{"id":18909,"date":"2024-02-29T14:20:23","date_gmt":"2024-02-29T06:20:23","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=18909"},"modified":"2024-02-29T14:20:23","modified_gmt":"2024-02-29T06:20:23","slug":"%e6%83%b3%e8%ae%ad%e7%bb%83%e7%b1%bbsora%e6%a8%a1%e5%9e%8b%e5%90%97%ef%bc%9f%e5%b0%a4%e6%b4%8b%e5%9b%a2%e9%98%9fopendit%e5%ae%9e%e7%8e%b080%e5%8a%a0%e9%80%9f","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2024\/02\/29\/18909\/","title":{"rendered":"\u60f3\u8bad\u7ec3\u7c7bSora\u6a21\u578b\u5417\uff1f\u5c24\u6d0b\u56e2\u961fOpenDiT\u5b9e\u73b080%\u52a0\u901f"},"content":{"rendered":"

\u4f5c\u4e3a 2024 \u5f00\u5e74\u738b\u70b8\uff0cSora \u7684\u51fa\u73b0\u6811\u7acb\u4e86\u4e00\u4e2a\u5168\u65b0\u7684\u8ffd\u8d76\u76ee\u6807\uff0c\u6bcf\u4e2a\u6587\u751f\u89c6\u9891\u7684\u7814\u7a76\u8005\u90fd\u60f3\u5728\u6700\u77ed\u7684\u65f6\u95f4\u5185\u590d\u73b0 Sora \u7684\u6548\u679c\u3002<\/p>\n

\u6839\u636e OpenAI \u62ab\u9732\u7684\u6280\u672f\u62a5\u544a\uff0cSora \u7684\u6838\u5fc3\u6280\u672f\u70b9\u4e4b\u4e00\u662f\u5c06\u89c6\u89c9\u6570\u636e\u8f6c\u5316\u4e3a patch \u7684\u7edf\u4e00\u8868\u5f81\u5f62\u5f0f\uff0c\u5e76\u901a\u8fc7 Transformer \u548c\u6269\u6563\u6a21\u578b\u7ed3\u5408\uff0c\u5c55\u73b0\u4e86\u5353\u8d8a\u7684\u6269\u5c55\uff08scale\uff09\u7279\u6027\u3002\u5728\u62a5\u544a\u516c\u5e03\u540e\uff0cSora \u6838\u5fc3\u7814\u53d1\u6210\u5458 William Peebles \u548c\u7ebd\u7ea6\u5927\u5b66\u8ba1\u7b97\u673a\u79d1\u5b66\u52a9\u7406\u6559\u6388\u8c22\u8d5b\u5b81\u5408\u8457\u7684\u8bba\u6587\u300aScalable Diffusion Models with Transformers\u300b\u5c31\u6210\u4e86\u4f17\u591a\u7814\u7a76\u8005\u5173\u6ce8\u7684\u91cd\u70b9\u3002\u5927\u5bb6\u5e0c\u671b\u80fd\u4ee5\u8bba\u6587\u4e2d\u63d0\u51fa\u7684 DiT \u67b6\u6784\u4e3a\u7a81\u7834\u53e3\uff0c\u63a2\u7d22\u590d\u73b0 Sora \u7684\u53ef\u884c\u8def\u5f84\u3002<\/p>\n

\u6700\u8fd1\uff0c\u65b0\u52a0\u5761\u56fd\u7acb\u5927\u5b66\u5c24\u6d0b\u56e2\u961f\u5f00\u6e90\u7684<\/span>\u4e00\u4e2a\u540d\u4e3a OpenDiT \u7684\u9879\u76ee\u4e3a\u8bad\u7ec3\u548c\u90e8\u7f72 DiT \u6a21\u578b\u6253\u5f00\u4e86\u65b0\u601d\u8def\u3002<\/p>\n

