Struct MultiGateResidual 

pub struct MultiGateResidual {
    pub w_beta: Param<Tensor<1>>,
    pub w_alpha: Param<Tensor<1>>,
    pub b_beta: Param<Tensor<1>>,
    pub d_model: usize,
    pub n_stream: usize,
}

One layer’s Multi-Gate Residual parameters: the mixer query w⁽ᵝ⁾ + bias b⁽ᵝ⁾, and the aggregator (AttnPool) query w⁽ᵅ⁾.

Fields

w_beta: Param<Tensor<1>>

Mixer query w⁽ᵝ⁾ ∈ ℝ^d (the per-stream sigmoid gate), [d_model].

w_alpha: Param<Tensor<1>>

Aggregator query w⁽ᵅ⁾ ∈ ℝ^d (the AttnPool softmax), [d_model].

b_beta: Param<Tensor<1>>

Per-stream mixer gate bias b⁽ᵝ⁾, [n_stream].

d_model: usize

Model width d.

n_stream: usize

Number of parallel residual streams n.

Implementations

impl MultiGateResidual

fn scale(&self) -> f32

fn rms_denom<const D: usize>(&self, x: Tensor<D>) -> Tensor<D>

The parameter-free RMS denominator d(x) ∈ [‥, 1] such that the RMSNorm (matching RmsNorm math with γ ≡ 1) is x / d(x). Returning the denominator rather than the normalised tensor lets Self::normed_score fold it out of the (feature-axis) score reduction, so the full-width normalised tensor is never built. The fp16 path keeps the same overflow-safe max-rescale, folded into the same scalar denominator.

fn normed_score<const R: usize>(&self, x: Tensor<R>, w: Tensor<R>) -> Tensor<R>

The RMSNorm-then-dot score scale · Σ_feat(x · w) / (rms(x)+eps), shape [‥, 1]. The RMS denominator is constant over the feature axis, so it is folded out of the reduction (via Self::rms_denom) — equal to Σ_feat(rms_norm(x) · w) · scale but without materialising the full-width normalised tensor.

fn mix_pool<const R: usize>( &self, layer_output: Tensor<R>, streams: Tensor<R>, ) -> (Tensor<R>, Tensor<R>)

The shared mix + pool, generic over the streams rank R (the stream axis is R-2, the feature axis R-1). Self::forward (R = 4) and Self::step (R = 3) only differ by that rank, so both lift their layer_output to a singleton stream axis, call this, and drop it again. All reductions keep their axis (size 1) for broadcasting, so scores/gates are […, n_stream, 1] throughout.

• layer_output: F_l lifted to a unit stream axis, […, 1, d_model]
• streams: the n_stream residual streams, […, n_stream, d_model]

Returns (h, streams') with h still carrying its unit stream axis ([…, 1, d_model]) and streams' the same shape as streams.

pub fn forward( &self, layer_output: Tensor<3>, streams: Tensor<4>, ) -> (Tensor<3>, Tensor<4>)

Full-sequence mix + pool.

• layer_output: this layer’s transform F_l, [batch, sequence, d_model]
• streams: the n_stream residual streams, [batch, sequence, n_stream, d_model]

Returns (h, streams'): the pooled input h for the next layer ([batch, sequence, d_model]) and the updated streams (same shape as in).

pub fn step( &self, layer_output: Tensor<2>, streams: Tensor<3>, ) -> (Tensor<2>, Tensor<3>)

Single-token mix + pool (the Self::forward math with the sequence axis dropped).

• layer_output: [batch, d_model]
• streams: [batch, n_stream, d_model]

Returns (h, streams'): [batch, d_model] and [batch, n_stream, d_model].

Trait Implementations

impl AutodiffModule for MultiGateResidual

fn valid(&self) -> Self

Returns the same module, but on the inner backend without auto-differentiation.

fn from_inner(module: Self) -> Self

Wraps an inner module back into an auto-diff module.

impl Clone for MultiGateResidual

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for MultiGateResidual

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

Formats the value using the given formatter.

impl Display for MultiGateResidual

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

Formats the value using the given formatter.

impl Module for MultiGateResidual

type Record = MultiGateResidualRecord

Type to save and load the module.

fn load_record(self, record: Self::Record) -> Self

Load the module state from a record.

fn into_record(self) -> Self::Record

Convert the module into a record containing the state.

fn num_params(&self) -> usize

Get the number of parameters the module has, including all of its sub-modules.

fn visit<Visitor: ModuleVisitor>(&self, visitor: &mut Visitor)

Visit each tensor parameter in the module with a visitor.

fn map<Mapper: ModuleMapper>(self, mapper: &mut Mapper) -> Self

Map each tensor parameter in the module with a mapper.

fn collect_devices(&self, devices: Devices) -> Devices

Return all the devices found in the underneath module tree added to the given vector without duplicates.

fn to_device(self, device: &Device) -> Self

Move the module and all of its sub-modules to the given device.

fn fork(self, device: &Device) -> Self

Fork the module and all of its sub-modules to the given device.

fn devices(&self) -> Vec<Device>

Return all the devices found in the underneath module tree without duplicates.

fn no_grad(self) -> Self

Each tensor in the module tree will not require grad.

fn train(self) -> Self

where Self: AutodiffModule,

Move the module and all of its sub-modules to the autodiff backend.

fn quantize_weights(self, quantizer: &mut Quantizer) -> Self

Quantize the weights of the module.

impl ModuleDisplay for MultiGateResidual

fn format(&self, passed_settings: DisplaySettings) -> String

Formats the module with provided display settings.

fn custom_settings(&self) -> Option<DisplaySettings>

Custom display settings for the module.

fn custom_content(&self, _content: Content) -> Option<Content>

Custom attributes for the module.

impl ModuleDisplayDefault for MultiGateResidual

fn content(&self, content: Content) -> Option<Content>

Attributes of the module used for display purposes.

fn num_params(&self) -> usize

Gets the number of the parameters of the module.