getDeviceId ( ) const

inlineoverridevirtualnoexcept

Get the compute device for this composite.

Returns the device from the shared execution context.

Returns: DeviceId for this composite and its children

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

◆ getExecutionContext()

template<DeviceType TDeviceType, TensorDataType TPrecision>

IExecutionContext * Mila::Dnn::Network< TDeviceType, TPrecision >::getExecutionContext ( ) const

inline

Public access to the network's shared execution context.

Exposes the Component-level context so model-level orchestrator tools (e.g. TokenSampler) can be constructed on the network's context. Wraps the protected Component accessor.

◆ getType()

template<DeviceType TDeviceType, TensorDataType TPrecision>

const ComponentType Mila::Dnn::Network< 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 >.

◆ load()

template<DeviceType TDeviceType, TensorDataType TPrecision>

void Mila::Dnn::Network< TDeviceType, TPrecision >::load	(	ModelArchive &	archive,
		SerializationMode	mode )

inline

Restore this network's parameters from an archive.

The inverse of save(). Restores into the LIVE graph: the network must already be constructed with the same topology and built, which is why this is a member and not a factory. Concrete models reconstruct themselves through their own fromCheckpoint() – read the config, construct, build, then call this.

The child manifest is not replayed. saveComponentGraph records names into per-child descriptor files rather than a list, so the live children are the authoritative enumeration; network/architecture.json is read only to check the count, which catches the common "loaded into a different model" mistake before any tensor is touched.

Parameters

archive	Archive to read from.
mode	Serialization mode.

Exceptions

std::runtime_error if the archive's component count disagrees with this network, or if any component's blobs are missing.

◆ load_()

template<DeviceType TDeviceType, TensorDataType TPrecision>

virtual void Mila::Dnn::Network< TDeviceType, TPrecision >::load_	(	ModelArchive &	archive,
		SerializationMode	mode )

inlineoverrideprotectedvirtual

Hook for concrete classes to validate type-specific state on load.

Deliberately NOT pure, where save_() is. A concrete network must write its config, because nothing else can; it rarely needs to read it back here, because the config is consumed by the static factory before construction – by the time this runs the network already exists with that geometry. Override only to assert the archive matches what was built.

The empty body also keeps this from being a breaking change for every existing Network subclass.

Parameters

archive	Archive to read from.
mode	Serialization mode.

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

Reimplemented in Mila::Dnn::GptTransformer< TDeviceType, TPrecision >.

◆ save()

template<DeviceType TDeviceType, TensorDataType TPrecision>

void Mila::Dnn::Network< TDeviceType, TPrecision >::save	(	ModelArchive &	archive,
		SerializationMode	mode ) const

inline

Save network to archive.

Saves component graph structure and delegates to concrete class via save_() hook for type-specific configuration.

Archive structure produced:

network/meta.json: Base metadata (name, version, num_components, timestamp)
network/architecture.json: Component topology (names, paths, ordering)
components/<name>/...: Child component state (recursive)
Concrete class writes additional files via save_() override

Parameters

archive	Archive to write to
mode	Serialization mode (Checkpoint, WeightsOnly, Architecture)

Exceptions

std::runtime_error if save_() is not overridden by concrete class

◆ save_()

template<DeviceType TDeviceType, TensorDataType TPrecision>

virtual void Mila::Dnn::Network< TDeviceType, TPrecision >::save_	(	ModelArchive &	archive,
		SerializationMode	mode ) const

overrideprotectedpure virtual

Hook for concrete classes to save type-specific state.

REQUIRED override for concrete networks. Must write:

Type identifier (e.g., "type": "MnistClassifier")
Configuration parameters (batch_size, architecture constants)
Shape metadata (for validation during Load())

This metadata enables the concrete class's Load() method to reconstruct the network.

Example implementation:

void save_(ModelArchive& archive, SerializationMode mode) const override
{
    json meta;
    meta["type"] = "MnistClassifier";  // Type identifier for runtime dispatch
    meta["batch_size"] = batch_size_;
    meta["input_shape"] = leading_shape_;
    // ... other configuration
    archive.writeJson("network/classifier_meta.json", meta);
}

Parameters

archive	Archive to write to
mode	Serialization mode (passed from save())

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

Implemented in Mila::Dnn::GemmaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >, Mila::Dnn::GptTransformer< TDeviceType, TPrecision >, and Mila::Dnn::LlamaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >.

◆ synchronize()

template<DeviceType TDeviceType, TensorDataType TPrecision>

void Mila::Dnn::Network< TDeviceType, TPrecision >::synchronize ( )

inlineoverridevirtual

Synchronize all child components.

Waits for outstanding device operations on all children.

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

◆ toString()

template<DeviceType TDeviceType, TensorDataType TPrecision>

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

inlineoverridevirtual

Generate a human-readable description.

Returns: String representation showing network name and children

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

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

Mila/Src/Dnn/Core/Network.ixx

Public Types

Public Member Functions

Protected Member Functions

Additional Inherited Members

Detailed Description

Constructor & Destructor Documentation

◆ Network()

Member Function Documentation

◆ createOptimizer()

◆ getDeviceId()

◆ getExecutionContext()

◆ getType()

◆ load()

◆ load_()

◆ save()

◆ save_()

◆ synchronize()

◆ toString()