⚡

LLM Execution Skeleton (Strict 8-Step Flow)

Sequential Pipeline

Click any step below to inspect data inputs, outputs, and core transformations.

Click step or press "Visualize"
↻
AUTOREGRESSIVE FEEDBACK LOOPStep 8 ➔ Re-enters Step 1

Stage 8 outputs the newly predicted word token ➔ Appends it onto the prompt sequence ➔ Re-enters Stage 1 to generate word #2, word #3, etc.

5

Step 5: Transformer BackboneRepeat 30-N Layers

Stage 5 of 8
Data Input (From Step 4)

Embedded Vectors

Data Output (To Step 6)

Contextual Hidden State (h_final)

💡
Student Takeaway:

THE BRAIN! Self-Attention mixes context across tokens, while FFN/MoE layers execute deep reasoning.