{"id":446,"date":"2025-11-25T09:17:26","date_gmt":"2025-11-25T01:17:26","guid":{"rendered":"https:\/\/high-flyer.in.suopu.cc\/?p=446"},"modified":"2025-11-25T09:17:26","modified_gmt":"2025-11-25T01:17:26","slug":"haiscale-%e5%b9%bb%e6%96%b9%e8%90%a4%e7%81%ab%e9%ab%98%e6%80%a7%e8%83%bd%e5%b9%b6%e8%a1%8c%e8%ae%ad%e7%bb%83%e5%b7%a5%e5%85%b7%e5%ba%93","status":"publish","type":"post","link":"https:\/\/high-flyer.in.suopu.cc\/en\/blog\/446\/","title":{"rendered":"haiscale | \u5e7b\u65b9\u8424\u706b\u9ad8\u6027\u80fd\u5e76\u884c\u8bad\u7ec3\u5de5\u5177\u5e93"},"content":{"rendered":"

haiscale (Highflyer AI Scale) \u662f\u4e00\u4e2a\u8f7b\u91cf\u7ea7\u7684\u9ad8\u6027\u80fd\u5e76\u884c\u8bad\u7ec3\u5de5\u5177\u5e93\uff0c\u5176\u6574\u5408\u4e86\u5e7b\u65b9 AI \u591a\u5e74\u7684\u5e76\u884c\u8bad\u7ec3\u7814\u53d1\u4f18\u5316\u7ecf\u9a8c\uff0c\u80fd\u591f\u5e2e\u52a9 PyTorch \u7528\u6237\u66f4\u52a0\u9ad8\u6548\u3001\u4fbf\u6377\u5730\u5728\u5927\u89c4\u6a21\u96c6\u7fa4\u4e0a\u8bad\u7ec3\u6a21\u578b\u3002<\/p>\n

haiscale \u4e2d\u5305\u542b\u4e86\u4ee5\u4e0b\u51e0\u79cd\u5de5\u5177\uff1a<\/p>\n

    \n
  1. haiscale.ddp<\/code>: \u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c\u5de5\u5177\uff0c\u4ee5\u5e7b\u65b9 AI \u81ea\u7814\u7684\u00a0hfreduce<\/a>\u00a0\u901a\u4fe1\u4e3a\u540e\u7aef\uff0c\u76f8\u6bd4\u4e8e NCCL \u80fd\u591f\u83b7\u5f97\u66f4\u597d\u7684\u591a\u5361\u62d3\u5c55\u6027\u80fd\uff1b<\/li>\n
  2. haiscale.fsdp<\/code>: \u6781\u81f4\u4f18\u5316\u00a0Fully Sharded Data Parallel (FSDP)<\/code>\u00a0\u7b97\u6cd5\u7684\u5b9e\u73b0\uff0c\u76f8\u6bd4\u4e8e PyTorch FSDP \u901f\u5ea6\u66f4\u5feb\u3001\u5360\u7528\u663e\u5b58\u66f4\u5c11\uff1b<\/li>\n
  3. haiscale.pipeline<\/code>: \u5206\u5e03\u5f0f\u6d41\u6c34\u7ebf\u5e76\u884c\uff08\u6216\u79f0\u6a21\u578b\u5e76\u884c\uff09\u5de5\u5177\u5305\uff0c\u5305\u542b GPipe, PipeDream \u7b49\u7b97\u6cd5\uff0c\u652f\u6301\u591a\u673a\u591a\u5361\u8bad\u7ec3\uff1b<\/li>\n
  4. haiscale.cpu_offload<\/code>: \u795e\u7ecf\u7f51\u7edc\u6a21\u578b Offload \u5de5\u5177\uff0c\u8282\u7701\u8bad\u7ec3\u5360\u7528\u7684\u663e\u5b58\u3002<\/li>\n<\/ol>\n