OpenDiT \u662f\u4e00\u4e2a\u6613\u4e8e\u4f7f\u7528\u3001\u5feb\u901f\u4e14\u5185\u5b58\u9ad8\u6548\u7684\u7cfb\u7edf\uff0c\u4e13\u95e8\u7528\u4e8e\u63d0\u9ad8 DiT \u5e94\u7528\u7a0b\u5e8f\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u6548\u7387\uff0c\u5305\u62ec\u6587\u672c\u5230\u89c6\u9891\u751f\u6210\u548c\u6587\u672c\u5230\u56fe\u50cf\u751f\u6210\u3002<\/p>\n

<\/div>\n

 <\/p>\n

\u9879\u76ee\u5730\u5740\uff1ahttps:\/\/github.com\/NUS-HPC-AI-Lab\/OpenDiT<\/span><\/p>\n

<\/div>\n

OpenDiT \u65b9\u6cd5\u4ecb\u7ecd<\/span><\/p>\n

OpenDiT \u63d0\u4f9b\u7531 Colossal-AI \u652f\u6301\u7684 Diffusion Transformer (DiT) \u7684\u9ad8\u6027\u80fd\u5b9e\u73b0\u3002<\/span>\u5728\u8bad\u7ec3\u65f6\uff0c\u89c6\u9891\u548c\u6761\u4ef6\u4fe1\u606f\u5206\u522b\u88ab\u8f93\u5165\u5230\u76f8\u5e94\u7684\u7f16\u7801\u5668\u4e2d\uff0c\u4f5c\u4e3aDiT\u6a21\u578b\u7684\u8f93\u5165\u3002<\/span>\u968f\u540e\uff0c\u901a\u8fc7\u6269\u6563\u65b9\u6cd5\u8fdb\u884c\u8bad\u7ec3\u548c\u53c2\u6570\u66f4\u65b0\uff0c\u6700\u7ec8\u5c06\u66f4\u65b0\u540e\u7684\u53c2\u6570\u540c\u6b65\u81f3EMA\uff08Exponential Moving Average\uff09\u6a21\u578b\u3002<\/span>\u63a8\u7406\u9636\u6bb5\u5219\u76f4\u63a5\u4f7f\u7528EMA\u6a21\u578b\uff0c\u5c06\u6761\u4ef6\u4fe1\u606f\u4f5c\u4e3a\u8f93\u5165\uff0c\u4ece\u800c\u751f\u6210\u5bf9\u5e94\u7684\u7ed3\u679c\u3002<\/span><\/p>\n

\"\u56fe\u6e90\uff1ahttps:\/\/www.zhihu.com\/people\/berkeley-you-yang\"<\/div>\n

OpenDiT \u5229\u7528\u4e86 ZeRO \u5e76\u884c\u7b56\u7565\uff0c\u5c06 DiT \u6a21\u578b\u53c2\u6570\u5206\u5e03\u5230\u591a\u53f0\u673a\u5668\u4e0a\uff0c\u521d\u6b65\u964d\u4f4e\u4e86\u663e\u5b58\u538b\u529b\u3002\u4e3a\u4e86\u53d6\u5f97\u66f4\u597d\u7684\u6027\u80fd\u4e0e\u7cbe\u5ea6\u5e73\u8861\uff0cOpenDiT \u8fd8\u91c7\u7528\u4e86\u6df7\u5408\u7cbe\u5ea6\u7684\u8bad\u7ec3\u7b56\u7565\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6a21\u578b\u53c2\u6570\u548c\u4f18\u5316\u5668\u4f7f\u7528 float32 \u8fdb\u884c\u5b58\u50a8\uff0c\u4ee5\u786e\u4fdd\u66f4\u65b0\u7684\u51c6\u786e\u6027\u3002\u5728\u6a21\u578b\u8ba1\u7b97\u7684\u8fc7\u7a0b\u4e2d\uff0c\u7814\u7a76\u56e2\u961f\u4e3a DiT \u6a21\u578b\u8bbe\u8ba1\u4e86 float16 \u548c float32 \u7684\u6df7\u5408\u7cbe\u5ea6\u65b9\u6cd5\uff0c\u4ee5\u5728\u7ef4\u6301\u6a21\u578b\u7cbe\u5ea6\u7684\u540c\u65f6\u52a0\u901f\u8ba1\u7b97\u8fc7\u7a0b\u3002<\/p>\n

