ByteDance Seed Proposes PISA: Block-Sparse Attention With O(Nlog N) Complexity
ByteDance-Seed · hf · 2026-09-28
ByteDance Seed introduces PISA, a block-sparse attention mechanism that reduces long-context attention cost from quadratic to O(N log N).
- Key idea: Conventional block selection scores all query-block pairs and stays quadratic. PISA uses a pyramid Top-K strategy: pooled keys form O(log N) coarse-to-fine levels, and LogSumExp scoring over bounded candidate sets narrows choices level by level.
- Engineering: Hardware-aware Triton kernels for training and inference fuse hierarchical routing and LogSumExp scoring without materializing the query-key score matrix.
- Results: Comparable performance to baselines on commonsense reasoning benchmarks, with better results on retrieval tasks in language modeling evaluations.
More from Infra
- Sarvam AI Launches Token Factory: India-Hosted Inference for Open Models — itsOmSarraf_ · 2026-11-16
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Jailbroken PS4 runs a 9B open-source model at up to 43 tok/s by talking directly to its 2013-era GPU — AIFlow_ML · 2026-11-16
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16
- Stratechery's Fictional Obituary: Imagining Jensen Huang's 2026 Exit From Nvidia — Gregory_C_Allen · 2026-11-16
- vLLM serves DeepSeek V4.1-Flash at 954K tokens/s per GPU, $38.96/GPU-hour on GB300 NVL72 — bookwormengr · 2026-11-16