    \u4e0b\u56fe\u5c55\u793a\u4e86 haiscale \u4e09\u79cd\u5e76\u884c\u65b9\u5f0f\u7684\u6027\u80fd\uff0c\u5176\u76f8\u6bd4 PyTorch \u5b98\u65b9\u81ea\u5e26\u5de5\u5177\u90fd\u6709\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\uff1a<\/p>\n

    \"bench\"<\/a><\/span><\/p>\n

    \u7528\u4e8e\u6d4b\u8bd5\u7684\u6a21\u578b\u662f\u00a0GPT-2 Medium<\/strong>\uff0c\u76f8\u5173\u4ee3\u7801\u5df2\u5f00\u6e90\u81f3 hfai \u6a21\u578b\u4ed3\u5e93\u3002\u4e0b\u9762\u5c06\u4e3a\u5927\u5bb6\u7b80\u8981\u4ecb\u7ecd\u3002<\/p>\n

    API\u6587\u6863<\/strong>\uff1ahttps:\/\/doc.hfai.high-flyer.cn\/api\/haiscale_ddp.html<\/a><\/p>\n

    \u793a\u4f8b\u6a21\u578b<\/strong>\uff1ahttps:\/\/github.com\/HFAiLab\/hfai-models\/tree\/main\/gpt<\/a><\/p>\n

    \u5b89\u88c5<\/h2>\n

    haiscale \u63d0\u4f9b Python \u63a5\u53e3\uff0c\u901a\u8fc7\u5982\u4e0b\u65b9\u5f0f\u5b89\u88c5\uff1a<\/p>\n

