A progressive map
Moves from foundations through core, extended, and advanced topics for sequential or reference reading.
transformer.ngjoo.com · ONLINE
Transformer Atlas
A four-part, 33-chapter path through the Transformer stack, from attention fundamentals to KV Cache, MoE, LoRA, quantization, and inference optimization.
Content structure
COREMechanisms unpacked
Explains Q/K/V, masking, multi-head attention, position encoding, and sampling through data flow and implementation.
Traceable sources
Connects key ideas to papers, historical implementations, or teaching code with clear source boundaries.
Who it is for
Useful for sequential study of Transformers and for topic lookup across attention, position encoding, KV Cache, MoE, LoRA, and quantization.
Read the boundary first
It is a learning atlas, not a model-training service. Explanations should be read alongside the linked papers and implementations.
Ways to use it
- Start with attention and position encoding when learning the stack in sequence.
- Jump to KV Cache, quantization, or sampling when restoring context for an inference issue.
- Follow sources back to papers and teaching code before implementing or presenting a mechanism.
How to review the result
- Separate conceptual explanation, engineering experience, and paper findings because they carry different evidence.
- Teaching examples explain a mechanism and do not stand in for a production framework's current implementation.
- Re-measure speed, memory, and quality on the actual model, hardware, and software version.
Current status and evidence
2026-09-23The public learning endpoint was reachable without sign-in on 2026-09-23.
The available evidence is first-party: deployment records, product documentation, and endpoint status at the review time. No independent evaluation was available. Reachability confirms access at that moment, not performance, accuracy, savings, or long-term availability.