{"id":11399,"date":"2026-08-25T08:07:09","date_gmt":"2026-08-25T00:07:09","guid":{"rendered":"https:\/\/infernews.com\/blog\/training-free-sparse-attention-for-video-generation-and-world-models\/"},"modified":"2026-08-25T08:07:09","modified_gmt":"2026-08-25T00:07:09","slug":"training-free-sparse-attention-for-video-generation-and-world-models","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/training-free-sparse-attention-for-video-generation-and-world-models\/","title":{"rendered":"SparsePR \u8b93\u5f71\u7247\u751f\u6210\u4ee5\u7a00\u758f\u6ce8\u610f\u529b\u52a0\u901f\uff0c\u6bcb\u9808\u91cd\u65b0\u8a13\u7df4"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/08\/pasted-d94b860732c4.jpg\" alt=\"Repository image for PardisTaghavi\/SparsePR\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f71\u7247\u751f\u6210\u6700\u5403\u8cc7\u6e90\u7684\u90e8\u5206\uff0c\u5f80\u5f80\u4e0d\u662f\u63d0\u793a\u8a5e\u6216\u8f38\u51fa\u683c\u5f0f\uff0c\u800c\u662f Video Transformer \u88cf\u5927\u91cf Attention \u8a08\u7b97\u3002SparsePR \u5c6c\u65bc training-free sparse attention \u53c3\u8003\u5be6\u4f5c\uff0c\u900f\u904e\u6e1b\u5c11\u4e0d\u5fc5\u8981\u7684 query\u3001key\/value \u4e92\u52d5\uff0c\u52a0\u901f HunyuanVideo-13B\u3001Wan2.2-I2V-A14B\u3001Cosmos-Predict2.5-14B \u53ca Cosmos3-Nano-16B\uff0c\u6bcb\u9808\u984d\u5916\u8a13\u7df4\u6a21\u578b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u6c92\u6709\u55ae\u7d14\u6309\u6ce8\u610f\u529b\u96c6\u4e2d\u7a0b\u5ea6\u522a\u8d70\u5340\u584a\uff0c\u800c\u662f\u4ee5 Response-Coupled Partitioning \u6309\u76ee\u524d\u56de\u61c9\u5206\u7d44\uff0c\u518d\u7528\u5c11\u91cf exact probe rows \u914d\u5408 Probe-Fitted Residual Reconstruction\uff0c\u88dc\u56de\u7a00\u758f\u8a08\u7b97\u907a\u6f0f\u7684\u8f38\u51fa\u3002\u4f7f\u7528\u8005\u53ef\u900f\u904e\u540c\u4e00\u500b\u4ecb\u9762\uff0c\u5728 dense baseline \u8207 SparsePR \u4e4b\u9593\u5207\u63db\uff0c\u76f4\u63a5\u6bd4\u8f03\u5f71\u7247\u8cea\u7d20\u3001\u901f\u5ea6\u548c\u986f\u5b58\u53d6\u6368\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u57f7\u884c pair density \u7d04 21.9% \u81f3 26.0%<\/li>\n<li>\u7aef\u5230\u7aef\u901f\u5ea6\u63d0\u5347\u7d04 1.48 \u81f3 2.61 \u500d<\/li>\n<li>\u6db5\u84cb\u6587\u5b57\u8f49\u5f71\u7247\u3001\u5716\u50cf\u8f49\u5f71\u7247\u53ca image-to-world \u60c5\u5883<\/li>\n<li>\u5728 VBench \u53ca PBench \u9032\u884c\u8a55\u4f30\uff0c\u90e8\u5206 Cosmos-Predict2.5-14B \u7d50\u679c\u9054 40.33 dB<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u9805\u76ee\u8981\u6c42 Linux\uff0c\u4e26\u5efa\u8b70\u4f7f\u7528 NVIDIA H100\uff1bHunyuanVideo\u3001Wan2.2\u3001Cosmos-Predict2.5 \u8207 Cosmos3 \u9700\u8981\u5206\u958b\u7684 CUDA 12.8 wheel \u74b0\u5883\uff0c\u539f\u56e0\u662f Cosmos3 \u4f9d\u8cf4\u8f03\u65b0\u7684 Diffusers \u548c Transformers\u3002\u53ef\u9078\u7684 fused CUDA kernels \u6709\u52a9\u9032\u4e00\u6b65\u57f7\u884c\uff0c\u4f46\u5b89\u88dd\u9580\u6abb\u660e\u986f\u9ad8\u65bc\u4e00\u822c\u5f71\u7247\u751f\u6210\u5de5\u5177\uff0c\u6a21\u578b checkpoint \u4ea6\u8981\u5f9e\u5b98\u65b9 Hugging Face \u5132\u5b58\u5eab\u53d6\u5f97\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">SparsePR \u9069\u5408\u7814\u7a76\u5f71\u7247\u751f\u6210\u6548\u7387\u3001\u5efa\u7acb\u9ad8\u7aef GPU \u63a8\u7406\u57fa\u6e96\uff0c\u6216\u9700\u8981\u5728\u4fdd\u7559\u756b\u9762\u8cea\u7d20\u4e0b\u6e1b\u5c11\u8a08\u7b97\u91cf\u7684\u5718\u968a\u3002\u5b83\u76ee\u524d\u4ecd\u662f\u53c3\u8003\u5be6\u4f5c\uff0c\u652f\u63f4\u6a21\u578b\u548c\u786c\u4ef6\u74b0\u5883\u8f03\u6709\u9650\uff1b\u5c0d\u53ea\u6709\u6d88\u8cbb\u7d1a GPU\u3001\u53ea\u60f3\u5feb\u901f\u8a66\u73a9\u5f71\u7247\u751f\u6210\u7684\u4f7f\u7528\u8005\uff0c\u6210\u672c\u8207\u74b0\u5883\u914d\u7f6e\u53ef\u80fd\u62b5\u92b7\u52a0\u901f\u5e36\u4f86\u7684\u597d\u8655\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/pardistaghavi.github.io\/SparsePR-website\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/PardisTaghavi\/SparsePR\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>SparsePR \u5c07\u7a00\u758f\u6ce8\u610f\u529b\u5e36\u5165\u56db\u6b3e\u5f71\u7247\u6a21\u578b\uff0c\u5728\u7dad\u6301\u751f\u6210\u8cea\u7d20\u7684\u540c\u6642\uff0c\u628a\u57f7\u884c\u901f\u5ea6\u63d0\u5347\u6700\u591a 2.61 \u500d\u3002<\/p>\n","protected":false},"author":8,"featured_media":11398,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","footnotes":""},"categories":[133,179,31,30,168,127,197],"tags":[],"class_list":["post-11399","post","type-post","status-publish","format-standard","hentry","category-133","category-nvidia","category-video","category-image","category-linux","category-127","category-framework"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11399","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=11399"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11399\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/11398"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=11399"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=11399"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=11399"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}