Zyphra发布层级注意力ZHA：稀疏搜索实现亚二次复杂度

Zyphra has released Zyphra Hierarchical Attention (ZHA), a sparse attention method built around a tree-structured indexer. ZHA targets a known gap in sparse attention: the indexer that picks tokens still scans the whole context. Zyphra’s technical paper shows ZHA can be distilled into frozen pretrained models with about 4B tokens. On Qwen3.6-35B-A3B it stays close to dense attention on reasoning, agentic coding and 256K retrieval benchmarks.

• Size: It is a method that adds 81,920 trainable parameters to Qwen3.6-35B-A3B (about 0.00024%). Tested up to 256K context.
• Runs on: Retrofits frozen pretrained models after a 4B-token KL warmup at 32K context. Speedups come from a modeled H100 BF16 roofline. No optimized kernels released. An FP4 router variant was also tested.
• Performance: Within a few points of dense attention on 7 reasoning, coding and retrieval benchmarks.
• Best: LongBench v2 hits 63.4% (deep tree) vs 58.0% dense, using about 1.1% of keys.
• Worst: On NoPE-based Nemotron-3.5-Lightning-30B-A3B, deep-tree RULER-13 drops 25.3 points below dense.

What is Zyphra Hierarchical Attention (ZHA)?

ZHA is a sparse attention method from Zyphra that finds relevant keys by searching a tree. It keeps the full KV cache. The tree only decides which cached keys each query reads.

Fixed-state methods like SSMs and sliding windows saturate at long contexts. Fine-grained sparse methods like DeepSeek Sparse Attention score every past token with an indexer. The final attention is sparse, but the indexer still grows with the full context.

Near 1M tokens, cited work shows indexing takes 60 to 90% of layer time. ZHA makes the search itself sparse.

How does ZHA’s tree indexer work?

Building the tree

ZHA splits keys into contiguous chunks and compresses each chunk into a summary key. Summaries are grouped and compressed again, level by level. With chunk size 4, 6 levels cover 4,096 tokens per top node.

Summaries come from a small set of learned reference queries. Each reference query gives a tangent-plane estimate of a chunk’s log attention mass. The max over these planes approximates how much attention any future query would give the chunk.

How does top-p routing pick tokens?

At decode time, a query scores the top-level nodes. It keeps the smallest set holding a fraction p of estimated attention mass. It then expands only those children, repeating until it reaches raw KV entries.

Top-p makes the budget adaptive. A single-fact lookup touches few branches. A query that needs many scattered facts expands more. The first token is always kept as an attention sink.

How does ZHA perform against dense attention and MSA?

Zyphra team converted Qwen3.6-35B-A3B with 2 trees: deep (chunk 4, 6 levels) and wide (chunk 256, 1 level). Both use 16 reference queries and p = 0.95. It also trained a MiniMax Sparse Attention (MSA) indexer on 40B tokens, 10x ZHA’s budget.

Source: arXiv 2610.42112, Figure 5. All numbers self-reported by Zyphra.

The biggest gap is on retrieval. MSA drops to 77.1% on RULER-13, while both ZHA trees stay above 93%. On LongBench v2, the deep tree and MSA read similar key shares (1.1% vs 1%). The deep tree scores 14.1 points higher.

Across 28 tree geometries, RULER accuracy held steady when attending to 10 to 30% of keys. At fixed compression, deeper trees kept accuracy at higher sparsity. A 4-parameter model predicts RULER accuracy from geometry with 1.44-point RMSE.