feat: banded local (Longformer) attention — fix O(T^2) long-audio OOM#9

Merged

mudler merged 7 commits into

feat/banded-local-attn

Jun 6, 2026

Commits on Jun 6, 2026

feat(attn): chunk-matmul banded local attention (O(1) graph nodes)
mudler
committed
feat(encoder): route local attention to chunk-matmul, lift window cap to 128
mudler
committed
docs(bench): add long-audio banded-attention section
mudler
committed