LLaMA-style transformer (decoder-only) for autoregressive token prediction. More...

Inheritance diagram for Mila::Dnn::LlamaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >:

Public Types
using	ComponentPtr = typename NetworkBase::ComponentPtr
using	LinearType = Linear<TDeviceType, TPrecision, TWeightQuantization>
using	LmHeadLinearType = Linear<TDeviceType, TPrecision>
using	MR = typename DeviceTypeTraits<TDeviceType>::memory_resource
using	NetworkBase = LanguageModelNetwork<TDeviceType, TPrecision>
using	RmsNormType = RmsNorm<TDeviceType, TPrecision>
using	TensorType = Tensor<TPrecision, MR>
using	TokenEmbeddingType = TokenEmbedding<TDeviceType, dtype_t::INT32, TPrecision>
using	TokenIndexType = Tensor<dtype_t::INT32, MR>
using	TransformerBlockType = LlamaBlock<TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy>
Public Types inherited from Mila::Dnn::LanguageModelNetwork< TDeviceType, TPrecision >
using	MR = typename DeviceTypeTraits<TDeviceType>::memory_resource
using	NetworkBase = Network<TDeviceType, TPrecision>
using	TensorType = Tensor<TPrecision, MR>
using	TokenIndexType = Tensor<TensorDataType::INT32, MR>
Public Types inherited from Mila::Dnn::Network< TDeviceType, TPrecision >
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
	LlamaTransformer (const std::string &name, const LlamaConfig &config, DeviceId device_id)
TokenIndexType &	backward (const TokenIndexType &input, const TensorType &output_grad) override
TensorType &	decode (const TokenIndexType &input, dim_t position) override
TensorType &	forward (const TokenIndexType &input) override
MemoryStats	getMemoryStats () const override
	Return the current memory allocation breakdown for this component.
ModelType	getModelType () const
MemoryStats	getRequiredMemory (const BuildContext &context) const override
	What build( context ) would allocate for the whole model, without allocating.
void	loadParameters (WeightsReader &reader)
TensorType &	prefill (const TokenIndexType &input) override
PrefillChunking	prefillChunking (int64_t B, int64_t T_ctx) const
	The prefill chunk this context length would use, and the largest the context permits.
std::string	toString () const override
	Generate a human-readable description.
void	zeroGradients () override
	Clear all model-owned gradients for this component.
Public Member Functions inherited from Mila::Dnn::LanguageModelNetwork< TDeviceType, TPrecision >
	LanguageModelNetwork (const std::string &name)
virtual TokenIndexType &	backward (const TokenIndexType &input, const TensorType &output_grad)=0
	Full backward pass (training).
virtual TensorType &	decode (const TokenIndexType &input, dim_t position)=0
	Inference decode – single-token autoregressive step.
virtual TensorType &	forward (const TokenIndexType &input)=0
	Full-sequence forward pass – the training path, not the generation path.
virtual TensorType &	prefill (const TokenIndexType &input)=0
	Inference prefill – process full prompt and populate the KV cache.
virtual TensorType &	prefillFrom (const TokenIndexType &input, dim_t start_offset)
	Chunked prefill starting at an absolute position (prompt-prefix reuse).
virtual bool	rewindKvCache (dim_t position)
virtual SequenceLogLikelihood	scoreTokens (const TokenIndexType &input)
	Rewind the KV caches to position for prompt-prefix reuse (PromptCaching.md).
Public Member Functions inherited from Mila::Dnn::Network< TDeviceType, TPrecision >
	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.
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.
std::vector< std::string >	componentPaths () const
	Every component path in this subtree, in traversal order.
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.
size_t	observe (std::string_view pattern, Compute::ComputePassMask passes, Compute::ActivationObserver sink)
	Attach sink to every component in this subtree whose path matches.
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)
	Remove a child component by name.
void	saveFlatTensors (Serialization::SafeTensorsWriter &writer, const std::string &prefix, Serialization::TensorSavePass pass) const override
	Recurse into children, extending the flat dotted prefix.
void	stopObserving ()
	Detach every observer in this subtree and clear the shared sink.
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).
IExecutionContext *	getExecutionContext () const
	Get the shared execution context.
const std::string	getName () const
	Get the component's name identifier.
virtual std::vector< ObservableStage >	getObservableStages () const
	Stages this component publishes, and the passes each is published on.
Compute::ComputePassMask	getObservedPasses () const noexcept
virtual std::vector< const ITensor * >	getOutputs () const
	Return non-owning pointers to the tensors this component produces.
virtual std::vector< std::string >	getParameterNames () const
	List all available parameter names for this component.
RuntimeMode	getRuntimeMode () const noexcept
TrainingMode	getTrainingMode () const noexcept
	The current runtime behavioral mode of this Component.
virtual bool	isBuilt () const final
	Returns true if build() has completed successfully.
bool	isInferenceMode () const noexcept
bool	isTrainingMode () const noexcept
virtual void	loadParameter (const std::string &, const Serialization::ITensorBlob &)
	Load a parameter from serialized tensor data.
void	setObservedPasses (Compute::ComputePassMask passes) noexcept
	Select which passes this component publishes on.
void	setTrainingMode (TrainingMode mode)
	Set the runtime behavioral mode for this Component.

