02.30 · UNIT 08 · Train, evaluate, and serve your transformer · Lesson
Checkpoints, inference, and responsible model reuse
Reliable reuse saves learned tensors plus the configuration that gives them meaning.
PLAIN-LANGUAGE INTRODUCTION
What is this?
Reliable reuse saves learned tensors plus the configuration that gives them meaning.
One simple example
A fixed input returns logits [1.2,−0.4,0.8] before saving. Reloading in evaluation mode should match them.
What goes in?
A checkpoint, matching model structure, tokenizer, and fixed test input.
What comes out?
Reproducible logits or documented fine-tuning results.
Why does it matter?
A reload check catches missing parameters and configuration mismatches.
What is it not?
A weight file alone does not define preprocessing or label meanings.
WORK THROUGH THE IDEA
See the idea in more detail
- Logits are raw output scores. This illustration uses
[1.2,−0.4,0.8]for one fixed input. - Save the state dictionary plus model configuration, vocabulary, and special-token IDs.
- Recreate the same architecture. Load the tensors. Set evaluation mode and disable gradient recording.
- Run the fixed input again. Compare all logits within a stated numeric tolerance.
- Common mistake: dropout in training mode can make two correct runs produce different logits.