      \n
    1. \u5982\u679c\u8981\u4f7f\u7528 haiscale DDP\uff0c\u9996\u5148\u9700\u8981\u5148\u5b89\u88c5 hfreduce \uff08\u5982\u679c\u4e0d\u9700\u8981\u4f7f\u7528 DDP \u53ef\u8df3\u8fc7\u8fd9\u6b65\uff09\uff1a\n
      \n
      sudo apt install libnuma-dev\r\nsudo apt install libibverbs-dev\r\npip install hfreduce --extra-index-url https:\/\/pypi.hfai.high-flyer.cn\/simple --trusted-host pypi.hfai.high-flyer.cn<\/code><\/pre>\n<\/div>\n<\/li>\n
    2. \u5b89\u88c5 haiscale\uff1a\n
      \n
      pip install haiscale --extra-index-url https:\/\/pypi.hfai.high-flyer.cn\/simple --trusted-host pypi.hfai.high-flyer.cn<\/code><\/pre>\n<\/div>\n<\/li>\n<\/ol>\n
      haiscale.ddp<\/h2>\n
      haiscale.ddp.DistributedDataParallel<\/code>\u00a0(haiscale DDP) \u662f\u4e00\u4e2a\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c\u8bad\u7ec3\u5de5\u5177\uff0c\u4f7f\u7528 hfreduce \u4f5c\u4e3a\u901a\u8baf\u540e\u7aef\uff0c\u53cd\u5411\u4f20\u64ad\u7684\u540c\u65f6\u4f1a\u5f02\u6b65\u5730\u5bf9\u8ba1\u7b97\u597d\u7684\u68af\u5ea6\u505a allreduce\u3002<\/p>\n
      \"ddp\"<\/a><\/span><\/p>\n
      haiscale DDP \u7684\u4f7f\u7528\u65b9\u5f0f\u548c pytorch DDP \u51e0\u4e4e\u76f8\u540c\uff0c\u4ee5\u4e0b\u662f\u4f7f\u7528\u793a\u4f8b\uff1a<\/p>\n
      \n
      from<\/span> haiscale.<\/span>ddp import<\/span> DistributedDataParallel\r\n\r\nmodel =<\/span> MyModel(<\/span>)<\/span>.<\/span>cuda(<\/span>)<\/span>\r\nmodel =<\/span> DistributedDataParallel(<\/span>model)<\/span>\r\noptimizer =<\/span> torch.<\/span>optim.<\/span>Adam(<\/span>model.<\/span>parameters(<\/span>)<\/span>,<\/span> lr=<\/span>0.01<\/span>)<\/span>\r\n\r\n# training ...<\/span>\r\nfor<\/span> step,<\/span> (<\/span>x,<\/span> y)<\/span> in<\/span> enumerate<\/span>(<\/span>dataloader)<\/span>:<\/span>\r\n    optimizer.<\/span>zero_grad(<\/span>)<\/span>\r\n    output =<\/span> model(<\/span>x)<\/span>\r\n    loss_fn(<\/span>y,<\/span> output)<\/span>.<\/span>backward(<\/span>)<\/span>\r\n    optimizer.<\/span>step(<\/span>)<\/span>\r\n\r\n# stop hfreduce<\/span>\r\nmodel.<\/span>reducer.<\/span>stop(<\/span>)<\/span><\/code><\/pre>\n<\/div>\n
      \u5982\u679c\u9700\u8981\u505a\u68af\u5ea6\u7d2f\u52a0\uff0c\u53ef\u4ee5\u4f7f\u7528\u00a0model.no_sync()<\/code>\u00a0\u6765\u51cf\u5c11\u901a\u8baf\u7684\u5f00\u9500\u3002\u6ce8\u610f\u53ea\u9700\u8981\u6700\u540e\u4e00\u6b21\u53cd\u5411\u4f20\u64ad\u65f6\u505a allreduce\uff1a<\/p>\n
      \n
      from<\/span> haiscale.<\/span>ddp import<\/span> DistributedDataParallel\r\n\r\nddp =<\/span> DistributedDataParallel(<\/span>model,<\/span> .<\/span>.<\/span>.<\/span>)<\/span>\r\nwith<\/span> ddp.<\/span>no_sync(<\/span>)<\/span>:<\/span>\r\n    for<\/span> input<\/span> in<\/span> inputs:<\/span>\r\n        ddp(<\/span>input<\/span>)<\/span>.<\/span>backward(<\/span>)<\/span>  # no synchronization, accumulate grads<\/span>\r\nddp(<\/span>another_input)<\/span>.<\/span>backward(<\/span>)<\/span>  # synchronize grads<\/span><\/code><\/pre>\n<\/div>\n
