Set diffusion factorizes likelihood over arbitrary token sets and uses a set-causal diffusion architecture to support KV caching and any-order decoding, yielding improved speed-quality tradeoffs versus prior diffusion LMs.
hub Mixed citations
Flex Attention: A Programming Model for Generating Optimized Attention Kernels
Mixed citation behavior. Most common role is background (50%).
35
Pith papers citing it
Background
50% of classified citations
abstract
Over the past 7 years, attention has become one of the most important primitives in deep learning. The primary approach to optimize attention is FlashAttention, which fuses the operation together, drastically improving both the runtime and the memory consumption. However, the importance of FlashAttention combined with its monolithic nature poses a problem for researchers aiming to try new attention variants -- a "software lottery". This problem is exacerbated by the difficulty of writing efficient fused attention kernels, resisting traditional compiler-based approaches. We introduce FlexAttention, a novel compiler-driven programming model that allows implementing the majority of attention variants in a few lines of idiomatic PyTorch code. We demonstrate that many existing attention variants (e.g. Alibi, Document Masking, PagedAttention, etc.) can be implemented via FlexAttention, and that we achieve competitive performance compared to these handwritten kernels. Finally, we demonstrate how FlexAttention allows for easy composition of attention variants, solving the combinatorial explosion of attention variants.
hub tools
citation-role summary
background 5
method 3
baseline 1
dataset 1
citation-polarity summary
representative citing papers
A 0.6B LM with length-aware attention adjustments performs competitive in-context retrieval at million-token scale on MS MARCO, NQ, and LIMIT benchmarks.
SmoothAgent introduces lookahead context engineering to eliminate transformation overhead in LLM agents, reducing TTFT by up to 11.9x through proactive KV cache preparation.
LongLive-2.0 delivers an NVFP4 parallel infrastructure that enables direct training of long multi-shot autoregressive diffusion video models and achieves up to 2.15x training and 1.84x inference speedups on Blackwell and other GPUs.
A survey that groups efficient video diffusion methods into four paradigms—step distillation, efficient attention, model compression, and cache/trajectory optimization—and outlines open challenges for practical use.
Dual Triangle Attention achieves effective bidirectional attention with built-in positional inductive bias via dual triangular masks, outperforming standard bidirectional attention on position-sensitive tasks and showing strong masked language modeling results with or without positional embeddings.
Exact Flow Linear Attention derives a closed-form exact update for delta-rule linear attention from continuous-time dynamics, removing Euler discretization error while preserving linear complexity and structure.
Causal-rCM unifies teacher-forcing and self-forcing distillation for autoregressive video diffusion, delivering a 2-step model with VBench-T2V score 84.63 and enabling interactive world models on Cosmos 3 using only synthetic data.
A geometry-aware dynamic-query transformer decoder with Local Strided Cross-Attention raises track reconstruction efficiency from 94.1% to 98.1%, halves latency, and cuts memory use by over 10x versus fixed-query baselines in a simplified HL-LHC simulation.
LCLMs are scaled 0.6B-encoder 4B-decoder compressors pre-trained on over 350B tokens that improve the Pareto frontier for general-task performance, compression speed, and peak memory in long-context language model inference.
Vortex provides a programmable frontend and backend for sparse attention in LLM serving, delivering up to 3.46x throughput over full attention while preserving accuracy.
UniPinRec unifies retrieval and ranking into a single model and pipeline deployed at Pinterest, reporting +1% engagement lift, 11.1% lower latency, and 63.6% higher QPS.
A new 30k-instance semantic segmentation dataset plus block distillation with sink tokens, dropout, and weighted loss lets block-attention models reach near full-attention performance on long texts.
Orthrus unifies autoregressive LLMs and diffusion models via shared KV cache and consensus to enable up to 7.8x parallel token generation speedup with O(1) memory overhead and lossless results.
FlashAR accelerates autoregressive image generation up to 22.9x by post-training a pre-trained raster-scan model with a complementary vertical head and dynamic fusion for two-way next-token prediction.
AdaSplash-2 introduces a histogram-based initialization for the α-entmax normalizer that cuts iterations to 1-2 and, with a sparsity-aware GPU kernel, matches or beats FlashAttention-2 training speed at moderate-to-high sparsity while delivering long-context gains.
Hybrid Forcing combines linear temporal attention for long-range retention, block-sparse attention for efficiency, and decoupled distillation to achieve real-time unbounded 832x480 streaming video generation at 29.5 FPS.
A single dense-pretrained RAT+ model can be adapted with 1B tokens to run at dilation sizes D=2..128, matching dense accuracy at D=16 and losing only 1-3 points at D=64 on reasoning and long-context benchmarks, though retrieval tasks drop far more.
SigLino distills SigLIP2 and DINOv3 into efficient vision models via asymmetric relation-knowledge distillation, token-balanced batching, and hierarchical data sampling on a new 200M-image corpus, yielding better transfer to grounding VLMs than training from scratch.
Flashlight is a compiler-native PyTorch framework that generates efficient fused kernels for arbitrary and data-dependent attention variants, supporting more cases than FlexAttention with competitive performance.
