Dnn.Components.GptBlock Module Reference

Exported Modules
module	Compute.Device
module	Compute.DeviceId
module	Dnn.Tensor
module	Compute.DeviceTypeTraits
module	Dnn.TensorHelpers
module	Dnn.ComponentType
module	Dnn.TensorInitializers
module	Serialization.Mode
module	Compute.ExecutionContext
module	Compute.DeviceType
module	Dnn.ITensor
module	Logging.Logger
module	Compute.IExecutionContext
module	Dnn.CompositeComponent
module	Dnn.Components.MultiHeadAttention
module	Serialization.ModelArchive
module	Dnn.Components.Linear
module	Dnn.TensorDataTypeTraits
module	Dnn.TensorDataType
module	Compute.MemoryResource
module	Dnn.Components.LayerNorm
module	Dnn.Components.MLP
module	Compute.CpuMemoryResource
module	Dnn.Component
module	Dnn.TensorOps
module	Compute.ExecutionContextFactory
module	Dnn.Components.Residual
module	Dnn.ActivationType
module	Dnn.TensorTypes

Classes
class	Mila::Dnn::GptBlock< TDeviceType, TPrecision >
	Transformer encoder block as a composite component. More...
class	Mila::Dnn::GptBlockConfig
	Configuration class for GPT transformer blocks. More...

Typedefs
using	AttentionType = MultiHeadAttention<TDeviceType, TPrecision>
using	ComponentPtr = typename CompositeComponentBase::ComponentPtr
using	CompositeComponentBase = CompositeComponent<TDeviceType, TPrecision>
using	ExecutionContextType = ExecutionContext<TDeviceType>
using	LayerNormType = LayerNorm<TDeviceType, TPrecision>
using	LinearType = Linear<TDeviceType, TPrecision>
using	MLPType = MLP<TDeviceType, TPrecision>
using	MR = typename DeviceTypeTraits<TDeviceType>::memory_resource
using	ResidualType = Residual<TDeviceType, TPrecision>
using	TensorType = Tensor<TPrecision, MR>

Functions
	GptBlock (const std::string &name, const GptBlockConfig &config, std::optional< DeviceId > device_id=std::nullopt)
	Construct GptBlock in shared or standalone mode.
	~GptBlock () override=default
TensorType &	backward (const TensorType &input, const TensorType &output_grad)
	Backward pass returning the input-gradient tensor.
void	createGraph ()
TensorType &	decode (const TensorType &input, int position)
	Inference-only single-token decode pass.
TensorType &	forward (const TensorType &input)
	Forward pass with optional KV cache dispatch.
MemoryStats	getMemoryStats () const override
	Return the current memory allocation breakdown for this component.
const ComponentType	getType () const override
	Get the component type identifier.
void	initializeKVCache (int64_t max_seq_len)
	Allocate KV cache buffers on the contained Attention component.
void	load_ (ModelArchive &archive, SerializationMode mode)
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.
void	resetKVCache ()
	Reset KV cache state on the contained Attention component.
void	save_ (ModelArchive &archive, SerializationMode mode) const override
	Save all child components recursively.
bool	supportsKVCache () const noexcept
	Returns true when the contained Attention supports KV caching.
std::string	toString () const override
	Generate a human-readable description.
void	validateBuildContext (const BuildContext &context) const
void	validateInputShape (const shape_t &input_shape) const
void	zeroGradients () override
	Clear all model-owned gradients for this component.

Variables
std::shared_ptr< AttentionType >	attn_ { nullptr }
shape_t	cached_input_shape_
GptBlockConfig	config_
std::shared_ptr< TensorType >	d_input_ { nullptr }
std::shared_ptr< TensorType >	d_res1_accum_ { nullptr }
std::shared_ptr< MLPType >	ffn_ { nullptr }
bool	forward_executed_ { false }
TensorType *	last_attn_out_ { nullptr }
TensorType *	last_ffn_out_ { nullptr }
TensorType *	last_ln1_out_ { nullptr }
TensorType *	last_ln2_out_ { nullptr }
TensorType *	last_out_proj_out_ { nullptr }
TensorType *	last_qkv_out_ { nullptr }
TensorType *	last_res1_out_ { nullptr }
TensorType *	last_res2_out_ { nullptr }
std::shared_ptr< LayerNormType >	ln1_ { nullptr }
std::shared_ptr< LayerNormType >	ln2_ { nullptr }
std::shared_ptr< LinearType >	out_proj_ { nullptr }
std::unique_ptr< IExecutionContext >	owned_exec_context_ { nullptr }
std::shared_ptr< LinearType >	qkv_proj_ { nullptr }
std::shared_ptr< ResidualType >	res1_ { nullptr }
std::shared_ptr< ResidualType >	res2_ { nullptr }

Files
file	/__w/Mila/Mila/Mila/Src/Dnn/Components/Transformers/Gpt/GptBlock.ixx
	Transformer encoder block implementation.
file	/__w/Mila/Mila/Mila/Src/Dnn/Components/Transformers/Gpt/GptBlock.Config.ixx
	Configuration for GPT-style transformer block (block-level).