TaylorSeer¶
TaylorSeer is a cross-step feature forecasting method. Instead of only replaying cached features, it extrapolates features from previous timesteps with a Taylor-series approximation to reduce error when reuse intervals become larger.
Role in Sol-Engine¶
TaylorSeer is part of the cache design space considered by the cache agent. It is useful as a candidate when plain residual replay produces too much drift at the desired skip interval.
Search knobs¶
- forecast order.
- warmup steps before forecasting.
- maximum cached interval.
- fallback recompute cadence.
When to use¶
Use TaylorSeer-style forecasting when the deployment needs more aggressive cross-step reuse than TeaCache or EasyCache can safely provide. Validate especially on motion consistency and high-frequency detail.