《Born》Appendix C: Glossary of Key Terms

Forum topic · QianXun · 2026-06-13

Summary

Appendix C of the technical book 《Born》 (serialized on zhichai.net) provides standardized definitions of the book's core terminology, organized into five categories. The Tensor and Data Structures section covers tensors (`Tensor[T, B]`), shape, stride, and broadcasting. The Compute Backend section defines backends (CPU/WebGPU), lazy evaluation for batching operations to hide communication latency, WGSL (WebGPU Shading Language), and workgroups as the basic unit of WebGPU compute dispatch. The Automatic Differentiation section explains autodiff, gradient tapes for recording forward operations, and backpropagation. The Neural Network section defines modules, multi-head attention (MHA), grouped-query attention (GQA), and residual connections. Finally, the Models and Deployment section covers checkpoints, GGUF (Georgi Gerganov Universal Format), KV-cache, and quantization. This glossary serves as a reference companion to the full book.

Appendix C of *Born* provides standard definitions for the core terms used throughout the book.

Tensors and Data Structures

| Term | Definition | |------|------------| | Tensor | A general term for multi-dimensional arrays; in Born, represented as Tensor[T, B] | | Shape | The list of lengths of a tensor along each axis | | Stride | The memory offset required to move one element along each axis | | Broadcasting | Rules for automatically expanding a smaller shape to match a larger one |

Compute Backends

| Term | Definition | |------|------------| | Backend | The concrete implementation that executes tensor operations (CPU/WebGPU) | | Lazy evaluation | Not executing operations immediately; batching submissions to hide communication latency | | WGSL | WebGPU Shading Language | | Workgroup | The basic unit of WebGPU compute dispatch |

Automatic Differentiation

| Term | Definition | |------|------------| | Automatic differentiation | An algorithm that automatically computes function gradients | | Gradient tape | A data structure that records the sequence of forward operations | | Backpropagation | Propagating gradients backward through the computation graph from the output loss |

Neural Networks

| Term | Definition | |------|------------| | Module | The basic building block of a neural network | | Multi-head attention (MHA) | Multiple attention heads computed in parallel | | Grouped-query attention (GQA) | Multiple query heads sharing key-value heads | | Residual connection | A shortcut connection that adds the input directly to the output |

Models and Deployment

| Term | Definition | |------|------------| | Checkpoint | A snapshot of model state saved during training | | GGUF | Georgi Gerganov Universal Format, a model format for LLMs | | KV-Cache | Caching key-value tensors to avoid redundant computation | | Quantization | Converting parameters from high precision to low precision |

---

📘 *Born* is a serialized technical book. This is Appendix C of 4 appendices.