Skip to content
mlmentorship

Book II · Chapter 3

Numerics and training stability

Understand normalization, precision, floating-point behavior, and loss failures.

Priority
Core
Difficulty
Intermediate
Useful for
MLE, RE, RS
Interview rounds
Training, Debugging

Read first: Optimization and schedules

Chapter contents

4 entries · read in order
  1. 01
    BatchNorm vs LayerNorm (and the transformer wrinkle)
    Concept
  2. 02
    Floating-point formats: FP32, FP16, BF16, FP8, TF32
    Concept
  3. 03
    Mixed precision training: FP16, BF16, and FP8
    Concept
  4. 04
    Loss spikes at scale
    Concept