Module state 

Per-sequence DSA indexer state — the cache the selector reads every decode step.

Why this exists at all

The indexer scores each pool from k_normed and gate, both projections of the hidden state (indexer.wk, index_kpool_compress_gate). Neither is recoverable from the MLA latent — wk · hidden cannot be inverted out of a rank-512 compression — so the indexer needs its own cache stream alongside the KV cache. HF does the same thing, keeping indexer state on a per-layer DynamicIndexedLayer via past_key_values.update_indexer.

No new subsystem

TransformerLayer::alloc_state is called once per sequence and LayerState is an Any downcast hook — the same mechanism qwen3_ssm uses for recurrent state. This is that, with a bigger buffer.

🪤 Flat, not paged

dsa_kpool_compress and dsa_index_scores index k[raw * D + d] linearly, and pools are built over absolute positions from the first valid token. So this is one contiguous per-sequence buffer, not block-table paged. The MLA latent stays paged; only the indexer stream is flat.

Structs

Glm5NextDsaState

One sequence’s indexer cache for one DSA layer.

Functions

dsa_capacity

THE authoritative indexer-capacity computation. Everything that needs to know how many rows a sequence’s indexer cache holds calls this — max_dsa_context for the allocation side, and the serve’s pre-model reserve for the budget side.

indexer_state_bytes

Bytes ONE sequence’s indexer cache occupies for ONE DSA layer at capacity rows.

max_dsa_context

Longest context DSA can select over, in tokens.