MultiLatentAttentionLayer<T>

Layers · AiDotNet.NeuralNetworks.Layers.SSM

Implements the Multi-Latent Attention (MLA) layer from DeepSeek-V2 (Aixin Liu et al., 2024).

Inheritance: LayerBase<T> → MultiLatentAttentionLayer<T>

Implements: IShapeContract

For Beginners

MLA is a memory-efficient attention mechanism used in DeepSeek-V2.

Standard attention is like keeping a complete notebook for every word you’ve read:

  • You write the full Key (what this word is about) and Value (what it says) for every word
  • When answering a question (Query), you look up all Key-Value pairs
  • For a 100K word document, that’s a LOT of notebook pages

MLA is like keeping compressed sticky notes instead:

  • For each word, you write just a short summary (the latent c_t)
  • When you need the full Key or Value, you expand the summary on the fly
  • You still get nearly the same answer quality, but use far less paper

The “Multi-Latent” name comes from having multiple attention heads, each with its own latent compression. This is what enables DeepSeek-V2 to handle very long contexts efficiently.

How It Works

Multi-Latent Attention compresses the KV cache via low-rank factorization, dramatically reducing memory usage during inference. Instead of caching full-dimensional K and V for every token, MLA caches a much smaller latent vector c_t, from which K and V are reconstructed on the fly.

The architecture:

The key insight is KV cache compression: during inference, you only need to store c_t (latentDim per token) instead of K and V (2 * modelDim per token). When latentDim = modelDim/4, this yields an 8x reduction in KV cache memory, which is the primary bottleneck for long-context LLM serving.

Reference: Aixin Liu et al., “DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model”, 2024. https://arxiv.org/abs/2405.04434

Constructors

Constructor Summary
MultiLatentAttentionLayer(int sequenceLength, int modelDimension, int numHeads, int latentDimension, IActivationFunction<T> activationFunction, IInitializationStrategy<T> initializationStrategy) Creates a new Multi-Latent Attention (MLA) layer.

Properties

Property Summary
bool SupportsTraining { get; }
int HeadDimension { get; } Gets the dimension per head.
int LatentDimension { get; } Gets the latent dimension for KV cache compression.
int ModelDimension { get; } Gets the model dimension.
int NumHeads { get; } Gets the number of attention heads.

Methods

Method Summary
Dictionary<string, string> GetParameterRoles() Returns parameter roles for per-role optimizer configuration (e.g., weight decay exemption for biases).
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank) Derived from this type’s [TensorLayout] declarations.
IReadOnlyList<Tensor<T>> GetTrainableParameters() Returns all trainable parameter tensors marked with [TrainableParameter].
Tensor<T> GetCompressionWeights() Gets the compression weights for external inspection.
Tensor<T> GetOutputProjectionWeights() Gets the output projection weights for external inspection.
Tensor<T> GetQueryWeights() Gets the query weights for external inspection.
void ClearGradients()
void ResetState()
void SetTrainableParameters(IReadOnlyList<Tensor<T>> parameters) Replaces trainable parameter tensors (e.g., with ParameterBuffer views).
void UpdateParameters(T learningRate)
void ZeroGrad() Clears all gradient fields discovered by convention ({paramName}Gradient).