FlashBlock: Attention Caching for Efficient Long-Context Block Diffusion

Generating long-form content, such as minute-long videos and extended texts, is increasingly important for modern generative models. Block diffusion improves inference efficiency via KV caching and block-wise causal inference and has been widely adopted in diffusion language models and video generation. However, in long-context settings, block diffusion still incurs substantial overhead from repeatedly computing attention over an ever-growing KV cache. We identify an underexplored property of block diffusion: cross-step redundancy of attention within a block. Our analysis shows that attention outputs from tokens outside the current block remain largely stable across diffusion steps, while block-internal attention varies significantly. Based on this observation, we propose FlashBlock, a cached block-external attention mechanism that reuses stable attention output, substantially reducing attention computation and KV cache access without modifying the diffusion process. Moreover, FlashBlock is orthogonal to sparse attention and can be combined as a complementary residual reuse strategy. When integrated, it substantially improves model accuracy under aggressive sparsification by offsetting much of the performance loss induced by sparsity. Experiments on diffusion language models and video generation demonstrate up to 1.44$\times$ higher token throughput and up to 1.6$\times$ reduction in attention time, with negligible impact on generation quality.

Modern generative models are increasingly used to create long outputs, such as long documents and videos, but generating such content can be slow and computationally expensive. This paper studies how to make a class of diffusion-based generative models more efficient when they work with long contexts. We find that, during generation, some attention computations over previously generated content remain largely stable across steps and do not need to be recomputed every time. Based on this observation, we propose FlashBlock, a method that caches and reuses these stable attention results while recomputing only the parts that are changing. This reduces unnecessary computation and memory access without changing the underlying generation process. Experiments on language and video generation models show that FlashBlock can speed up generation while maintaining similar output quality. This work may help make long-context generative models faster and more practical to use.