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:2609.13612v1 Announce Type: new Abstract: Modern AI systems are built on the Transformer architecture, whose core operation, attention, accounts for the majority of computation and memory cost....
FlashBoB introduces an I/O‑efficient algorithm for exact backward‑over‑backward (BoB) in softmax attention, enabling precise second‑order differentiation without large intermediate tensors. By exploiting a hierarchical affine structure, the method confines computation to on‑chip tiles and limits off‑chip memory traffic, achieving θ(N² d²/M) HBM usage. Experiments show FlashBoB scales to sequence lengths of 262K on a single A100 GPU, outperforming prior exact baselines and FlashBack by up to 6.3×.
arXiv:2608. 20210v1 Announce Type: cross Abstract: Small language models are usually built like large ones and then squeezed onto a CPU afterwards.
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.
arXiv:2508. 18224v3 Announce Type: replace-cross Abstract: Recent advances in sparse attention mechanisms have demonstrated strong potential for reducing the computational cost of long-context training and inference in large language models (LLMs).
The paper introduces TACO, a new optimizer for fine‑tuning large language models that drastically reduces optimizer state memory while preserving first‑order gradients. TACO selects the sign of the largest magnitude entry in each column of weight matrices, achieving a 174× reduction in persistent optimizer memory compared to AdamW8bit and a 2.9× decrease in peak training memory on OPT‑13B. This allows full‑parameter fine‑tuning of 30–32B‑parameter models on a single 80 GB GPU across multiple model families and tasks, with comparable accuracy and runtime to existing methods.
TileMix is a tile‑centric mixed‑precision attention kernel that routes score‑tile groups within fused dense attention to either FP16 or INT8 computation, using compact bitmasks to decide precision per tile. By partitioning the attention matrix into hardware‑aligned tiles and updating a shared online‑softmax state, TileMix preserves dense token connectivity without requiring training and supports grouped‑query attention, variable‑length batches, and INT8 key/value caches. Benchmarks on LLaMA, Qwen, and Vicuna show that TileMix restores long‑context quality lost with uniform INT8 and improves prefill throughput over FP16, offering a controllable accuracy‑efficiency trade‑off across model families.
arXiv:2605. 01910v2 Announce Type: replace-cross Abstract: Autoregressive decoding becomes bandwidth-limited at long contexts, as generating each token requires reading all $n_k$ key and value vectors from KV cache.
The paper introduces TANGO, a Token‑Aggregated Nonlinear Gating Operator that blends cross‑token mixing and token‑wise transformation in transformer architectures. By computing a nonlinear gate per source token and averaging these gates for each destination, TANGO forms a source‑conditioned linear operator that improves predictive performance. Experiments on web text, formal mathematics, and code show that full‑prefix TANGO achieves the lowest test negative log‑likelihood across 16 settings, while a narrower variant offers substantial throughput gains with only a modest increase in loss.
arXiv:2609.15810v1 Announce Type: new Abstract: Diffusion Transformers deliver state-of-the-art video generation, but their long spatiotemporal sequences make attention the dominant deployment cost,...
The paper introduces a reproducible benchmark for evaluating attention mechanisms in tabular foundation models, focusing on the distinct row and column attention patterns that differ from language model attention. It compares several backends—Torch SDPA, FlashAttention variants, vLLM, and SageAttention—across realistic tabular shapes on A100, H100, and B200 GPUs, revealing that optimal backend choice varies by attention type, hardware, and model specifics. The study finds FlashAttention generally performs best, but CuDNN can outperform it for column attention on longer sequences, while SageAttention excels for large row sequences beyond 16k rows.
arXiv:2606. 07713v1 Announce Type: cross Abstract: The attention mechanism is the dominant computational bottleneck in modern transformer-based AI.