arXiv:2608. 20210v1 Announce Type: cross Abstract: Small language models are usually built like large ones and then squeezed onto a CPU afterwards.
Small language models are usually built like large ones and then squeezed onto a CPU afterwards. We did the opposite: we fixed the target first, one user, one token at a time, 4-bit weights, ordinary CPU, and chose the architecture to suit it.
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:2607. 21535v1 Announce Type: new Abstract: Speculative decoding accelerates autoregressive generation by having a cheap draft propose tokens that a target verifies in parallel.
Speculative decoding accelerates autoregressive generation by having a cheap draft propose tokens that a target verifies in parallel. Frontier models increasingly ship a built-in Multi-Token-Prediction (MTP/NEXTN) draft head under the assumption that the draft is negligibly cheap.
The paper investigates whether training Mixture-of-Experts (MoE) routers can improve memory‑bandwidth locality on consumer GPUs. Using a new zero‑surgery telemetry tool, the authors measure that a large Qwen3‑235B model is bottlenecked by disk‑based expert access, and that an LRU cache can serve a majority of requests. They pre‑register experiments training 137 M‑parameter MoE models with locality‑aware losses, finding that while cache misses can drop up to 60 % (99 % static‑pin hit rate), every configuration fails to meet a strict 1 % perplexity threshold, indicating a tight coupling between cache efficiency and model quality.
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: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.
Task-Aware Spectral Pruning (TASP) is a post‑training framework that tailors sparse masks to specific tasks by calibrating module‑level spectral descriptors against task‑specific ablation effects. It constructs masks that close grouped‑query‑attention and SwiGLU dependencies, routing each user turn to a single compiled mask that remains fixed during prefill and decoding. In experiments, TASP achieves a 43% active‑FLOP reduction while preserving 97.7% of the dense BF16 performance on Llama‑3‑70B, and delivers a 1.44× speedup on an A100 80GB with INT8‑weight/BF16‑compute, reducing decode latency from 45.2 to 31.3 ms/token.
The paper investigates using short polynomial approximations to accelerate special‑function operations in large language models on NVIDIA Blackwell GPUs. By replacing native sigmoid, tanh, and SiLU with degree‑3 or degree‑4 bfloat16 programs, the authors achieve up to 2.19× speed‑ups in isolated FP16 benchmarks and modest training‑step throughput gains (2.7–8.0%) across four integration tasks. The study also evaluates model behavior, finding negligible training‑loss differences within 100 billion tokens.
arXiv:2609.03949v2 Announce Type: replace-cross Abstract: A long-lived KV cache must be compressed before the queries that will read it exist. Selection by observed attention collapses there: on a No...
arXiv:2607. 21927v1 Announce Type: new Abstract: Full self-attention in large language models scales as O(N^2), which limits long-context document analysis to 65,536 tokens and requires costly GPU clusters.
arXiv:2606. 30709v1 Announce Type: cross Abstract: Hierarchical Global Attention (HGA) is a drop-in replacement for dense causal attention in pretrained long-context transformers.