nx_attn_window.nx

nx_attn_window.nx -- sliding-window (local) attention kernel. Ships VRAM-track W-001 per docs/VRAM_OPTIMIZATION_REALISTIC_TRACKING.md: sub-quadratic attention via a fixed-width local receptive field. Each query token i attends only to K/V positions within [i - W, i + W], so the score matrix shrinks from [n_q, n_kv] to [n_q, 2W+1]. For n=4096 tokens, W=256 -> 8x smaller scores buffer (4096 * 513 vs 4096 * 4096). Working memory savings dominate at long context. Composition (bits-up cardinal): * Operates on `*NxTensor` -- canonical L1 container from nx_tensor.nx (the same surface nx_attention.nx uses). * Reuses nx_attn_softmax_row_q10 from nx_attention.nx -- the softmax kernel is row-wise so it works on the banded [n_q, 2W+1] layout unchanged. * Sentinel-mask for invalid slots: NX_ATTN_W_NEG_INF (chosen so the existing _attn_exp_q10 clamps it to 0). ===== Math ======================================================= Dense attention (existing): scores[i, j] = Q[i] . K[j] for all i, j weights[i, *] = softmax(scores[i, *]) out[i, d] = sum_j weights[i, j] * V[j, d] Sliding-window: band_start(i) = max(0, i - W) band_len(i) = min(2W + 1, n_kv - band_start(i)) scores[i, j] = Q[i] . K[band_start(i) + j] for j in [0, band_len(i)) scores[i, j] = NEG_INF for j in [band_len(i), 2W+1) weights[i, *] = softmax(scores[i, *]) (NEG_INF slots -> 0 weight) out[i, d] = sum_{j < band_len(i)} weights[i, j] * V[band_start(i) + j, d] Identity check: with W >= n_kv - 1, every query has band_len == n_kv, no invalid slots, and the result equals dense attention exactly. ===== Quality envelope (the honest measurement basis) ============= Sliding-window matches dense attention quality within ~1% on