{"id":9790,"date":"2026-06-30T03:58:12","date_gmt":"2026-06-29T19:58:12","guid":{"rendered":"https:\/\/infernews.com\/blog\/pytorch-implementation-of-parallel-rollout-approximation\/"},"modified":"2026-06-30T03:58:12","modified_gmt":"2026-06-29T19:58:12","slug":"pytorch-implementation-of-parallel-rollout-approximation","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/pytorch-implementation-of-parallel-rollout-approximation\/","title":{"rendered":"PRA\uff1a\u50cf\u7d20\u7d1a\u81ea\u56de\u6b78\u751f\u5716\u7684\u65b0\u8def\u7dda"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/06\/result.jpg\" alt=\"result\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">PRA \u662f\u4e00\u500b\u4ee5 PyTorch \u5be6\u4f5c\u7684\u5f71\u50cf\u751f\u6210\u7814\u7a76\u9805\u76ee\uff0c\u5c6c\u65bc class-conditional pixel-space autoregressive image generation \u6a21\u578b\u8207\u8a13\u7df4\u6846\u67b6\u3002\u5b83\u8981\u89e3\u6c7a\u7684\u554f\u984c\uff0c\u662f\u76f4\u63a5\u5728\u50cf\u7d20\u7a7a\u9593\u9010\u6b65\u751f\u6210\u5716\u7247\u6642\uff0c\u55ae\u6b65\u8aa4\u5dee\u5927\u3001\u800c\u4e14 teacher-forced training \u8207\u63a8\u7406\u6d41\u7a0b\u4e0d\u4e00\u81f4\uff0c\u4ee4\u8aa4\u5dee\u4e00\u8def\u7d2f\u7a4d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73fe\u6709 pixel-space continuous-token autoregressive \u505a\u6cd5\uff0c\u591a\u6578\u76f4\u63a5\u9810\u6e2c\u9ad8\u7dad\u50cf\u7d20 patch\uff0c\u6216\u7528 x-prediction\u3001input noise injection \u6e1b\u8f15\u8aa4\u5dee\uff0c\u4f46\u6539\u5584\u6709\u9650\uff1bexact rollout training \u96d6\u7136\u66f4\u8cbc\u8fd1\u63a8\u7406\uff0c\u4ee3\u50f9\u53c8\u592a\u9ad8\u3002PRA \u7684\u53d6\u5411\u662f\u52a0\u5165 Parallel Rollout Approximation (PRA)\uff1a\u5148\u751f\u6210\u4f4e\u7dad intermediate states\uff0c\u518d\u7d93 pixel decoder \u6620\u5c04\u56de pixel-space tokens\uff0c\u540c\u6642\u7528\u8fd1\u4f3c\u63a8\u7406\u6642\u7684 pixel-feedback \u65b9\u5f0f\u4fdd\u7559\u5e73\u884c\u8a13\u7df4\u6548\u7387\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u500b\u8a2d\u8a08\u7684\u91cd\u9ede\uff0c\u4e0d\u662f\u55ae\u7d14\u8ffd\u6c42\u66f4\u5927\u6a21\u578b\uff0c\u800c\u662f\u91cd\u65b0\u8655\u7406\u300c\u8a13\u7df4\u898b\u5230\u7684\u8f38\u5165\u300d\u8207\u300c\u751f\u6210\u6642\u771f\u6b63\u6536\u5230\u7684\u56de\u994b\u300d\u4e4b\u9593\u7684\u843d\u5dee\u3002\u8ad6\u6587\u8cc7\u6599\u986f\u793a\uff0c\u5b83\u5728 ImageNet-1K 256\u00d7256 \u7684 class-conditional \u751f\u6210\u4e0a\uff0cPRA-S 135M \u53c3\u6578\u5df2\u505a\u5230 FID 2.58\uff0c\u512a\u65bc\u5148\u524d billion-scale pixel-space AR \u7684 3.60\uff1bPRA-L 511M \u9032\u4e00\u6b65\u5230 1.94\uff0c\u5b9a\u4f4d\u5f88\u6e05\u695a\uff0c\u5c31\u662f\u885d\u7740 pixel-space AR \u7684 SOTA \u800c\u4f86\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u90e8\u7f72\u7406\u89e3\u4e0a\uff0c\u5132\u5b58\u5eab\u5df2\u63d0\u4f9b environment.yml\u3001requirements.txt\u3001\u9810\u8a13\u7df4\u6b0a\u91cd\u8207 sample_ddp.py\uff0c\u4ee3\u8868\u4f5c\u8005\u9810\u8a2d\u4f60\u6703\u7528\u591a\u5361\u5206\u6563\u5f0f\u62bd\u6a23\u8207\u8a55\u6e2c\u3002\u8a55\u4f30\u6307\u6a19\u5305\u62ec FID\u3001Inception Score\u3001precision\u3001recall\uff0c\u53e6\u6709 ImageNet