Atenção à janela deslizante

Ordinary dense self-attention forms token-pair interactions across a sequence. A sliding-window mask limits the allowed interactions to a local neighborhood, with the exact convention stated by the model or implementation. If N positions each consider at most W keys, the pairwise work is proportional to N times W, which is linear in N when W remains fixed. Other computations still contribute to a complete layer’s cost. For a causal convention in which W includes the current token, a query at position i can see positions i−W+1 through i, clipped at the beginning of the sequence. Doubling N with the same W doubles this pairwise term instead of quadrupling it. Stacking local layers can propagate information beyond a single layer’s direct neighborhood, but an expanded theoretical receptive field is not a guarantee of accurate recall over that entire span. The original Mistral 7B report specifies a 4096 window size and a rolling buffer. Its architecture table lists context_len as 8192, separate from the larger theoretical span from stacking layers. It maps cached keys and values for timestep i to slot i modulo W so older entries can be overwritten when appropriate. This bounded local-layer cache should not be generalized to every layer in a hybrid model. Gemma 2’s report describes alternating local and global attention layers with their own specified spans. Longformer instead assigns global attention to selected token locations: those tokens attend across the sequence and other tokens can attend to them. These are distinct ways to add nonlocal interactions. Their accuracy, memory, and latency require evaluation in the actual architecture.