- Splits K/V into 128-token segments - Runs FMHA per segment, merges with exp(lse) weighted sum - Tests: s_k=256 (2 tiles), s_k=512 (4 tiles) - Uses reference attn_sum for normalization
- Splits K/V into 128-token segments - Runs FMHA per segment, merges with exp(lse) weighted sum - Tests: s_k=256 (2 tiles), s_k=512 (4 tiles) - Uses reference attn_sum for normalization