Time-series foundation models can skip whole Transformer blocks: A100 latency drops 11% to 13% at moderate sparsity, with dataset-level variation

A pretrained time-series foundation model can skip whole Transformer blocks at inference instead of running full depth for every input. These models previously executed a fixed network for every input, while global compression shrinks a network for all inputs at once. Across 1,088 held-out windows from ETTh1, Traffic, Weather, and Exchange, Chronos-Bolt with 7 of 11 candidate blocks and TimesFM with 16 of 20 blocks show little change in mean relative forecasting quality, while median A100 batch-1 latency falls 11.2% and 13.4%.

Which blocks can be removed is decided by greedy selection on calibration windows, and a skipped block returns its input before attention and feed-forward computation. Selection uses only calibration windows disjoint from the evaluation windows, so no evaluation window affects selection. Chronos-Bolt is T5-style with its first encoder block mandatory and blocks 1 to 11 as candidates, while TimesFM is decoder-only with 20 candidate Transformer blocks; counters verify inactive blocks are not called, and with all candidates active the forecasts agree exactly with the untouched model.

The work covers two pretrained models, four univariate benchmarks, and one GPU type. It establishes latency and throughput savings while all weights remain resident, so it does not address model-memory or energy savings, and it reports no edge-device benchmark.

More aggressive TimesFM skipping accelerates inference further but substantially harms Traffic and Weather, and TimesFM quality degrades sharply below 80% active blocks on those datasets. Dataset-specific masks transfer inconsistently, and Chronos representations do not support reliable unseen-domain routing, so opportunity does not imply routability.