HyScale is a long‑context training system designed to accelerate learned sparse attention for million‑token contexts. It introduces HyTopK, which computes each query’s exact top‑k indices on the fly without a full score table, and selection replay, which reuses forward‑pass selections during recomputation. These techniques, combined with system‑wide optimizations, achieve a 2.07× speedup for GLM‑5.1 training on 1M‑token sequences compared to Megatron‑LM, and similar gains for DeepSeek‑V4‑Flash and Qwen3.8‑Flash‑Next on large GPU clusters.
arXiv:2610.42147v1 Announce Type: new Abstract: Agentic workflows are increasingly shifting towards long-horizon tasks, such as open-ended scientific work, where models benefit from long context wind...
HISA: Efficient Hierarchical Indexing for Fine-Grained Sparse Attention proposes a two-stage hierarchical indexer that replaces the flat token scan used in token-level sparse attention mechanisms like DeepSeek Sparse Attention. The method first performs block-level coarse filtering to discard irrelevant regions, then applies the original token-level indexer only within the retained candidate blocks, preserving the same top-sparse pattern for downstream attention. Benchmarks show HISA achieves significant speedups at 64K context and matches the quality of DeepSeek-V3.2 and GLM-5 without additional training.
arXiv:2606. 19348v1 Announce Type: cross Abstract: We present a preview version of DeepSeek-V4 series, including two strong Mixture-of-Experts (MoE) language models -- DeepSeek-V4-Pro with 1.
Learned sparse attention, used in production models such as DeepSeek-V4 and GLM-5, lets each query attend to only k keys chosen by a lightweight indexer, but the indexer still scores every key in the...
arXiv:2606. 13392v1 Announce Type: new Abstract: Ultra-long-context capability is becoming indispensable for frontier LLMs: agentic workflows, repository-scale code reasoning, and persistent memory all require the model to jointly attend over hundreds of thousands to millions of tokens, yet the quadratic cost of softmax attention makes this untenable at deployment scale.
arXiv:2605. 16928v2 Announce Type: replace-cross Abstract: Long-context inference in large language models is bottlenecked by the quadratic cost of full attention.
arXiv:2608.30320v1 Announce Type: new Abstract: We describe the architecture and ablations of Qwen3.8-Flash-Next, a sparse mixture-of-experts model with 125B parameters, 6B activated per token, and a...
The report introduces A.X K2, a 688‑parameter Mixture‑of‑Experts language model designed for agentic applications. Trained on 8.5 trillion tokens, it surpasses its predecessor A.X K1 by over 30 percentage points on several benchmarks, thanks to a higher‑quality data mix and improved token efficiency. Key innovations include Sparse Gated Attention for efficient long‑context handling, Gated Norm for training stability, and a Think‑Fusion recipe that allows switching between thinking and non‑thinking modes within the same model.
The paper reports that chain‑of‑thought (CoT) supervised fine‑tuning (SFT) improves reasoning but systematically harms long‑context recall in hybrid linear‑attention models such as HypeNet and Jet‑Nemotron. Retrieval performance on the Needle‑In‑A‑Haystack benchmark drops dramatically after CoT‑SFT, especially with harder settings and longer contexts. The authors introduce QK‑Restore, a training‑free method that reinstates the query‑key projection matrices from the pre‑SFT checkpoint, which recovers long‑range recall while preserving reasoning gains.
arXiv:2606. 09079v1 Announce Type: cross Abstract: Conventional LLMs keep the full KV cache loaded during decoding, causing a severe GPU memory bottleneck for ultra-long context serving.
Block Sparse Flash Attention (BSFA) is a drop‑in replacement for FlashAttention that speeds up long‑context inference by pruning about 50% of computation and memory transfers. It selects the top‑k most important value blocks for each query using exact query‑key similarities and calibrated per‑layer, per‑head thresholds, requiring only a one‑time training‑free calibration. On Llama‑3.1‑8B, BSFA delivers up to 1.13× speedup on LongBench with a 1.1% accuracy drop and up to 1.24× on Needle‑in‑a‑Haystack retrieval with a 1% drop, while the attention kernel itself accelerates by up to 1.38×.
arXiv:2512.16391v2 Announce Type: replace-cross Abstract: Attention is the dominant source of latency during long-context LLM inference, an increasingly popular workload with reasoning models and RAG...