DiT \u6a21\u578b\u4e2d\u4f7f\u7528\u7684 EMA \u65b9\u6cd5\u662f\u4e00\u79cd\u7528\u4e8e\u5e73\u6ed1\u6a21\u578b\u53c2\u6570\u66f4\u65b0\u7684\u7b56\u7565\uff0c\u53ef\u4ee5\u6709\u6548\u63d0\u9ad8\u6a21\u578b\u7684\u7a33\u5b9a\u6027\u548c\u6cdb\u5316\u80fd\u529b\u3002\u4f46\u662f\u4f1a\u989d\u5916\u4ea7\u751f\u4e00\u4efd\u53c2\u6570\u7684\u62f7\u8d1d\uff0c\u589e\u52a0\u4e86\u663e\u5b58\u7684\u8d1f\u62c5\u3002\u4e3a\u4e86\u8fdb\u4e00\u6b65\u964d\u4f4e\u8fd9\u90e8\u5206\u663e\u5b58\uff0c\u7814\u7a76\u56e2\u961f\u5c06 EMA \u6a21\u578b\u5206\u7247\uff0c\u5e76\u5206\u522b\u5b58\u50a8\u5728\u4e0d\u540c\u7684 GPU \u4e0a\u3002\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6bcf\u4e2a GPU \u53ea\u9700\u8ba1\u7b97\u548c\u5b58\u50a8\u81ea\u5df1\u8d1f\u8d23\u7684\u90e8\u5206 EMA \u6a21\u578b\u53c2\u6570\uff0c\u5e76\u5728\u6bcf\u6b21 step \u540e\u7b49\u5f85 ZeRO \u5b8c\u6210\u66f4\u65b0\u540e\u8fdb\u884c\u540c\u6b65\u66f4\u65b0\u3002<\/p>\n

FastSeq<\/span><\/p>\n

\u5728 DiT \u7b49\u89c6\u89c9\u751f\u6210\u6a21\u578b\u9886\u57df\uff0c\u5e8f\u5217\u5e76\u884c\u6027\u5bf9\u4e8e\u6709\u6548\u7684\u957f\u5e8f\u5217\u8bad\u7ec3\u548c\u4f4e\u5ef6\u8fdf\u63a8\u7406\u662f\u5fc5\u4e0d\u53ef\u5c11\u7684\u3002<\/p>\n

 <\/p>\n

\u7136\u800c\uff0cDeepSpeed-Ulysses\u3001Megatron-LM Sequence Parallelism \u7b49\u73b0\u6709\u65b9\u6cd5\u5728\u5e94\u7528\u4e8e\u6b64\u7c7b\u4efb\u52a1\u65f6\u9762\u4e34\u5c40\u9650\u6027 \u2014\u2014 \u8981\u4e48\u662f\u5f15\u5165\u8fc7\u591a\u7684\u5e8f\u5217\u901a\u4fe1\uff0c\u8981\u4e48\u662f\u5728\u5904\u7406\u5c0f\u89c4\u6a21\u5e8f\u5217\u5e76\u884c\u65f6\u7f3a\u4e4f\u6548\u7387\u3002<\/p>\n

 <\/p>\n

\u4e3a\u6b64\uff0c\u7814\u7a76\u56e2\u961f\u63d0\u51fa\u4e86 FastSeq\uff0c\u4e00\u79cd\u9002\u7528\u4e8e\u5927\u5e8f\u5217\u548c\u5c0f\u89c4\u6a21\u5e76\u884c\u7684\u65b0\u578b\u5e8f\u5217\u5e76\u884c\u3002FastSeq \u901a\u8fc7\u4e3a\u6bcf\u4e2a transformer \u5c42\u4ec5\u4f7f\u7528\u4e24\u4e2a\u901a\u4fe1\u8fd0\u7b97\u7b26\u6765\u6700\u5c0f\u5316\u5e8f\u5217\u901a\u4fe1\uff0c\u5229\u7528 AllGather \u6765\u63d0\u9ad8\u901a\u4fe1\u6548\u7387\uff0c\u5e76\u7b56\u7565\u6027\u5730\u91c7\u7528\u5f02\u6b65 ring \u5c06 AllGather \u901a\u4fe1\u4e0e qkv \u8ba1\u7b97\u91cd\u53e0\uff0c\u8fdb\u4e00\u6b65\u4f18\u5316\u6027\u80fd\u3002<\/p>\n