      haiscale.fsdp<\/h2>\n
      Fully Sharded Data Parallel (FSDP) \u662f META \u5728 ZERO-3 \u7684\u57fa\u7840\u4e0a\u63d0\u51fa\u7684\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c\u5de5\u5177\uff0c\u5b83\u628a\u6a21\u578b\u7684\u53c2\u6570\u8fdb\u884c\u5207\u5206\u5e76\u5206\u6563\u5230\u4e0d\u540c\u7684 GPU \u4e0a\uff0c\u6bcf\u5757 GPU \u4e0a\u53ea\u6709\u00a01\/ngpus<\/code>\u00a0\u7684\u53c2\u6570\u3002\u5728\u505a\u524d\u5411\u548c\u53cd\u5411\u4f20\u64ad\u65f6\uff0cFSDP \u4f1a\u5148\u505a allgather \u83b7\u5f97\u5b8c\u6574\u7684\u53c2\u6570\uff0c\u7136\u540e\u5728\u524d\u5411\u548c\u53cd\u5411\u4f20\u64ad\u7ed3\u675f\u540e\u91ca\u653e\u6389\uff0c\u53ea\u4fdd\u7559\u00a01\/ngpus<\/code>\u00a0\u7684\u53c2\u6570\u548c\u68af\u5ea6\u3002FSDP \u901a\u8fc7\u53c2\u6570\u5206\u7247<\/strong>\u7684\u65b9\u5f0f\uff0c\u80fd\u591f\u51cf\u5c11\u6a21\u578b\u53c2\u6570\u3001\u68af\u5ea6\u3001\u4f18\u5316\u5668\u72b6\u6001\u7684\u663e\u5b58\u5360\u7528\uff0c\u5e2e\u52a9\u6211\u4eec\u8bad\u7ec3\u66f4\u5927\u89c4\u6a21\u7684\u6a21\u578b\u3002<\/p>\n
      \"fsdp\"<\/a><\/span><\/p>\n
      haiscale.fsdp.FullyShardedDataParallel<\/code>\u00a0\u7684\u4f7f\u7528\u65b9\u6cd5\u548c DDP \u7c7b\u4f3c\uff0c\u4f46\u4f18\u5316\u5668\u5fc5\u987b\u5728 FSDP \u4e4b\u540e\u521b\u5efa<\/strong>\uff0c\u5e76\u4e14\u4fdd\u5b58\u6a21\u578b\u53c2\u6570\u7684\u65f6\u5019\u9700\u8981\u5148\u8c03\u7528\u00a0summon_full_params<\/code>\u3002\u4ee5\u4e0b\u662f\u4f7f\u7528\u793a\u4f8b\uff1a<\/p>\n
      \n
      from<\/span> haiscale.<\/span>fsdp import<\/span> FullyShardedDataParallel\r\n\r\nmodel =<\/span> MyModel(<\/span>)<\/span>.<\/span>cuda(<\/span>)<\/span>\r\nmodel =<\/span> FullyShardedDataParallel(<\/span>model)<\/span>\r\noptimizer =<\/span> torch.<\/span>optim.<\/span>AdamW(<\/span>model.<\/span>parameters(<\/span>)<\/span>,<\/span> lr=<\/span>0.01<\/span>)<\/span>\r\n\r\n# training ...<\/span>\r\nfor<\/span> step,<\/span> (<\/span>x,<\/span> y)<\/span> in<\/span> enumerate<\/span>(<\/span>dataloader)<\/span>:<\/span>\r\n    optimizer.<\/span>zero_grad(<\/span>)<\/span>\r\n    output =<\/span> model(<\/span>x)<\/span>\r\n    loss_fn(<\/span>y,<\/span> output)<\/span>.<\/span>backward(<\/span>)<\/span>\r\n    optimizer.<\/span>step(<\/span>)<\/span>\r\n\r\n# save checkpoint<\/span>\r\nwith<\/span> model.<\/span>summon_full_params(<\/span>)<\/span>:<\/span>\r\n    if<\/span> rank ==<\/span> 0<\/span>:<\/span>\r\n        state =<\/span> model.<\/span>state_dict(<\/span>)<\/span>\r\n        torch.<\/span>save(<\/span>state,<\/span> 'model.pt'<\/span>)<\/span><\/code><\/pre>\n<\/div>\n
