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:2606. 20005v1 Announce Type: cross Abstract: Attention distillation, which trains one attention distribution to match another by minimizing their Kullback-Leibler (KL) divergence, is widely used in knowledge distillation, model compression, continual learning, and sparse-attention LLM training.
FlyWheel introduces optimized TPU kernels for softmax and linear attention, reducing vector work and improving matrix operand reuse. By overlapping independent computations, FlyWheel achieves up to 4.14× speedups over the RPA kernel and 1.80× over Splash Attention, reaching 81% of peak BF16 throughput on TPU v6e. It also boosts Gated DeltaNet prefill by up to 3.61× and enhances vLLM serving of Qwen3.8-27B prefill throughput by up to 1.56× while preserving accuracy.
arXiv:2608.30386v1 Announce Type: cross Abstract: Hybrid linear-attention architectures have recently scaled to large open-weight models, offering quality competitive with full attention while substa...
arXiv:2607. 14568v1 Announce Type: cross Abstract: A companion study ran a 35B mixture-of-experts model on a 2011 NVIDIA Tesla C2075 (Fermi, sm_20, 6GB) as a GPU-prefill/CPU-decode hybrid, because the 4-bit model did not fit in device memory (arXiv:2606.
The paper introduces Group-Query Latent Attention (GQLA), a variant of Multi-head Latent Attention that shares group‑sharded up‑projections to enable a joint latent representation and a shardable expanded cache. GQLA supports both MQA‑absorb and GQA execution using the same weights and existing kernels, and roofline analysis shows how absorption trades traffic for computation while GQA shards KV groups. Experiments on DeepSeek‑V3.1‑Base and other models demonstrate that GQLA achieves comparable training losses to MLA, maintains high accuracy after checkpoint conversion, and improves output throughput by up to 35% on eight H20‑3e GPUs, while also allowing larger batch sizes due to a smaller per‑rank cache.
DeltaLLM is a training‑free framework that accelerates on‑device large language model inference by introducing delta‑key sparse attention and a vectorized structured sparse online softmax. It reuses delta keys across query heads, preserves exact attention near the diagonal while approximating non‑local regions, and exploits hardware‑friendly sparsity with a specialized vector‑processor architecture. Experiments on BitNet‑b1.58‑2B‑4T and LLaMA3.2‑1B‑Instruct show up to 60% delta‑key sparsity, 63% structured attention‑score sparsity, and a 1.62× speedup on the KV260 FPGA with minimal accuracy loss.
BellowsKV introduces a training‑free approach that maps the hidden states of multiple adjacent layers in a large language model to a single shared latent representation per token, allowing each layer to retrieve its own keys and values via a layer‑specific projection. The method retains the latent at full dimension and offers two compression strategies: BellowsKV‑Quant, which stores the latent at low precision, and BellowsKV‑Evict, which makes a shared eviction decision per block. Experiments on Llama‑3.1‑8B, Mistral‑7B, and Qwen2.5‑7B show that BellowsKV‑Quant preserves about 99% of full‑cache accuracy at 3.6× compression and 97‑98% at 13‑14×, outperforming existing cross‑layer methods that drop to at most 84% accuracy at 16× compression.
arXiv:2607. 15456v1 Announce Type: new Abstract: Looped, weight-tied Transformers reduce parameters by reusing a block, but decoding still stores a separate K/V cache for every recurrence step.
The paper investigates how block‑diffusion language models can use a constant‑size cache to enable efficient parallel decoding. By employing sequence mixers that summarize completed blocks into a reusable state and a block‑causal training objective, the authors pretrain three 3B block‑diffusion denoisers (attention, Mamba, and hybrid) on 300 B tokens. The resulting state‑space cache remains O(1) in memory and latency regardless of context length, yielding significant speed‑up and memory savings compared to traditional attention‑based caches, especially at very long sequences.
arXiv:2605. 15250v3 Announce Type: replace-cross Abstract: Multi-head Latent Attention (MLA), the attention used in DeepSeek-V2/V3, jointly compresses keys and values into a low-rank latent and matches the H100 roofline almost perfectly.
arXiv:2610. 23796v1 Announce Type: new Abstract: Transformer prefill reuses weights across tokens and is dominated by matrix multiplication and activation data movement; single-token decode has limited reuse and is dominated by weight and KV-cache reads.