 <\/p>\n

<\/div>\n

 <\/p>\n

<\/span><\/div>\n

 <\/p>\n

 <\/p>\n

\u7b97\u5b50\u4f18\u5316<\/span><\/p>\n

\u5728 DiT \u6a21\u578b\u4e2d\u5f15\u5165 adaLN \u6a21\u5757\u5c06\u6761\u4ef6\u4fe1\u606f\u878d\u5165\u89c6\u89c9\u5185\u5bb9\uff0c\u867d\u7136\u8fd9\u4e00\u64cd\u4f5c\u5bf9\u6a21\u578b\u7684\u6027\u80fd\u63d0\u5347\u81f3\u5173\u91cd\u8981\uff0c\u4f46\u4e5f\u5e26\u6765\u4e86\u5927\u91cf\u7684\u9010\u5143\u7d20\u64cd\u4f5c\uff0c\u5e76\u4e14\u5728\u6a21\u578b\u4e2d\u88ab\u9891\u7e41\u8c03\u7528\uff0c\u964d\u4f4e\u4e86\u6574\u4f53\u7684\u8ba1\u7b97\u6548\u7387\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u7814\u7a76\u56e2\u961f\u63d0\u51fa\u4e86\u9ad8\u6548\u7684 Fused adaLN Kernel\uff0c\u5c06\u591a\u6b21\u64cd\u4f5c\u5408\u5e76\u6210\u4e00\u6b21\uff0c\u4ece\u800c\u589e\u52a0\u4e86\u8ba1\u7b97\u6548\u7387\uff0c\u5e76\u4e14\u51cf\u5c11\u4e86\u89c6\u89c9\u4fe1\u606f\u7684 I\/O \u6d88\u8017\u3002<\/p>\n

\"\u56fe\u6e90\uff1ahttps:\/\/www.zhihu.com\/people\/berkeley-you-yang\"<\/div>\n

\u7b80\u5355\u6765\u8bf4\uff0cOpenDiT \u5177\u6709\u4ee5\u4e0b\u6027\u80fd\u4f18\u52bf\uff1a<\/p>\n

1\u3001\u5728 GPU \u4e0a\u52a0\u901f\u9ad8\u8fbe 80%\uff0c50%\u7684\u5185\u5b58\u8282\u7701<\/p>\n

    \n
  • \u8bbe\u8ba1\u4e86\u9ad8\u6548\u7684\u7b97\u5b50\uff0c\u5305\u62ec\u9488\u5bf9DiT\u8bbe\u8ba1\u7684 Fused AdaLN\uff0c\u4ee5\u53ca FlashAttention\u3001Fused Layernorm \u548cHybridAdam\u3002<\/li>\n
  • \u91c7\u7528\u6df7\u5408\u5e76\u884c\u65b9\u6cd5\uff0c\u5305\u62ec ZeRO\u3001Gemini \u548c DDP\u3002\u5bf9 ema \u6a21\u578b\u8fdb\u884c\u5206\u7247\u4e5f\u8fdb\u4e00\u6b65\u964d\u4f4e\u4e86\u5185\u5b58\u6210\u672c\u3002<\/li>\n<\/ul>\n

    2\u3001FastSeq\uff1a\u4e00\u79cd\u65b0\u9896\u7684\u5e8f\u5217\u5e76\u884c\u65b9\u6cd5<\/p>\n