Kimi Linear hybridizes linear attention with a new KDA module to beat full attention on tasks while slashing KV cache by 75% and speeding decoding up to 6x.
Self Forcing trains autoregressive video diffusion models by performing autoregressive rollout with KV caching during training to close the exposure bias gap, using a holistic video-level loss and few-step diffusion for efficiency.
MAGI-1 is a 24B-parameter autoregressive video world model that predicts denoised frame chunks sequentially with increasing noise to enable causal, scalable, streaming generation up to 4M token contexts.
Titans combine attention for current context with a learnable neural memory for long-term history, achieving better performance and scaling to over 2M-token contexts on language, reasoning, genomics, and time-series tasks.
citing papers explorer
-
Set Diffusion: Interpolating Token Orderings Between Autoregression and Diffusion for Fast and Flexible Decoding
Set diffusion factorizes likelihood over arbitrary token sets and uses a set-causal diffusion architecture to support KV caching and any-order decoding, yielding improved speed-quality tradeoffs versus prior diffusion LMs.
-
Can Language Models Actually Retrieve In-Context? Drowning in Documents at Million Token Scale
A 0.6B LM with length-aware attention adjustments performs competitive in-context retrieval at million-token scale on MS MARCO, NQ, and LIMIT benchmarks.
-
SmoothAgent: Efficient Long-Horizon LLM-Based Agent Serving with Lookahead Context Engineering
SmoothAgent introduces lookahead context engineering to eliminate transformation overhead in LLM agents, reducing TTFT by up to 11.9x through proactive KV cache preparation.
-
LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation
LongLive-2.0 delivers an NVFP4 parallel infrastructure that enables direct training of long multi-shot autoregressive diffusion video models and achieves up to 2.15x training and 1.84x inference speedups on Blackwell and other GPUs.
-
Efficient Video Diffusion Models: Advancements and Challenges
A survey that groups efficient video diffusion methods into four paradigms—step distillation, efficient attention, model compression, and cache/trajectory optimization—and outlines open challenges for practical use.
-
Dual Triangle Attention: Effective Bidirectional Attention Without Positional Embeddings
Dual Triangle Attention achieves effective bidirectional attention with built-in positional inductive bias via dual triangular masks, outperforming standard bidirectional attention on position-sensitive tasks and showing strong masked language modeling results with or without positional embeddings.
-
Exact Flow Linear Attention: Exact Solution from Continuous-Time Dynamics
Exact Flow Linear Attention derives a closed-form exact update for delta-rule linear attention from continuous-time dynamics, removing Euler discretization error while preserving linear complexity and structure.
-
Causal-rCM: A Unified Teacher-Forcing and Self-Forcing Open Recipe for Autoregressive Diffusion Distillation in Streaming Video Generation and Interactive World Models
Causal-rCM unifies teacher-forcing and self-forcing distillation for autoregressive video diffusion, delivering a 2-step model with VBench-T2V score 84.63 and enabling interactive world models on Cosmos 3 using only synthetic data.
-
Better Queries, Cheaper Attention: Adapting Transformers for Efficient Sparse Reconstruction
A geometry-aware dynamic-query transformer decoder with Local Strided Cross-Attention raises track reconstruction efficiency from 94.1% to 98.1%, halves latency, and cuts memory use by over 10x versus fixed-query baselines in a simplified HL-LHC simulation.
-
End-to-End Context Compression at Scale
LCLMs are scaled 0.6B-encoder 4B-decoder compressors pre-trained on over 350B tokens that improve the Pareto frontier for general-task performance, compression speed, and peak memory in long-context language model inference.
-
Vortex: Efficient and Programmable Sparse Attention Serving for AI Agents
Vortex provides a programmable frontend and backend for sparse attention in LLM serving, delivering up to 3.46x throughput over full attention while preserving accuracy.
-
UniPinRec: Unifying Generative Retrieval and Ranking at Pinterest Scale
UniPinRec unifies retrieval and ranking into a single model and pipeline deployed at Pinterest, reporting +1% engagement lift, 11.1% lower latency, and 63.6% higher QPS.
-
Towards Generalization of Block Attention via Automatic Segmentation and Block Distillation
A new 30k-instance semantic segmentation dataset plus block distillation with sink tokens, dropout, and weighted loss lets block-attention models reach near full-attention performance on long texts.
-
Orthrus: Memory-Efficient Parallel Token Generation via Dual-View Diffusion
Orthrus unifies autoregressive LLMs and diffusion models via shared KV cache and consensus to enable up to 7.8x parallel token generation speedup with O(1) memory overhead and lossless results.
-
FlashAR: Efficient Post-Training Acceleration for Autoregressive Image Generation
FlashAR accelerates autoregressive image generation up to 22.9x by post-training a pre-trained raster-scan model with a complementary vertical head and dynamic fusion for two-way next-token prediction.
-
AdaSplash-2: Faster Differentiable Sparse Attention
AdaSplash-2 introduces a histogram-based initialization for the α-entmax normalizer that cuts iterations to 1-2 and, with a sparsity-aware GPU kernel, matches or beats FlashAttention-2 training speed at moderate-to-high sparsity while delivering long-context gains.
