Struct MultiHeadAttention

pub struct MultiHeadAttention {
    pub num_heads: i32,
    pub query_proj: MaybeQuantized<Linear>,
    pub key_proj: MaybeQuantized<Linear>,
    pub value_proj: MaybeQuantized<Linear>,
    pub output_proj: MaybeQuantized<Linear>,
}

Implements the scaled dot product attention with multiple heads.

Fields

num_heads: i32

Number of attention heads

query_proj: MaybeQuantized<Linear>

Query projection layer

key_proj: MaybeQuantized<Linear>

Key projection layer

value_proj: MaybeQuantized<Linear>

Value projection layer

output_proj: MaybeQuantized<Linear>

Output projection layer

Implementations

impl MultiHeadAttention

pub fn new( dims: impl Into<i32>, num_heads: impl Into<i32>, ) -> Result<Self, <<Self as Buildable>::Builder as Builder<Self>>::Error>

Creates a new instance of MultiHeadAttention.

impl MultiHeadAttention

pub const DEFAULT_BIAS: bool = false

Default value for the bias field

pub fn create_additive_causal_mask<T>(n: i32) -> Result<Array, Exception>

where T: ArrayElement + LowerBounded, Array: FromScalar<T>,

Creates an attention mask for use with MultiHeadAttention.

Trait Implementations

impl Buildable for MultiHeadAttention

type Builder = MultiHeadAttentionBuilder

The builder type for this buildable type

impl Builder<MultiHeadAttention> for MultiHeadAttentionBuilder

type Error = MultiHeadAttentionBuildError

Error with building

fn build(self) -> Result<MultiHeadAttention, Self::Error>

Build the type

impl Clone for MultiHeadAttention

fn clone(&self) -> MultiHeadAttention

Returns a copy of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for MultiHeadAttention

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<'a, Input> Module<Input> for MultiHeadAttention

where Input: Into<MultiHeadAttentionInput<'a>>,

type Error = Exception

Error type for the module.

type Output = Array

Output type of the module.

fn forward(&mut self, input: Input) -> Result<Self::Output, Self::Error>

Forward pass of the module.

fn training_mode(&mut self, mode: bool)

Set whether the module is in training mode.

impl ModuleParameters for MultiHeadAttention

fn freeze_parameters(&mut self, recursive: bool)

Freeze all parameters in the module.

fn unfreeze_parameters(&mut self, recursive: bool)

Unfreeze all parameters in the module.

fn parameters(&self) -> ModuleParamRef<'_>

Get references to the module parameters.

fn parameters_mut(&mut self) -> ModuleParamMut<'_>

Get mutable references to the module parameters.

fn trainable_parameters(&self) -> ModuleParamRef<'_>

Get references to the trainable parameters. A parameter is trainable if it is NOT frozen.

fn all_frozen(&self) -> Option<bool>

Check if all parameters in the module are frozen. Returns None if there are no parameters.

fn any_frozen(&self) -> Option<bool>

Check if any parameter in the module is frozen. Returns None if there are no parameters.

fn update(&mut self, parameters: ModuleParam)

Update the module parameters.

fn update_flattened(&mut self, flattened_parameters: FlattenedModuleParam)

Update the module parameters from a flattened representation.

impl Quantizable for MultiHeadAttention

type Quantized = MultiHeadAttention

The quantized type.

type QuantizationError = Exception

The error type for quantization.

fn try_into_quantized( self, group_size: i32, bits: i32, ) -> Result<Self::Quantized, Self::QuantizationError>

Quantize the module with the specified group size and number of bits.

const DEFAULT_GROUP_SIZE: i32 = 64i32

The default group size for quantization.

const DEFAULT_BITS: i32 = 4i32

The default number of bits for quantization.