      \n
    • \u4e13\u4e3a\u7c7b\u4f3c DiT \u7684\u5de5\u4f5c\u8d1f\u8f7d\u800c\u8bbe\u8ba1\uff0c\u5728\u8fd9\u4e9b\u5e94\u7528\u4e2d\uff0c\u5e8f\u5217\u901a\u5e38\u8f83\u957f\uff0c\u4f46\u53c2\u6570\u76f8\u6bd4\u4e8e LLM \u8f83\u5c0f\u3002<\/li>\n
    • \u8282\u70b9\u5185\u5e8f\u5217\u5e76\u884c\u53ef\u8282\u7701\u9ad8\u8fbe 48% \u7684\u901a\u4fe1\u91cf\u3002<\/li>\n
    • \u6253\u7834\u5355\u4e2a GPU \u7684\u5185\u5b58\u9650\u5236\uff0c\u51cf\u5c11\u6574\u4f53\u8bad\u7ec3\u548c\u63a8\u7406\u65f6\u95f4\u3002<\/li>\n<\/ul>\n

      3\u3001\u6613\u4e8e\u4f7f\u7528<\/p>\n

        \n
      • \u53ea\u9700\u51e0\u884c\u4ee3\u7801\u7684\u4fee\u6539\uff0c\u5373\u53ef\u83b7\u5f97\u5de8\u5927\u7684\u6027\u80fd\u63d0\u5347\u3002<\/li>\n
      • \u7528\u6237\u65e0\u9700\u4e86\u89e3\u5206\u5e03\u5f0f\u8bad\u7ec3\u7684\u5b9e\u73b0\u65b9\u5f0f\u3002<\/li>\n<\/ul>\n

        4\u3001\u6587\u672c\u5230\u56fe\u50cf\u548c\u6587\u672c\u5230\u89c6\u9891\u751f\u6210\u5b8c\u6574 pipeline<\/p>\n

          \n
        • \u7814\u7a76\u4eba\u5458\u548c\u5de5\u7a0b\u5e08\u53ef\u4ee5\u8f7b\u677e\u4f7f\u7528 OpenDiT pipeline \u5e76\u5c06\u5176\u5e94\u7528\u4e8e\u5b9e\u9645\u5e94\u7528\uff0c\u800c\u65e0\u9700\u4fee\u6539\u5e76\u884c\u90e8\u5206\u3002<\/li>\n
        • \u7814\u7a76\u56e2\u961f\u901a\u8fc7\u5728 ImageNet \u4e0a\u8fdb\u884c\u6587\u672c\u5230\u56fe\u50cf\u8bad\u7ec3\u6765\u9a8c\u8bc1 OpenDiT \u7684\u51c6\u786e\u6027\uff0c\u5e76\u53d1\u5e03\u4e86\u68c0\u67e5\u70b9\uff08checkpoint\uff09\u3002<\/li>\n<\/ul>\n

          \u5b89\u88c5\u4e0e\u4f7f\u7528<\/span><\/p>\n

          \u8981\u4f7f\u7528 OpenDiT\uff0c\u9996\u5148\u8981\u5b89\u88c5\u5148\u51b3\u6761\u4ef6\uff1a<\/p>\n

            \n
          • Python >= 3.10<\/li>\n
          • PyTorch >= 1.13\uff08\u5efa\u8bae\u4f7f\u7528 >2.0 \u7248\u672c\uff09<\/li>\n
          • CUDA >= 11.6<\/li>\n<\/ul>\n

            \u5efa\u8bae\u4f7f\u7528 Anaconda \u521b\u5efa\u4e00\u4e2a\u65b0\u73af\u5883\uff08Python >= 3.10\uff09\u6765\u8fd0\u884c\u793a\u4f8b\uff1a<\/p>\n

