fla-kda

Identify the affected path (safe vs non-safe intra, inter-solve, or backward) and keep both paths consistent unless the change is explicitly scoped to one. Then run correctness tests covering only the axes your change touches, such as varlen layouts, gate modes, and numerical extremes.

With safe_gate, activation is lower_bound * sigmoid(exp(A_log) * (g + dt_bias)), keeping per-token gates in [lower_bound, 0). Without it, activation is -exp(A_log) * softplus(g + dt_bias). safe_gate requires use_gate_in_kernel=True and -5 <= lower_bound < 0.

With lower_bound=-5, a 16-token sub-chunk can accumulate -80 in natural-log units, which would overflow exp2 if fed the full cumulative span. The midpoint offset b_gm = b_g - b_gn limits each exponent operand to about half the sub-chunk, staying below the kernel's exp2 safety threshold.

Yes. With use_gate_in_kernel=False, g is already the log-space decay tensor and A_log, dt_bias, and lower_bound are not part of the gate activation. With use_gate_in_kernel=True, A_log is required and dt_bias is optional.

Cover dense and varlen layouts, forward and backward, all three gate modes, raw and post-sigmoid beta, MHA and GVA, and gate numerical extremes including lower_bound=-5, extreme A_log, long-sequence cumulative decay, chunk boundaries, and ragged varlen boundaries.

KDA supports the FlashKDABackend and the KDATileLangBackend. Changes affecting either backend should include backend verifier behavior in the correctness checklist.