Tags
Controlled tags for browsing topics across modules, concepts, models, and glossary pages. Each tag links to a landing page listing every published resource with that tag.
Architecture
- Foundationsfoundations
Foundational glossary and taxonomy pages that later model-family and module reference pages build on.
Architecture
- Taxonomytaxonomy
Cross-family glossary terms that explain how models, architectures, modules, components, modalities, and model roles fit together in Model Atlas.
Architecture
- Token-to-probability chaintoken-to-probability-chain
Beginner glossary path from tokens through embeddings, tensors, logits, softmax, and training fundamentals.
Architecture
Module type
- Activationactivation
Activation functions and activation-focused pages across glossary and module docs.
Module type
- Attentionattention
Self-attention and attention mechanisms used in transformers, modules, and related reference pages.
Module type
- Feed-forwardfeed-forward
Feed-forward blocks, FFN variants, and nearby module or glossary pages.
Module type
- Normalizationnormalization
Normalization layers, normalization variants, and related glossary pages.
Module type
- Position Encodingposition-encoding
Position-encoding methods, positional embedding variants, and long-context encoding pages.
Module type
- State Spacestate-space
State-space and selective state-space sequence modules that carry compact recurrent state across tokens.
Module type
- Tokenizationtokenization
Tokenizer algorithms, token boundaries, and related reference pages about how models split text into tokens.
Module type
Training
Inference
- Context Windowcontext-window
How far a model can attend within a sequence—local windows, sparse reach, or full dense attention over prior tokens.
Inference
- Inferenceinference
Runtime concepts, systems, and hardware guidance for serving or running models.
Inference
- KV Cachekv-cache
Key-value caching for faster autoregressive inference and lower memory bandwidth during decoding.
Inference
- Local modelslocal-models
Guides and references for running language models on local laptops, desktops, and workstations.
Inference
- Quantizationquantization
Lower-precision storage and runtime strategies that trade memory, bandwidth, and latency against possible quality loss.
Inference