#include <opennn/network/operators/attention_operator.h>

Inheritance diagram for opennn::AttentionOperator:

Classes
struct	SdpaBf16Pack

Public Member Functions
void	set (Index, Index, Index, Index, CausalMask, Type)

vector< TensorSpec >	forward_scratch_specs (Index) const

TensorSpec	backward_scratch_spec (Index) const

vector< TensorSpec >	sdpa_gradient_scratch_specs (Index) const

TensorSpec	sdpa_qkv_pack_spec (Index) const

vector< TensorSpec >	sdpa_state_specs (Index) const

void	forward_propagate (ForwardPropagation &, size_t, ForwardPropagationMode) override

void	back_propagate (ForwardPropagation &, BackPropagation &, size_t) const override

void	concatenate_output_heads (ForwardPropagation &, size_t) const

void	split_output_delta (ForwardPropagation &, BackPropagation &, size_t) const

	AttentionOperator ()

	~AttentionOperator () override

	AttentionOperator (AttentionOperator &&) noexcept

AttentionOperator &	operator= (AttentionOperator &&) noexcept

	AttentionOperator (const AttentionOperator &)=delete

AttentionOperator &	operator= (const AttentionOperator &)=delete

Public Member Functions inherited from opennn::Operator
virtual	~Operator ()=default

virtual vector< TensorSpec >	parameter_specs () const

virtual vector< TensorSpec >	state_specs () const

virtual vector< SlotQuantization >	parameter_quantization () const

virtual vector< ParameterSlot >	parameter_slots ()

virtual void	link_parameters (span< const TensorView > views)

virtual void	link_gradients (span< const TensorView > views)

virtual void	link_states (span< const TensorView >)

virtual void	link_parameter_scales (span< const TensorView >)

bool	link_slots (span< const TensorView > views, TensorView ParameterSlot::member)

virtual void	initialize_states ()

virtual void	set_weights_dtype (Type new_weights_dtype)

virtual void	set_parameters_random ()

virtual void	set_parameters_glorot ()

virtual void	set_parameters_pytorch ()

virtual void	to_JSON (JsonWriter &) const

virtual void	from_JSON (const Json *)

virtual void	load_state_from_JSON (const Json *)

TensorView &	get_input (ForwardPropagation &forward_propagation, size_t layer, size_t slot_index=0) const noexcept

vector< TensorView > &	get_inputs (ForwardPropagation &forward_propagation, size_t layer) const noexcept

TensorView &	get_output (ForwardPropagation &forward_propagation, size_t layer, size_t slot_index=0) const noexcept

TensorView &	get_output_delta (BackPropagation &back_propagation, size_t layer, size_t slot_index=0) const noexcept

TensorView &	get_input_delta (BackPropagation &back_propagation, size_t layer, size_t slot_index=0) const noexcept

Static Public Member Functions
static bool	sdpa_supported (Type, Device)

Public Attributes
Index	heads_number = 0

Index	head_dimension = 0

Index	query_sequence_length = 0

Index	source_sequence_length = 0

bool	use_causal_mask = false

bool	use_sdpa = false

bool	interleaved_heads = false

bool	zero_padded_queries = false

MatrixR	causal_mask

DropoutOperator	dropout

size_t	scratch_slot = 0

size_t	attention_output_slot = 0

size_t	concatenated_output_delta_slot = 0

size_t	sdpa_gradient_slot = 0

size_t	sdpa_qkv_pack_slot = 0

size_t	sdpa_state_slot = 0

size_t	dropout_mask_slot = 0

Public Attributes inherited from opennn::Operator
Type	compute_dtype = Type::FP32

Type	weights_dtype = Type::FP32

vector< size_t >	input_slots = {0}

vector< size_t >	output_slots = {1}

vector< size_t >	input_delta_slots = {1}

vector< size_t >	output_delta_slots = {0}

Static Public Attributes
static constexpr size_t	sdpa_scratch_slots_count = 8

static constexpr size_t	sdpa_state_slots_count = 4

Constructor & Destructor Documentation

◆ AttentionOperator() [1/3]

opennn::AttentionOperator::AttentionOperator ( )

