pub fn indexer_state_bytes(capacity: usize, index_head_dim: usize) -> usizeExpand description
Bytes ONE sequence’s indexer cache occupies for ONE DSA layer at capacity rows.
SSOT for the three allocations in Glm5NextDsaState::alloc and for the serve’s
per-sequence reserve. = capacity * (4 * index_head_dim + 1), i.e. 513 B/token/layer at
GLM-5.3’s index_head_dim = 128. Replicated — EP does NOT halve it.