opennn::GroupedQueryAttentionOperator Struct Reference

#include <opennn/network/layers/grouped_query_attention_layer.h>

Inheritance diagram for opennn::GroupedQueryAttentionOperator:

Public Member Functions
	GroupedQueryAttentionOperator ()
	~GroupedQueryAttentionOperator () override
void	set (Index new_sequence_length, Index new_hidden, Index new_q_heads, Index new_kv_heads, Index new_head_dim, float new_rope_theta, float new_rms_epsilon, bool new_use_qk_norm)
vector< TensorSpec >	parameter_specs () const override
vector< TensorSpec >	forward_scratch_specs () const
vector< SlotQuantization >	parameter_quantization () const override
void	link_parameters (span< const TensorView >) override
void	link_parameter_scales (span< const TensorView >) override
void	set_parameters_random () override
const Buffer *	get_rope_tables () const noexcept
void	forward_propagate (ForwardPropagation &, size_t, ForwardPropagationMode) override
void	back_propagate (ForwardPropagation &, BackPropagation &, size_t) const override
Public Member Functions inherited from opennn::Operator
virtual	~Operator ()=default
virtual vector< TensorSpec >	state_specs () const
virtual vector< ParameterSlot >	parameter_slots ()
virtual void	link_gradients (span< const TensorView > views)
virtual void	link_states (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_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 Attributes
static constexpr size_t	forward_scratch_slots_count = 9

Friends
class	GroupedQueryAttention

Additional Inherited Members
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}

Constructor & Destructor Documentation

GroupedQueryAttentionOperator()

opennn::GroupedQueryAttentionOperator::GroupedQueryAttentionOperator

(

)

~GroupedQueryAttentionOperator()

opennn::GroupedQueryAttentionOperator::~GroupedQueryAttentionOperator ( )

override

Member Function Documentation

back_propagate()

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

overridevirtual

Reimplemented from opennn::Operator.

forward_propagate()

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

overridevirtual

Reimplemented from opennn::Operator.

forward_scratch_specs()

vector< TensorSpec > opennn::GroupedQueryAttentionOperator::forward_scratch_specs

(

)

const

get_rope_tables()

const Buffer * opennn::GroupedQueryAttentionOperator::get_rope_tables ( ) const

inlinenoexcept

link_parameter_scales()

void opennn::GroupedQueryAttentionOperator::link_parameter_scales ( span< const TensorView > )

overridevirtual

Reimplemented from opennn::Operator.

link_parameters()

void opennn::GroupedQueryAttentionOperator::link_parameters ( span< const TensorView > )

overridevirtual

Reimplemented from opennn::Operator.

parameter_quantization()

vector< SlotQuantization > opennn::GroupedQueryAttentionOperator::parameter_quantization ( ) const

overridevirtual

Reimplemented from opennn::Operator.

parameter_specs()

vector< TensorSpec > opennn::GroupedQueryAttentionOperator::parameter_specs ( ) const

overridevirtual

Reimplemented from opennn::Operator.

set()

void opennn::GroupedQueryAttentionOperator::set	(	Index	new_sequence_length,
		Index	new_hidden,
		Index	new_q_heads,
		Index	new_kv_heads,
		Index	new_head_dim,
		float	new_rope_theta,
		float	new_rms_epsilon,
		bool	new_use_qk_norm )

set_parameters_random()

void opennn::GroupedQueryAttentionOperator::set_parameters_random ( )

overridevirtual

Reimplemented from opennn::Operator.

Friends And Related Symbol Documentation

GroupedQueryAttention

friend class GroupedQueryAttention

friend

Member Data Documentation

forward_scratch_slots_count

size_t opennn::GroupedQueryAttentionOperator::forward_scratch_slots_count = 9

staticconstexpr