◆ ~AttentionOperator()

opennn::AttentionOperator::~AttentionOperator ( )

override

◆ AttentionOperator() [2/3]

opennn::AttentionOperator::AttentionOperator ( AttentionOperator && )

noexcept

◆ AttentionOperator() [3/3]

opennn::AttentionOperator::AttentionOperator ( const AttentionOperator & )

delete

Member Function Documentation

◆ back_propagate()

void opennn::AttentionOperator::back_propagate	(	ForwardPropagation &	,
		BackPropagation &	,
		size_t	) const

overridevirtual

Reimplemented from opennn::Operator.

◆ backward_scratch_spec()

TensorSpec opennn::AttentionOperator::backward_scratch_spec ( Index ) const

◆ concatenate_output_heads()

void opennn::AttentionOperator::concatenate_output_heads	(	ForwardPropagation &	,
		size_t	) const

◆ forward_propagate()

void opennn::AttentionOperator::forward_propagate	(	ForwardPropagation &	,
		size_t	,
		ForwardPropagationMode	)

overridevirtual

Reimplemented from opennn::Operator.

◆ forward_scratch_specs()

vector< TensorSpec > opennn::AttentionOperator::forward_scratch_specs ( Index ) const

◆ operator=() [1/2]

AttentionOperator & opennn::AttentionOperator::operator= ( AttentionOperator && )

noexcept

◆ operator=() [2/2]

AttentionOperator & opennn::AttentionOperator::operator= ( const AttentionOperator & )

delete

◆ sdpa_gradient_scratch_specs()

vector< TensorSpec > opennn::AttentionOperator::sdpa_gradient_scratch_specs ( Index ) const

◆ sdpa_qkv_pack_spec()

TensorSpec opennn::AttentionOperator::sdpa_qkv_pack_spec ( Index ) const

◆ sdpa_state_specs()

vector< TensorSpec > opennn::AttentionOperator::sdpa_state_specs ( Index ) const

◆ sdpa_supported()

static bool opennn::AttentionOperator::sdpa_supported	(	Type	,
		Device	)

static

◆ set()

void opennn::AttentionOperator::set	(	Index	,
		Index	,
		Index	,
		Index	,
		CausalMask	,
		Type	)

◆ split_output_delta()

void opennn::AttentionOperator::split_output_delta	(	ForwardPropagation &	,
		BackPropagation &	,
		size_t	) const

Member Data Documentation

◆ attention_output_slot

size_t opennn::AttentionOperator::attention_output_slot = 0

◆ causal_mask

MatrixR opennn::AttentionOperator::causal_mask

◆ concatenated_output_delta_slot

size_t opennn::AttentionOperator::concatenated_output_delta_slot = 0

◆ dropout

DropoutOperator opennn::AttentionOperator::dropout

◆ dropout_mask_slot

size_t opennn::AttentionOperator::dropout_mask_slot = 0

◆ head_dimension

Index opennn::AttentionOperator::head_dimension = 0

◆ heads_number

Index opennn::AttentionOperator::heads_number = 0

◆ interleaved_heads

bool opennn::AttentionOperator::interleaved_heads = false

◆ query_sequence_length

Index opennn::AttentionOperator::query_sequence_length = 0

◆ scratch_slot

size_t opennn::AttentionOperator::scratch_slot = 0

◆ sdpa_gradient_slot

size_t opennn::AttentionOperator::sdpa_gradient_slot = 0

◆ sdpa_qkv_pack_slot

size_t opennn::AttentionOperator::sdpa_qkv_pack_slot = 0

◆ sdpa_scratch_slots_count

size_t opennn::AttentionOperator::sdpa_scratch_slots_count = 8

staticconstexpr

◆ sdpa_state_slot

size_t opennn::AttentionOperator::sdpa_state_slot = 0

◆ sdpa_state_slots_count

size_t opennn::AttentionOperator::sdpa_state_slots_count = 4