LLM INFERENCE · VECTOR FLOW
Residual Stream
Attention
MLP / FFN
LayerNorm
GPT-style · d=768 · h=12 · L=12
Input Tokens
Component Legend
Attention (cross-token)
MLP / FFN (per-token)
Layer Norm
Residual Stream
Current Component
Embedding Layer
Residual Stream · d=768 → PCA(32)
每條代表一個 PCA 主成分方向的投影強度
Attention Pattern (Layer
—
)
列=Query token，欄=Key token
Operation Notes
按下 PLAY 開始模擬…
▶ PLAY
STEP →
↺ RESET
Speed
Layer
0
/ 12
Phase
embed
Particles
0