             <\/p>\n

            conda create -n opendit python=3.10 -y<\/code>\r\n<\/code>conda activate opendit<\/code><\/span><\/pre>\n
             <\/p>\n
             <\/p>\n
            \u5b89\u88c5 ColossalAI\uff1a<\/p>\n
             <\/p>\n
            git clone https:\/\/github.com\/hpcaitech\/ColossalAI.gitcd ColossalAI<\/code>\r\n<\/code>git checkout adae123df3badfb15d044bd416f0cf29f250bc86<\/code>\r\n<\/code>pip install -e .<\/code><\/span><\/pre>\n
             <\/p>\n
             <\/p>\n
            \u5b89\u88c5 OpenDiT\uff1a<\/p>\n
             <\/p>\n
            git clone https:\/\/github.com\/oahzxl\/OpenDiTcd OpenDiT<\/code>\r\n<\/code>pip install -e .<\/code><\/span><\/pre>\n
             <\/p>\n
             <\/p>\n
            \uff08\u53ef\u9009\u4f46\u63a8\u8350\uff09\u5b89\u88c5\u5e93\u4ee5\u52a0\u5feb\u8bad\u7ec3\u548c\u63a8\u7406\u901f\u5ea6\uff1a<\/p>\n
             <\/p>\n
            # Install Triton for fused adaln kernel<\/code>\r\n<\/code>pip install triton<\/code>\r\n<\/code># Install FlashAttention<\/code>\r\n<\/code>pip install flash-attn<\/code>\r\n<\/code># Install apex for fused layernorm kernel<\/code>\r\n<\/code>git clone https:\/\/github.com\/NVIDIA\/apex.gitcd apex<\/code>\r\n<\/code>git checkout 741bdf50825a97664db08574981962d66436d16a<\/code>\r\n<\/code>pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --config-settings \"--build-option=--cpp_ext\" --config-settings \"--build-option=--cuda_ext\" .\/--global-option=\"--cuda_ext\" --global-option=\"--cpp_ext\"<\/code><\/span><\/pre>\n
             <\/p>\n
             <\/p>\n
            \u56fe\u50cf\u751f\u6210<\/span><\/p>\n
            \u4f60\u53ef\u4ee5\u901a\u8fc7\u6267\u884c\u4ee5\u4e0b\u547d\u4ee4\u6765\u8bad\u7ec3 DiT \u6a21\u578b\uff1a<\/p>\n
             <\/p>\n
            # Use script<\/code>\r\n<\/code>bash train_img.sh# Use command line<\/code>\r\n<\/code>torchrun --standalone --nproc_per_node=2 train.py \\<\/code>\r\n<\/code>    --model DiT-XL\/2 \\<\/code>\r\n<\/code>    --batch_size 2<\/code><\/span><\/pre>\n
             <\/p>\n
             <\/p>\n
             <\/p>\n
            \u9ed8\u8ba4\u7981\u7528\u6240\u6709\u52a0\u901f\u65b9\u6cd5\u3002\u4ee5\u4e0b\u662f\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4e00\u4e9b\u5173\u952e\u8981\u7d20\u7684\u8be6\u7ec6\u4fe1\u606f\uff1a<\/p>\n
              \n
            • plugin: \u652f\u6301 ColossalAI\u3001zero2 \u548c ddp \u4f7f\u7528\u7684 booster \u63d2\u4ef6\u3002\u9ed8\u8ba4\u662f zero2\uff0c\u5efa\u8bae\u542f\u7528 zero2\u3002<\/li>\n
            • mixed_ precision\uff1a\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u7684\u6570\u636e\u7c7b\u578b\uff0c\u9ed8\u8ba4\u662f fp16\u3002<\/li>\n
            • grad_checkpoint: \u662f\u5426\u542f\u7528\u68af\u5ea6\u68c0\u67e5\u70b9\u3002\u8fd9\u8282\u7701\u4e86\u8bad\u7ec3\u8fc7\u7a0b\u7684\u5185\u5b58\u6210\u672c\u3002\u9ed8\u8ba4\u503c\u4e3a False\u3002\u5efa\u8bae\u5728\u5185\u5b58\u8db3\u591f\u7684\u60c5\u51b5\u4e0b\u7981\u7528\u5b83\u3002<\/li>\n