-
Long-Horizon Streaming Video Generation via Hybrid Attention with Decoupled Distillation
Hybrid Forcing combines linear temporal attention for long-range retention, block-sparse attention for efficiency, and decoupled distillation to achieve real-time unbounded 832x480 streaming video generation at 29.5 FPS.
-
RAT+: Train Dense, Infer Sparse -- Recurrence Augmented Attention for Dilated Inference
A single dense-pretrained RAT+ model can be adapted with 1B tokens to run at dilation sizes D=2..128, matching dense accuracy at D=16 and losing only 1-3 points at D=64 on reasoning and long-context benchmarks, though retrieval tasks drop far more.
-
SigLino: Efficient Multi-Teacher Distillation for Agglomerative Vision Foundation Models
SigLino distills SigLIP2 and DINOv3 into efficient vision models via asymmetric relation-knowledge distillation, token-balanced batching, and hierarchical data sampling on a new 200M-image corpus, yielding better transfer to grounding VLMs than training from scratch.
-
Flashlight: PyTorch Compiler Extensions to Accelerate Attention Variants
Flashlight is a compiler-native PyTorch framework that generates efficient fused kernels for arbitrary and data-dependent attention variants, supporting more cases than FlexAttention with competitive performance.
-
Kimi Linear: An Expressive, Efficient Attention Architecture
Kimi Linear hybridizes linear attention with a new KDA module to beat full attention on tasks while slashing KV cache by 75% and speeding decoding up to 6x.
-
Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion
Self Forcing trains autoregressive video diffusion models by performing autoregressive rollout with KV caching during training to close the exposure bias gap, using a holistic video-level loss and few-step diffusion for efficiency.
-
MAGI-1: Autoregressive Video Generation at Scale
MAGI-1 is a 24B-parameter autoregressive video world model that predicts denoised frame chunks sequentially with increasing noise to enable causal, scalable, streaming generation up to 4M token contexts.
-
Titans: Learning to Memorize at Test Time
Titans combine attention for current context with a learnable neural memory for long-term history, achieving better performance and scaling to over 2M-token contexts on language, reasoning, genomics, and time-series tasks.
-
LightTransfer: Your Long-Context LLM is Secretly a Hybrid Model with Effortless Adaptation
LightTransfer identifies lazy layers in LLMs like LLaMA and replaces their attention with streaming attention to form hybrid models, delivering up to 2.17x throughput with under 1.5% drop on LongBench and strong results on reasoning benchmarks.
-
ScalingAttention: Discovering Intrinsic Sparse Attention Topology for Video Diffusion Transformers
ScalingAttention discovers a prompt-agnostic intrinsic sparse attention topology in video DiTs via weight-encoded extraction (WEST) and fidelity-aware tuning (FAST), delivering up to 1.9x speedup.
-
NVIDIA OmniDreams: Real-Time Generative World Model for Closed-Loop Autonomous Vehicle Simulation
OmniDreams turns the Cosmos video-diffusion model into a real-time, action-conditioned driving simulator and shows that its internal representations can be fine-tuned into a competitive driving policy.
-
$R^3$: 3D Reconstruction via Relative Regression
R³ uses relative regression with confidence-weighted constraints from an MLP to support long-context offline and streaming 3D reconstruction without global coordinate assumptions.
-
Bernini: Latent Semantic Planning for Video Diffusion
Bernini is a framework that uses an MLLM planner to output semantic representations for a DiT renderer to generate or edit videos, reporting SOTA benchmark performance.
-
OSP-Next: Efficient High-Quality Video Generation with Sparse Sequence Parallelism, HiF8 Quantization, and Reinforcement Learning
OSP-Next reports 83.73% VBench score and up to 2.27x speedup via hybrid sparse attention, SSP parallelism, HiF8 quantization, and Mix-GRPO on diffusion transformers.
-
ZAYA1-VL-8B Technical Report
ZAYA1-VL-8B is a new MoE vision-language model with vision-specific LoRA adapters and bidirectional image attention that reports competitive performance against several 3B-4B models on image, reasoning, and counting benchmarks.
-
Better Models, Faster Training: Sigmoid Attention for single-cell Foundation Models
Sigmoid attention replaces softmax in single-cell foundation models to deliver better representations, faster training, and stability, backed by bounded derivatives, diagonal Jacobian, and a new efficient GPU kernel.
-
On The Application of Linear Attention in Multimodal Transformers
Linear attention delivers significant computational savings in multimodal transformers and follows the same scaling laws as softmax attention on ViT models trained on LAION-400M with ImageNet-21K zero-shot validation.
-
Toward Native Multimodal Modeling: A Roadmap
A roadmap that defines architectural nativity for multimodal models and categorizes them into Multi-to-Text, Multi-to-Target, and Multi-to-Multi types while outlining an industrial pipeline toward unified transformer-based native multimodal modeling.
- Accelerating Sparse Transformer Inference on GPU