VynarisEarly beta Kimi K3Get your API key

Attention Mechanisms & Transformer Architecture

22 plain-English definitions.

Attention sink
The empirical phenomenon where initial tokens in a sequence absorb disproportionate attention mass, exploited by some long-context and streaming methods to preserve stability.
Causal attention mask
A mask applied during attention so each token can only attend to itself and earlier tokens, enforcing autoregressive generation.
Cross-attention
An attention mechanism where queries come from one sequence and keys/values from another, used to let a decoder attend to encoder outputs or other modalities.
Decoder-only architecture
A transformer design that generates tokens autoregressively using only a stack of decoder (causal-attention) blocks; the standard architecture for modern chat LLMs.
Encoder-decoder architecture
A two-part transformer design (e.g., T5) where an encoder processes the full input and a decoder generates output conditioned on it, common in translation and summarization.
Encoder-only architecture
A transformer design (e.g., BERT) that processes the whole input bidirectionally without generation, used for classification and embeddings.
Feed-forward network (FFN)
The position-wise two-layer (or gated) MLP block in a transformer layer that transforms each token's representation independently after attention.
Gated linear unit (GLU)
A feed-forward variant that multiplies one linear projection by a gated (activated) second projection, improving expressiveness over a plain MLP.
GeGLU
A gated linear unit variant using the GELU activation function in place of SwiGLU's Swish.
Layer normalization
A normalization technique that rescales activations across the feature dimension for each token, stabilizing training.
Linear attention
An attention approximation that reformulates the softmax kernel to achieve linear rather than quadratic time/memory complexity in sequence length.
Load balancing loss
An auxiliary training loss added to MoE models to discourage the router from sending too many tokens to a small subset of experts.
Multi-head attention
An attention layer that runs several attention "heads" in parallel on different learned projections of the input so the model can attend to different relationship types at once.
Pre-norm vs post-norm
Two placements of normalization layers relative to the residual connection; pre-norm (before the sublayer) is now standard for training stability at scale.
Residual connection
A skip connection that adds a layer's input directly to its output, easing gradient flow in deep networks.
RMSNorm
A simplified layer normalization that scales by the root-mean-square of activations without subtracting the mean, cheaper and widely used in modern LLMs.
Scaled dot-product attention
The core attention formula that divides the dot product of query and key vectors by the square root of the key dimension before softmax, to keep gradients stable.
Self-attention
A mechanism where each token in a sequence computes a weighted combination of all other tokens in the same sequence to build context-aware representations.
Shared expert
An MoE design where one or more experts are always active for every token alongside the sparsely routed experts, capturing common knowledge.
Sparse attention
Any attention variant that skips computing full pairwise attention scores, instead attending to a subset of tokens, to scale to longer contexts cheaply.
Sparse MoE routing
The gating mechanism that decides which top-k experts process each token in a mixture-of-experts layer.
SwiGLU
A gated linear unit variant using the Swish/SiLU activation function, widely used in modern transformer feed-forward blocks (e.g., LLaMA).