FG-Attn: Leveraging Fine-Grained Sparsity In Diffusion Transformers

• URL: http://arxiv.org/abs/2509.16518v1
• Date: Sat, 20 Sep 2025 03:48:32 GMT
• Title: FG-Attn: Leveraging Fine-Grained Sparsity In Diffusion Transformers
• Abstract summary: We propose FG-Attn, a sparse attention mechanism for long-context diffusion transformers.<n>Our approach skips computations at the granularity of Mx1 slices of the attention map.<n>It achieves an average 1.55X speedup for 5 second, 480p videos, and an average 1.41X for 5 second, 720p videos on a single H100 GPU.
• Score: 6.260564859775371
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: Generating realistic videos with diffusion transformers demands significant computation, with attention layers the central bottleneck; even producing a short clip requires running a transformer over a very long sequence of embeddings, e.g., more than 30K embeddings for a 5-second video, incurring significant latency. Prior work aims to mitigate this bottleneck by exploiting sparsity in the attention layers to reduce computation. However, these works typically rely on block-sparse attention, which skips score computation only when all entries in a block of attention scores (corresponding to M queries and M keys, with M = 64 typically) are zero. This coarse-granular skipping of attention scores does not fully exploit sparsity in the attention map and leaves room for improvement. In this work, we propose FG-Attn, a sparse attention mechanism for long-context diffusion transformers that leverages sparsity at a fine granularity. Unlike block-sparse attention, which skips entire MxM blocks, our approach skips computations at the granularity of Mx1 slices of the attention map. Each slice is produced by query-key dot products between a block of query vectors and a single key. To implement our proposed sparse attention mechanism, we develop a new efficient bulk-load operation called asynchronous-gather load. This load operation gathers a sparse set of relevant key-value vectors from memory and arranges them into packed tiles in the GPU's shared memory. Only a sparse set of keys relevant to those queries are loaded into shared memory when computing attention for a block of queries, in contrast to loading full blocks of key tokens in block-sparse attention. Our fine-grained sparse attention, applied to video diffusion models, achieves an average 1.55X (up to 1.65X) speedup for 5 second, 480p videos, and an average 1.41X (up to 1.49X) for 5 second, 720p videos on a single H100 GPU.

