LLM-Visualized

L o a d i n g ...

Layer 1 / 12

Parameters used

0.00 M

Equations

Selection

Loading token matrix…

Token -- Dim -- Band -- Value --

Transformer inference

KV Cache

Pre-fill

Pre-fill Reuse cache Decode

Stores Keys and values for prior tokens

Current pass Build the prompt cache

Why faster Avoids recomputing the full prefix every token

≤-2 0 ≥+2

Raw token Token ID Position

-- -- --

Input dimension —

Output dimension —

Bias dimension —

Number of parameters —

Current stage —

Token note —

Attention score matrix

Pre-softmax

Source Token

Query Token

<= -5

>= +5

No attention scores available.

Query token Source token Score

-- -- --

DEV MODE

Vector sampling

Settings

Show equations overlay Displays the equation panel for the current transformer step. Enable KV cache mode Caches attention keys and values for reuse across prefill and decode passes.

KV cache enabled for prefill and decode.

Enable dev mode Shows vector sampling values and indices in the selection panel.

Global animation speed

%

Adjusts the overall speed multiplier for scene playback.

Global brightness

%

About the project

LLM-Visualized

Skip options

Choose how far to fast-forward from this point in the animation.

No skip actions are available right now.