      haiscale.pipeline<\/h2>\n
      \"pipeline\"<\/a><\/span><\/p>\n
      \u5982\u4e0a\u56fe\u6240\u793a\uff0chaiscale.pipeline<\/code>\u00a0\u5de5\u5177\u5305\u4e2d\u63d0\u4f9b\u4e86\u4e09\u79cd\u6d41\u6c34\u7ebf\u5e76\u884c\u7684\u7b97\u6cd5\uff1a<\/p>\n
        \n
      1. GPipe<\/code>: \u628a\u6a21\u578b\u5207\u5206\u6210\u00a0ngpus<\/code>\u00a0\u4efd\uff0c\u6240\u6709 microbatch \u7684\u524d\u5411\u4f20\u64ad\u7ed3\u675f\u4e4b\u540e\u518d\u505a\u53cd\u5411\u4f20\u64ad\uff1b<\/li>\n
      2. PipeDream<\/code>: \u628a\u6a21\u578b\u5207\u5206\u6210\u00a0ngpus<\/code>\u00a0\u4efd\uff0c\u524d\u5411\u548c\u53cd\u5411\u4f20\u64ad\u4ea4\u66ff\u6267\u884c\uff08non-interleaved 1F1B\uff09\uff1b<\/li>\n
      3. Interleaved1F1B<\/code>: \u628a\u6a21\u578b\u5207\u5206\u6210\u00a0ngpus * num_model_chunks<\/code>\u00a0\u4efd\uff0c\u524d\u5411\u548c\u53cd\u5411\u4f20\u64ad\u4ea4\u66ff\u6267\u884c\u3002<\/li>\n<\/ol>\n
        \u5bf9\u4e8e\u4e2d\u7b49\u89c4\u6a21\u7684\u6a21\u578b\uff08\u6bd4\u5982 GPT-2 Medium\uff09\uff0c\u6211\u4eec\u63a8\u8350\u4f18\u5148\u4f7f\u7528 PipeDream\uff0c\u76f8\u6bd4\u4e8e GPipe \u548c Interleaved1F1B \u5360\u7528\u663e\u5b58\u66f4\u5c11\uff0c\u901f\u5ea6\u66f4\u5feb\u3002 \u5bf9\u4e8e\u8d85\u5927\u89c4\u6a21\u7684\u6a21\u578b\uff08\u6bd4\u5982 GPT-3\uff09\uff0c\u6211\u4eec\u53ef\u4ee5\u66f4\u52a0\u5747\u5300\u3001\u7ec6\u7c92\u5ea6\u7684\u5207\u5206\u5b83\uff0c\u8fd9\u65f6\u5019\u63a8\u8350\u4f7f\u7528 Interleaved1F1B\u3002<\/p>\n
        haiscale.pipeline<\/code>\u00a0\u63d0\u4f9b\u4e86\u4e00\u4e2a\u7edf\u4e00\u7684\u00a0forward_backward<\/code>\u00a0\u63a5\u53e3\uff0c\u6211\u4eec\u9700\u8981\u4f20\u5165\u635f\u5931\u51fd\u6570\u00a0criterion<\/code>\u00a0\u548c\u6807\u7b7e\u6570\u636e\u00a0labels<\/code>\uff0c\u635f\u5931\u51fd\u6570\u4f1a\u901a\u8fc7\u00a0loss = criterion(*outputs, *labels)<\/code>\u00a0\u7684\u65b9\u5f0f\u8c03\u7528\uff0cforward_backward<\/code>\u00a0\u63a5\u53e3\u4f1a\u8fd4\u56de\u4e00\u4e2a\u5143\u7ec4\u00a0(losses, outputs)<\/code>\uff0c\u5176\u4e2dlosses<\/code>\u00a0\u4ee3\u8868\u6bcf\u4e2a microbatch \u7684 loss \u503c\uff0coutputs<\/code>\u00a0\u4ee3\u8868\u6a21\u578b\u7684\u8f93\u51fa\u3002\u53ea\u6709\u6700\u540e\u4e00\u4e2a rank \u7684\u8fdb\u7a0b\u80fd\u591f\u83b7\u5f97 loss \u548c\u8f93\u51fa\uff0c\u5176\u4ed6\u8fdb\u7a0b\u5f97\u5230\u7684\u662f\u00a0(None, None)<\/code>\u3002<\/p>\n
        \u4e0b\u9762\u901a\u8fc7\u793a\u4f8b\u5c55\u793a GPipe \u548c PipeDream \u7684\u7528\u6cd5\uff1a<\/p>\n
        \n