Related papers

• HERALD: High-Throughput Block Diffusion LLM Serving via CPU-GPU Cooperative KV Cache Retrieval [31.594043836825403]
  Diffusion LLMs (dLLMs) improve GPU utilization over autoregressive decoding by generating multiple tokens per forward pass.<n>Their KV cache still grows linearly with context, limiting throughput at long contexts.<n>We present HERALD, a KV offloading system for block dLLMs.
  arXiv  Detail & Related papers  (2026-06-19T17:36:51Z)
• ZoomR: Memory Efficient Reasoning through Multi-Granularity Key Value Retrieval [58.575695990976136]
  Large language models (LLMs) have shown great performance on complex reasoning tasks.<n>LLMs often require generating long intermediate thoughts before reaching a final answer.<n>We introduce ZoomR, a novel approach that enables LLMs to adaptively compress verbose reasoning thoughts into summaries.
  arXiv  Detail & Related papers  (2026-04-13T02:00:35Z)
• Small Vision-Language Models are Smart Compressors for Long Video Understanding [73.65465038390771]
  Long-form video understanding relies on intent-driven efficiency rather than greedily padded context.<n>We propose Tempo, an efficient query-aware framework compressing long videos for downstream understanding.<n>Tests show Tempo compresses hour-long videos substantially below theoretical limits, proving true long-form video understanding relies on intent-driven efficiency.
  arXiv  Detail & Related papers  (2026-04-09T11:40:25Z)
• Fast Autoregressive Video Diffusion and World Models with Temporal Cache Compression and Sparse Attention [37.91838955436801]
  Autoregressive video diffusion models enable streaming generation, opening the door to long-form synthesis, video world models, and interactive neural game engines.<n>As generation progresses, the KV cache grows, causing both increasing latency and escalating GPU memory, which in turn restricts usable temporal context and harms long-range consistency.<n>We propose a unified, training-free attention framework for autoregressive diffusion: TempCache compresses the KV cache via temporal correspondence to bound cache growth; AnnCA accelerates cross-attention by selecting frame-relevant prompt tokens using fast approximate nearest neighbor matching; and AnnSA sparsifies self-attention by restricting each query
  arXiv  Detail & Related papers  (2026-02-02T08:31:21Z)
• Spava: Accelerating Long-Video Understanding via Sequence-Parallelism-aware Approximate Attention [63.69228529380251]
  Spava is a sequence-parallel framework with optimized attention for long-video inference.<n>Spava delivers speedups of 12.72x, 1.70x, and 1.18x over FlashAttn, ZigZagRing, and APB, without notable performance loss.
  arXiv  Detail & Related papers  (2026-01-29T09:23:13Z)
• Block Sparse Flash Attention [29.499030734003952]
  Block-Sparse FlashAttention is a drop-in replacement for FlashAttention.<n>It computes exact query-key similarities to select the top-k most important value blocks for each query.<n>It achieves up to 1.10x speedup on real-world reasoning benchmarks and up to 1.24x needle-in-a-haystack retrieval tasks.
  arXiv  Detail & Related papers  (2025-12-07T21:20:12Z)
• SLA: Beyond Sparsity in Diffusion Transformers via Fine-Tunable Sparse-Linear Attention [88.47701139980636]
  In Diffusion Transformer (DiT) models, particularly for video generation, attention latency is a major bottleneck.<n>We find that attention weights can be separated into two parts: a small fraction of large weights with high rank and the remaining weights with very low rank.<n>We propose SLA, a trainable attention method that fuses sparse and linear attention to accelerate diffusion models.
  arXiv  Detail & Related papers  (2025-09-28T17:58:59Z)
• Video-XL-2: Towards Very Long-Video Understanding Through Task-Aware KV Sparsification [9.615466029246694]
  Video-XL-2 is a novel MLLM that delivers superior cost-effectiveness for long-video understanding based on task-aware KV sparsification.<n>It is capable of processing over 10,000 frames on a single NVIDIA A100 (80GB) GPU and thousands of frames in just a few seconds.
  arXiv  Detail & Related papers  (2025-06-24T01:19:56Z)
• Minute-Long Videos with Dual Parallelisms [57.22737565366549]
  Diffusion Transformer (DiT)-based video diffusion models generate high-quality videos at scale but incur prohibitive processing latency and memory costs for long videos.<n>We propose a novel distributed inference strategy, termed DualParal.<n>Instead of generating an entire video on a single GPU, we parallelize both temporal frames and model layers across GPUs.
  arXiv  Detail & Related papers  (2025-05-27T11:55:22Z)
• Delta Attention: Fast and Accurate Sparse Attention Inference by Delta Correction [52.14200610448542]
  A transformer has a quadratic complexity, leading to high inference costs and latency for long sequences.<n>We propose a simple, novel, and effective procedure for correcting this distributional shift.<n>Our method can maintain approximately 98.5% sparsity over full quadratic attention, making our model 32 times faster than Flash Attention 2 when processing 1M token prefills.
  arXiv  Detail & Related papers  (2025-05-16T13:48:33Z)
• MagicPIG: LSH Sampling for Efficient LLM Generation [41.75038064509643]
  We show that TopK attention itself suffers from quality degradation in certain downstream tasks because attention is not always as sparse as expected.<n>We propose MagicPIG, a heterogeneous system based on Locality Sensitive Hashing (LSH)<n>MagicPIG significantly reduces the workload of attention while preserving high accuracy for diverse tasks.
  arXiv  Detail & Related papers  (2024-10-21T16:44:51Z)
• vTensor: Flexible Virtual Tensor Management for Efficient LLM Serving [53.972175896814505]
  Large Language Models (LLMs) are widely used across various domains, processing millions of daily requests. Large Language Models (LLMs) are widely used across various domains, processing millions of daily requests.
  arXiv  Detail & Related papers  (2024-07-22T14:37:58Z)
• FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning [11.508362885430133]
  We exploit the asymmetric GPU memory hierarchy to bring significant memory saving and runtime speedup. FlashAttention is still not nearly as fast as optimized matrix-multiply (GEMM) operations, reaching only 25-40% of the theoretical maximum FLOPs/s. We propose FlashAttention-2, with better work partitioning to address these issues.
  arXiv  Detail & Related papers  (2023-07-17T17:50:36Z)
• EVEREST: Efficient Masked Video Autoencoder by Removing Redundant Spatiotemporal Tokens [57.354304637367555]
  We present EVEREST, a surprisingly efficient MVA approach for video representation learning. It finds tokens containing rich motion features and discards uninformative ones during both pre-training and fine-tuning. Our method significantly reduces the computation and memory requirements of MVA.
  arXiv  Detail & Related papers  (2022-11-19T09:57:01Z)
• FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness [80.3586155104237]
  FlashAttention is an IO-aware exact attention algorithm for Transformers. It reduces the number of memory reads/writes between GPU high bandwidth memory (HBM) and GPU on-chip. FlashAttention and block-sparse FlashAttention enable longer context in Transformers.
  arXiv  Detail & Related papers  (2022-05-27T17:53:09Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.