pub const MASKED_ATTN_MAX_KEYS: usize = 12_288;Expand description
2026-09-25: The most keys dsa_mla_masked_attn takes: it stages the [S] f32 score row
in dynamic shared memory, and 4 · S <= 49,152 B gives 12,288.
examples/glm5next_dsa_decode_gate.rs checks the serve’s decode kernel against it, and
examples/dsa_indexer_microtest.rs launches it with dsa_topk_to_mask. The serve’s
decode is glm5next_dsa_mla_decode_fp8 (attend.rs), which reads the selected
tokens through the block table and has no S-sized shared-memory row.