33种量化方法实测，找到视频生成长序列推理的内存优化最佳方案。
KV Cache Quantization for Self-Forcing Video Generation: A 33-Method Empirical Study
- 采用软剪枝INT4量化压缩KV缓存，实现5.49倍压缩率
- 峰值显存从19.28GB降至11.7GB，仅小幅增加运行时间
- 揭示当前框架内存瓶颈，指导未来优化方向
自强迫视频生成通过反复回传生成内容扩展短时模型以支持更长视频生成，但随之带来系统瓶颈：键值（KV）缓存随生成长度线性增长。本文基于Wan2.1的自强迫架构，对33种量化与缓存策略进行了全面实证研究，涵盖610次提示级观察和63项基准评估，覆盖MovieGen单次10秒生成与StoryEval长叙事稳定性两种场景。联合评估峰值显存、运行时间、压缩比、VBench图像质量及BF16参照下的保真度（SSIM、LPIPS、PSNR）和终端漂移。三大发现稳健：其一，受FlowCache启发的软剪枝INT4方案在实际部署中表现最优，达5.42-5.49倍压缩，显存由19.28GB降至约11.7GB，运行时间增幅有限；其二，高保真压缩方法如PRQ_INT4与QUAROT_KV_INT4虽质量好，但因严重牺牲性能或内存而难以实用；其三，名义压缩不足为据：部分方法虽减小存储量，却仍超出BF16峰值显存，因现有集成在注意力与刷新阶段仍重构或保留大型BF16缓冲区。本研究构建了基准测试工具、分析流程与量化实践地图，明确了当前可行方案与未来研究方向。代码、数据与展示仪表盘见https://github.com/suraj-ranganath/kv-quant-longhorizon/
原文摘要 · Abstract (English)
Self-forcing video generation extends a short-horizon video model to longer rollouts by repeatedly feeding generated content back in as context. This scaling path immediately exposes a systems bottleneck: the key-value (KV) cache grows with rollout length, so longer videos require not only better generation quality but also substantially better memory behavior. We present a comprehensive empirical study of KV-cache compression for self-forcing video generation on a Wan2.1-based Self-Forcing stack. Our study covers 33 quantization and cache-policy variants, 610 prompt-level observations, and 63 benchmark-level summaries across two evaluation settings: MovieGen for single-shot 10-second generation and StoryEval for longer narrative-style stability. We jointly evaluate peak VRAM, runtime, realized compression ratio, VBench imaging quality, BF16-referenced fidelity (SSIM, LPIPS, PSNR), and terminal drift. Three findings are robust. First, the strongest practical operating region is a FlowCache-inspired soft-prune INT4 adaptation, which reaches 5.42-5.49x compression while reducing peak VRAM from 19.28 GB to about 11.7 GB with only modest runtime overhead. Second, the highest-fidelity compressed methods, especially PRQ_INT4 and QUAROT_KV_INT4, are not the best deployment choices because they preserve quality at severe runtime or memory cost. Third, nominal compression alone is not sufficient: several methods shrink KV storage but still exceed BF16 peak VRAM because the current integration reconstructs or retains large BF16 buffers during attention and refresh stages. The result is a benchmark harness, analysis workflow, and empirical map of which KV-cache ideas are practical today and which are promising research directions for better memory integration. Code, data products, and the presentation dashboard are available at https://github.com/suraj-ranganath/kv-quant-longhorizon/.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。