MagnetoDecoderLayer<T>

Layers · AiDotNet.NeuralNetworks.Layers

KOSMOS’s MAGNETO causal language model (reference Kosmos2TextTransformer + tied lm_head). It runs these steps:

  • Token embeddings are looked up, image embeddings are written into their slots, and everything is

scaled by sqrt(d).

  • Positions: fairseq sinusoidal positions starting at 2, or xPos inside attention.
  • Pre-LN blocks, where attention carries a Sub-LN on its output before out_proj, and the FFN

is fc1, GELU, a Sub-LN, then fc2.

  • A final LayerNorm, then logits from the tied embedding matrix.

Inheritance: LayerBase<T> → MagnetoDecoderLayer<T>

Implements: IShapeContract

For Beginners

This is the language model that writes text one token at a time, reading the image’s embeddings as if they were words at the start of the sentence.

How It Works

The single-input forward takes token ids [S] and returns next-token logits [S, vocab].

Constructors

Constructor Summary
MagnetoDecoderLayer(int vocabSize, int dim, int numLayers, int numHeads, int ffnDim, int maxPositions, KosmosPositionEncoding positions) Creates the decoder.

Properties

Property Summary
IReadOnlyList<LayerPort> InputPorts { get; } Generated from [TensorPort] declarations.
IReadOnlyList<LayerPort> OutputPorts { get; } Generated from [TensorPort] declarations.
bool SupportsTraining { get; }
int Dim { get; } Model width.
int VocabSize { get; } Vocabulary size (logit width).

Methods

Method Summary
Dictionary<string, string> GetParameterRoles() Returns parameter roles for per-role optimizer configuration (e.g., weight decay exemption for biases).
IReadOnlyList<ILayer<T>> GetSubLayers()
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank)
IReadOnlyList<Tensor<T>> GetTrainableParameters() Returns all trainable parameter tensors marked with [TrainableParameter].
LayerInputDomain GetInputDomain(int[] inputShape) Generated from the primary [TensorPort] declaration.
void ResetState()
void SetTrainableParameters(IReadOnlyList<Tensor<T>> parameters) Replaces trainable parameter tensors (e.g., with ParameterBuffer views).
void ZeroGrad() Clears all gradient fields discovered by convention ({paramName}Gradient).