Host-driven loop
launch 0 / 15
CPU hostready
GPU computeone step
launch → wait → update
KV cache
grows per step
0 / 15
Code Predictor · fixed 15-step frame
continuous replay
Captured fixed loop
replay 0 / 1
CPU hostready
CUDA graphTriton attention
one replay runs all steps
KV cache
fixed storage
preallocated · 15 steps
The CPU submits once; the fixed GPU program completes the full frame.