Protected Member Functions
void	onBuilding (const BuildContext &context) override
	Hook invoked by build() to allocate component buffers.
void	onTrainingModeChanging (TrainingMode training_mode) override
	Hook invoked when training mode is about to change.
std::size_t	ropeCacheBytes (dim_t head_dim, dim_t T_ctx, std::size_t granularity) const noexcept
	Bytes one RoPE cos/sin cache occupies for a given head width and context length.
void	save_ (ModelArchive &archive, SerializationMode) const override
	Hook for concrete classes to save type-specific state.
Protected Member Functions inherited from Mila::Dnn::Network< TDeviceType, TPrecision >
virtual void	load_ (ModelArchive &archive, SerializationMode mode) override
	Hook for concrete classes to validate type-specific state on load.
void	verifyArchitectureCompatibility (const WeightsMetadata &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.
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 >
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.
void	publish (Compute::ComputePass pass, std::string_view stage, const ITensor &value) const
	Publish one activation to the observer, if this pass is being observed.
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 = typename DeviceTypeTraits<TDeviceType>::host_staging_memory_resource
	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, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>
requires PrecisionSupportedOnDevice<TPrecision, TDeviceType>
class Mila::Dnn::LlamaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >

LLaMA-style transformer (decoder-only) for autoregressive token prediction.

Graph: TokenEmbedding -> RoPE -> LlamaBlock x N -> RmsNorm -> Linear (lm_head). RoPE is applied to the full embedding stream after the token lookup; each LlamaBlock receives rotary-encoded embeddings as input.

Template parameters:

TDeviceType: device type (Cpu/Cuda)
TPrecision: tensor precision

Member Function Documentation

◆ getMemoryStats()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

MemoryStats Mila::Dnn::LlamaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::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 >.

◆ getRequiredMemory()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

MemoryStats Mila::Dnn::LlamaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::getRequiredMemory ( const BuildContext & context ) const

inlineoverridevirtual

What build( context ) would allocate for the whole model, without allocating.

Mirrors onBuilding(): resolve the prefill chunk first, recurse with the same per-child contexts, then add the shared GQA workspace this transformer owns.

Two corrections Gemma needs are absent here, and their absence is the finding rather than an omission. Llama does not pool per-block activations, so there is no installed-output adjustment; and it does not tie the embedding to the head, so the two largest tensors are counted separately and in full. See BACKLOG, Models.

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

◆ onBuilding()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

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

inlineoverrideprotectedvirtual

Hook invoked by build() to allocate component buffers.

Receives the stored BuildContext. Implementations must use config.allocationSeqLen() when sizing output buffers – this is the single call that makes Inference and Training allocate the correct buffer sizes automatically without per-component logic.

// Example -- Linear component:
shape_t out_shape =
{
    config.batchSize(),
    config.allocationSeqLen(),   // 1 for Inference, T for Training
    config_.getOutputFeatures()
};
output_ = std::make_unique<TensorType>( device, out_shape,
    this->getName() + ".output" );

The default implementation forwards to the legacy onBuilding( const shape_t& ) overload for backwards compatibility. New components should override this overload directly.

Note: Do not call build() or onBuilding() from within this hook.; Implementations should either succeed fully or leave no partial state, as a failed build() may be retried.

Takes the build-time configuration; use its allocationSeqLen() to obtain the correct output buffer sequence dimension.

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

◆ onTrainingModeChanging()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

void Mila::Dnn::LlamaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::onTrainingModeChanging ( TrainingMode training_mode )

inlineoverrideprotectedvirtual

Hook invoked when training mode is about to change.

Propagates the new mode to all child components. The hook runs with the Component's training mutex held; it MUST NOT call setTrainingMode().

Parameters

training_mode New training mode (Normal or Eval)

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

◆ prefillChunking()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

PrefillChunking Mila::Dnn::LlamaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::prefillChunking	(	int64_t	B,
		int64_t	T_ctx ) const

inline

The prefill chunk this context length would use, and the largest the context permits.

The largest rung whose whole predicted footprint fits the device's free memory, read now. Allocates nothing, so a caller may ask before the network is built; two calls made at different moments can disagree. See Specifications/MemoryFootprint.md section 11.

◆ ropeCacheBytes()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

std::size_t Mila::Dnn::LlamaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::ropeCacheBytes	(	dim_t	head_dim,
		dim_t	T_ctx,
		std::size_t	granularity ) const

inlineprotectednoexcept

Bytes one RoPE cos/sin cache occupies for a given head width and context length.

MUST match CudaRopeOp::getRequiredStateMemorySize – FP32 regardless of the model precision, one row per context position, half the head dimension, two caches. Duplicated here because the deduplication is the transformer's to apply and it needs the per-key size; the model-level comparison against getMemoryStats is what holds the two together.

◆ save_()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

void Mila::Dnn::LlamaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::save_	(	ModelArchive &	archive,
		SerializationMode	mode ) const

inlineoverrideprotectedvirtual

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())

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

◆ toString()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

std::string Mila::Dnn::LlamaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::toString ( ) const

inlineoverridevirtual

Generate a human-readable description.

Returns: String representation showing children

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

◆ zeroGradients()

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuantization = NoWeightQuant, KvCachePolicy TKvCachePolicy = NoKvCompression>

void Mila::Dnn::LlamaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >::zeroGradients ( )

inlineoverridevirtual

Clear all model-owned gradients for this component.

Default implementation is a no-op. Composite components should override to recurse to children. Leaf components should override to zero their parameter and activation gradients using device-aware helpers.

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

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

Mila/Src/Dnn/Components/Transformers/LlaMa/Llama.ixx

Public Types

Public Member Functions

Protected Member Functions

Additional Inherited Members

Detailed Description

Member Function Documentation

◆ getMemoryStats()

◆ getRequiredMemory()

◆ onBuilding()

◆ onTrainingModeChanging()

◆ prefillChunking()

◆ ropeCacheBytes()

◆ save_()

◆ toString()

◆ zeroGradients()