IndexShare lets nearby layers reuse the token positions selected by a DeepSeek Sparse Attention (DSA) indexer.

Architecture gallery GLM-5.2 note Z.ai release post IndexCache paper

Figure 1. One full indexer result supplies four GLM-5.2 layers.

What is shared

Selected token positions

GLM-5.2 pattern

full, shared, shared, shared

Example architectures

GLM-5.2

The DSA selection being reused

The k in top-k, not to be confused with the k that is used for the keys in the DSA equation, is a hyperparameter that is set to 2048 in the DeepSeek model code. The indexer and token selector result in each token attending to a few past tokens that the model has learned to consider most relevant, rather than all tokens or a fixed local window.

The goal was not to improve performance over the dense base model but to reduce performance degradation from the sparse attention mechanism while benefiting from improved efficiency.

The four-layer reuse pattern

indexer_type = [full, shared, shared, shared]

if indexer_type == full:
    selected = top_k(indexer(hidden_state))
else:
    selected = previous_selected

output = attention(hidden_state, selected)  # New output in every layer

Figure 2. GLM-5.2 marks each DSA indexer as full or shared across its 78-layer stack (Original source GLM-5.2 and IndexShare for Long-Context Sparse Attention).

Evidence and tradeoff

Item	Reported detail	What it means
Adjacent-layer overlap	The IndexCache study measured 70% to 100% overlap between adjacent top-k sets.	A new indexer often rediscovers positions selected one layer earlier.
GLM-5.2 training	The four-layer pattern was introduced during continued mid-training with 128K-token sequences.	Retained indexers can adapt to serving the following layers.
GLM-5.2 result	Z.ai reports 2.9× fewer per-token FLOPs at a one-million-token context.	The reported gain is specific to GLM-5.2 and this context length.
IndexCache result	Removing 75% of indexer computations on a 30B DSA model produced up to a 1.82× prefill speedup and a 1.48× decode speedup at 200K.	These are separate system measurements and should not be compared directly with the GLM-5.2 FLOP result.
Tradeoff	A shared layer cannot select fresh positions from its current hidden states.	It uses the most recent full layer’s positions, while computing its own attention output.

Sources

GLM-5.2 release post GLM-5.2 configuration IndexCache paper GLM-5.2 quick note DeepSeek Sparse Attention explainer

Back to architecture gallery