GQLSA-V2: Content-Aware Global Block Retrieval for Linear-Time Sparse Attention

We introduce GQLSA-V2, a revision of Grouped-Query Latent Sparse Attention that replaces the static uniform-stride global block selection of the original formulation with a content-aware retrieval mechanism. GQLSA-V1 combines latent compression of key-value representations, grouped-query parameter sharing, and block-sparse attention with a constant selection size, achieving linear O(T) complexity while maintaining language modeling quality comparable to dense attention. Its single documented limitation was that global block selection was deterministic and position-based, insensitive to content. V2 removes this limitation by building a random-projection inverted index over block summaries once per forward pass, then retrieving for each query the global blocks whose content is most similar to the query's own. The mechanism uses no learnable parameters, preserves the complexity class of the forward pass, preserves causality through a strictly-past anchor block, and preserves the state dict of V1 exactly. Existing V1 checkpoints load into V2 at strict=True with no retraining. Empirically, on an NVIDIA Tesla T4, V2 achieves 91.12 ms at sequence length 4096, within 0.3% of V1's 90.89 ms. Memory is bit-identical to V1 (2384.32 MB vs 2384.29 MB at T = 4096), and WikiText-2 quality is preserved within observed run-to-run noise. The documented limitation of V1 is eliminated at no measurable cost to quality, speed, or memory.