            • enable_modulate_kernel: \u662f\u5426\u542f\u7528 modulate \u5185\u6838\u4f18\u5316\uff0c\u4ee5\u52a0\u5feb\u8bad\u7ec3\u8fc7\u7a0b\u3002\u9ed8\u8ba4\u503c\u4e3a False\uff0c\u5efa\u8bae\u5728 GPU < H100 \u65f6\u542f\u7528\u5b83\u3002<\/li>\n
            • enable_layernorm_kernel: \u662f\u5426\u542f\u7528 layernorm \u5185\u6838\u4f18\u5316\uff0c\u4ee5\u52a0\u5feb\u8bad\u7ec3\u8fc7\u7a0b\u3002\u9ed8\u8ba4\u503c\u4e3a False\uff0c\u5efa\u8bae\u542f\u7528\u5b83\u3002<\/li>\n
            • enable_flashattn: \u662f\u5426\u542f\u7528 FlashAttention\uff0c\u4ee5\u52a0\u5feb\u8bad\u7ec3\u8fc7\u7a0b\u3002\u9ed8\u8ba4\u503c\u4e3a False\uff0c\u5efa\u8bae\u542f\u7528\u3002<\/li>\n
            • sequence_parallel_size\uff1a\u5e8f\u5217\u5e76\u884c\u5ea6\u5927\u5c0f\u3002\u5f53\u8bbe\u7f6e\u503c > 1 \u65f6\u5c06\u542f\u7528\u5e8f\u5217\u5e76\u884c\u3002\u9ed8\u8ba4\u503c\u4e3a 1\uff0c\u5982\u679c\u5185\u5b58\u8db3\u591f\uff0c\u5efa\u8bae\u7981\u7528\u5b83\u3002<\/li>\n<\/ul>\n
              \u5982\u679c\u4f60\u60f3\u4f7f\u7528 DiT \u6a21\u578b\u8fdb\u884c\u63a8\u7406\uff0c\u53ef\u4ee5\u8fd0\u884c\u5982\u4e0b\u4ee3\u7801\uff0c\u9700\u8981\u5c06\u68c0\u67e5\u70b9\u8def\u5f84\u66ff\u6362\u4e3a\u4f60\u81ea\u5df1\u8bad\u7ec3\u7684\u6a21\u578b\u3002<\/p>\n
               <\/p>\n
              # Use script<\/code>\r\n<\/code>bash sample_img.sh# Use command line<\/code>\r\n<\/code>python sample.py --model DiT-XL\/2 --image_size 256 --ckpt .\/model.pt<\/code><\/span><\/pre>\n
               <\/p>\n
              \u89c6\u9891\u751f\u6210<\/span><\/p>\n
               <\/p>\n
              \u4f60\u53ef\u4ee5\u901a\u8fc7\u6267\u884c\u4ee5\u4e0b\u547d\u4ee4\u6765\u8bad\u7ec3\u89c6\u9891 DiT \u6a21\u578b\uff1a<\/p>\n
               <\/p>\n
              # train with scipt<\/code>\r\n<\/code>bash train_video.sh# train with command line<\/code>\r\n<\/code>torchrun --standalone --nproc_per_node=2 train.py \\<\/code>\r\n<\/code>    --model vDiT-XL\/222 \\<\/code>\r\n<\/code>    --use_video \\<\/code>\r\n<\/code>    --data_path .\/videos\/demo.csv \\<\/code>\r\n<\/code>    --batch_size 1 \\<\/code>\r\n<\/code>    --num_frames 16 \\<\/code>\r\n<\/code>    --image_size 256 \\<\/code>\r\n<\/code>    --frame_interval 3<\/code>\r\n<\/code># preprocess<\/code>\r\n<\/code># our code read video from csv as the demo shows<\/code>\r\n<\/code># we provide a code to transfer ucf101 to csv format<\/code>\r\n<\/code>python preprocess.py<\/code><\/span><\/pre>\n
               <\/p>\n
               <\/p>\n
              \u4f7f\u7528 DiT \u6a21\u578b\u6267\u884c\u89c6\u9891\u63a8\u7406\u7684\u4ee3\u7801\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n
               <\/p>\n
              # Use script<\/code>\r\n<\/code>bash sample_video.sh# Use command line<\/code>\r\n<\/code>python sample.py \\<\/code>\r\n<\/code>    --model vDiT-XL\/222 \\<\/code>\r\n<\/code>    --use_video \\<\/code>\r\n<\/code>    --ckpt ckpt_path \\<\/code>\r\n<\/code>    --num_frames 16 \\<\/code>\r\n<\/code>    --image_size 256 \\<\/code>\r\n<\/code>    --frame_interval 3<\/code><\/span><\/pre>\n