classification probing accuracy \u4f5c\u70ba\u751f\u6210\u4ee5\u5916\u7684\u88dc\u5145\u89c0\u5bdf\uff0c\u8868\u793a\u4f5c\u8005\u4e5f\u5728\u6e2c\u8a66\u8868\u5fb5\u80fd\u529b\uff0c\u800c\u4e0d\u53ea\u770b\u51fa\u5716\u6f02\u4eae\u8207\u5426\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u9805\u76ee\u985e\u578b<\/strong>\uff1a\u7814\u7a76\u539f\u578b\u517c\u6a21\u578b\u5be6\u4f5c\uff0c\u96c6\u4e2d\u5c55\u793a PRA \u8a13\u7df4\u8207\u53d6\u6a23\u6d41\u7a0b<\/li>\n<li><strong>\u76f8\u95dc\u6a21\u578b<\/strong>\uff1aPRA-S\u3001PRA-B\u3001PRA-L\uff0c\u53c3\u6578\u91cf\u7d04 135M\u3001250M\u3001511M<\/li>\n<li><strong>\u4e3b\u8981\u5dee\u7570<\/strong>\uff1a\u4e0d\u7528\u96e2\u6563 tokenizer\uff0c\u7dad\u6301 pixel-in\u3001pixel-out AR \u4ecb\u9762<\/li>\n<li><strong>\u8f03\u9069\u5408\u60c5\u5883<\/strong>\uff1a\u7814\u7a76 pixel-space AR\u3001\u6bd4\u8f03\u751f\u6210\u6307\u6a19\u3001\u91cd\u73fe ImageNet \u985e\u689d\u4ef6\u751f\u5716\u7d50\u679c<\/li>\n<li><strong>\u9700\u8981\u7559\u610f<\/strong>\uff1a\u76ee\u524d\u516c\u958b\u5167\u5bb9\u504f\u5411\u7814\u7a76\u91cd\u73fe\uff0c\u4e0d\u662f\u5373\u88dd\u5373\u7528\u7684\u7d42\u7aef\u61c9\u7528\u5de5\u5177<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u53d7\u76ca\u6700\u5927\u7684\uff0c\u6703\u662f\u505a\u5f71\u50cf\u751f\u6210\u7814\u7a76\u7684\u5718\u968a\u3001\u60f3\u6bd4\u8f03 autoregressive \u8207 diffusion \u8def\u7dda\u7684\u4eba\uff0c\u4ee5\u53ca\u8981\u7814\u7a76\u50cf\u7d20\u7a7a\u9593\u5efa\u6a21\u53d6\u6368\u7684\u5b78\u8853\u9805\u76ee\u3002\u5c0d\u4e00\u822c\u958b\u767c\u8005\u4f86\u8aaa\uff0c\u9019\u500b\u9805\u76ee\u8f03\u50cf\u9ad8\u6c34\u6e96\u5be6\u9a57\u5e73\u53f0\uff1b\u6709\u73fe\u6210 checkpoint \u548c\u8a55\u6e2c\u6d41\u7a0b\u662f\u512a\u9ede\uff0c\u4f46 CUDA\u3001PyTorch \u8207 FlashAttention \u76f8\u5bb9\u6027\u4ecd\u9700\u81ea\u884c\u8655\u7406\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/github.com\/MangataX\/PRA\" rel=\"noopener noreferrer\" target=\"_blank\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/arxiv.org\/pdf\/2606.27978\" rel=\"noopener noreferrer\" target=\"_blank\"><strong>Paper<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>PRA \u662f\u4e00\u500b PyTorch \u5f71\u50cf\u751f\u6210\u7814\u7a76\u9805\u76ee\uff0c\u4e3b\u6253 pixel-space autoregressive \u751f\u6210\u3002\u5b83\u5617\u8a66\u7528\u8f03\u7d30\u6a21\u578b\u898f\u6a21\uff0c\u63db\u5230\u66f4\u4f4e FID \u8868\u73fe\u3002<\/p>\n","protected":false},"author":8,"featured_media":9789,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[30,152,157,127,195],"tags":[],"class_list":["post-9790","post","type-post","status-publish","format-standard","hentry","category-image","category-python","category-157","category-127","category-195"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9790","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=9790"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9790\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/9789"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=9790"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=9790"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=9790"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}