EDGE0 // ASYNC MOE PREFETCH
1. Decode Now
2. Predict Next
3. Prefetch Next
|
Pause
Prediction: HIT
Reset
Storage Latency Penalty:
0.1ms (Hidden)
STAGE 1: CURRENT DECODE
STAGE 2: PREROUTER PREDICT
STAGE 3: PREFETCH & HANDOFF
STATE
HISTORY
COMPUTE
TOP-K IDS
MMAP LOAD
READY
BASE WEIGHTS
HIDDEN STATE
EXPERT HISTORY
RESIDENT MODEL
PREROUTER HEAD
LOAD IN FLIGHT
SSD EXPERT STORE
NEXT-TOKEN DECODE
SYSTEM OVERVIEW
Select a stage above or click any system node to inspect how Edge0 overlaps storage I/O with current token computation.