Fast-WAM → Faster-WAM

Fast-WAM vs Faster-WAM · one action chunk

From 300 cache reads to 10

Both models run the same 30-layer video model once and denoise 32 actions in 10 steps. In Fast-WAM each of 30 action layers reads its own video layer's cache at every step, so that's 30 × 10 = 300 reads. Faster-WAM mixes the 30 caches into one, once, and its single action layer reads that mix 10 times.

One action chunk, two models, one clock

video layer K V action layer text + state fused cache
model time
0
video-cache reads
0
noise level σ
1
denoiser kernels
0
Clock: phase means from a 20-chunk LeRobot run on an A100 in bfloat16 with random weights Measured. The phases leave out about 15 ms (Fast-WAM) and 9 ms (Faster-WAM) of host work, so the race ends at 945.6 against 169.5 ms, where the end-to-end timer gives 938.0 against 175.6 ms. Same clock is to scale for each phase and step, and Step by step gives every step the same screen time. σ is the noise level, and the real 10-step schedule keeps it near 1 until the last steps. The action curves are illustrative (3 of 7 values), and the fusion lines are shaded by the paper's Figure 4 head average, read by eye. Kernel counts come from a separate profiler run.

Sources