               <\/p>\n
               <\/p>\n
              DiT \u590d\u73b0\u7ed3\u679c<\/span><\/p>\n
              \u4e3a\u4e86\u9a8c\u8bc1 OpenDiT \u7684\u51c6\u786e\u6027\uff0c\u7814\u7a76\u56e2\u961f\u4f7f\u7528 OpenDiT \u7684 origin \u65b9\u6cd5\u5bf9 DiT \u8fdb\u884c\u4e86\u8bad\u7ec3\uff0c\u5728 ImageNet \u4e0a\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u6a21\u578b\uff0c\u5728 8xA100 \u4e0a\u6267\u884c 80k step\u3002\u4ee5\u4e0b\u662f\u7ecf\u8fc7\u8bad\u7ec3\u7684 DiT \u751f\u6210\u7684\u4e00\u4e9b\u7ed3\u679c\uff1a<\/p>\n
              <\/div>\n
              \u635f\u5931\u4e5f\u4e0e DiT \u8bba\u6587\u4e2d\u5217\u51fa\u7684\u7ed3\u679c\u4e00\u81f4\uff1a<\/p>\n
              <\/div>\n
              \u8981\u590d\u73b0\u4e0a\u8ff0\u7ed3\u679c\uff0c\u9700\u8981\u66f4\u6539 train_img.py \u4e2d\u7684\u6570\u636e\u96c6\u5e76\u6267\u884c\u4ee5\u4e0b\u547d\u4ee4\uff1a<\/p>\n
               <\/p>\n
              torchrun --standalone --nproc_per_node=8 train.py \\<\/code>\r\n<\/code>    --model DiT-XL\/2 \\<\/code>\r\n<\/code>    --batch_size 180 \\<\/code>\r\n<\/code>    --enable_layernorm_kernel \\<\/code>\r\n<\/code>    --enable_flashattn \\<\/code>\r\n<\/code>    --mixed_precision fp16<\/code><\/span>\r\n\r\n\u6587\u7ae0\u6765\u81ea\uff1a\u65b0\u6d6a\u8d22\u7ecf\u5ba2\u6237\u7aef<\/pre>\n
              <\/div>\n
              <\/g><\/svg><\/i> \"Loading\"<\/p>\n
              <\/div>\n","protected":false},"excerpt":{"rendered":"
              \u4f5c\u4e3a 2024 \u5f00\u5e74\u738b\u70b8\uff0cSora \u7684\u51fa\u73b0\u6811\u7acb\u4e86\u4e00\u4e2a\u5168\u65b0\u7684\u8ffd\u8d76\u76ee\u6807\uff0c\u6bcf\u4e2a\u6587\u751f\u89c6\u9891\u7684\u7814\u7a76\u8005\u90fd\u60f3\u5728\u6700\u77ed\u7684\u65f6\u95f4\u5185\u590d\u73b0 […]<\/p>\n
              <\/div>\n
              <\/g><\/svg><\/i> \"Loading\"<\/p>\n
              <\/div>\n","protected":false},"author":56,"featured_media":18910,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3,23,80],"tags":[1028],"class_list":["post-18909","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-23","category-80","tag-soraopendit80"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/18909","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/56"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=18909"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/18909\/revisions"}],"predecessor-version":[{"id":18918,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/18909\/revisions\/18918"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/18910"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=18909"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=18909"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=18909"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}