        from<\/span> haiscale.<\/span>pipeline import<\/span> GPipe,<\/span> PipeDream,<\/span> partition\r\n\r\ndist.<\/span>init_process_group(<\/span>.<\/span>.<\/span>.<\/span>)<\/span>\r\ntorch.<\/span>cuda.<\/span>set_device(<\/span>local_rank)<\/span>\r\nrank,<\/span> world_size =<\/span> dist.<\/span>get_rank(<\/span>)<\/span>,<\/span> dist.<\/span>get_world_size(<\/span>)<\/span>\r\ntorch.<\/span>manual_seed(<\/span>12345<\/span>)<\/span>\r\n\r\ndef<\/span> loss_fn<\/span>(<\/span>out,<\/span> y)<\/span>:<\/span>\r\n    return<\/span> (<\/span>(<\/span>out -<\/span> y)<\/span>**<\/span>2<\/span>)<\/span>.<\/span>sum<\/span>(<\/span>)<\/span>\r\n\r\nmodel =<\/span> nn.<\/span>Sequential(<\/span>.<\/span>.<\/span>.<\/span>)<\/span>\r\nmodel =<\/span> partition(<\/span>model,<\/span> rank,<\/span> world_size)<\/span>\r\n\r\n# chunks: number of microbatches<\/span>\r\nmodel =<\/span> PipeDream(<\/span>model.<\/span>cuda(<\/span>)<\/span>,<\/span> chunks=<\/span>32<\/span>)<\/span>\r\n# or model = GPipe(model.cuda(), chunks=32)<\/span>\r\n\r\nfor<\/span> x,<\/span> y in<\/span> dataloader:<\/span>\r\n    losses,<\/span> outputs =<\/span> model.<\/span>forward_backward(<\/span>x,<\/span> criterion=<\/span>loss_fn,<\/span> labels=<\/span>(<\/span>y,<\/span>)<\/span>,<\/span> return_outputs=<\/span>True<\/span>)<\/span>\r\n    if<\/span> rank ==<\/span> world_size -<\/span> 1<\/span>:<\/span>\r\n        loss =<\/span> losses.<\/span>sum<\/span>(<\/span>)<\/span>.<\/span>item(<\/span>)<\/span>  # losses: torch.Size([32])<\/span>\r\n\r\n# eval<\/span>\r\nwith<\/span> torch.<\/span>no_grad(<\/span>)<\/span>:<\/span>\r\n    out =<\/span> model(<\/span>x)<\/span>\r\n    if<\/span> rank ==<\/span> world_size -<\/span> 1<\/span>:<\/span>\r\n        # calculate metrics ...<\/span><\/code><\/pre>\n<\/div>\n
        \u4ee5\u4e0b\u662f Interleaved1F1B \u7684\u4f7f\u7528\u793a\u4f8b\uff1a<\/p>\n
        \n
        from<\/span> haiscale.<\/span>pipeline import<\/span> Interleaved1F1B,<\/span> partition\r\n\r\ndist.<\/span>init_process_group(<\/span>.<\/span>.<\/span>.<\/span>)<\/span>\r\ntorch.<\/span>cuda.<\/span>set_device(<\/span>local_rank)<\/span>\r\nrank,<\/span> world_size =<\/span> dist.<\/span>get_rank(<\/span>)<\/span>,<\/span> dist.<\/span>get_world_size(<\/span>)<\/span>\r\ntorch.<\/span>manual_seed(<\/span>12345<\/span>)<\/span>\r\n\r\ndef<\/span> loss_fn<\/span>(<\/span>out,<\/span> y)<\/span>:<\/span>\r\n    return<\/span> (<\/span>(<\/span>out -<\/span> y)<\/span>**<\/span>2<\/span>)<\/span>.<\/span>sum<\/span>(<\/span>)<\/span>\r\n\r\nmodel =<\/span> nn.<\/span>Sequential(<\/span>.<\/span>.<\/span>.<\/span>)<\/span>\r\nmodules =<\/span> partition(<\/span>model,<\/span> rank,<\/span> world_size,<\/span> num_model_chunks=<\/span>2<\/span>)<\/span>  # len(modules) = 2<\/span>\r\nmodules =<\/span> [<\/span>m.<\/span>cuda(<\/span>)<\/span> for<\/span> m in<\/span> modules]<\/span>\r\nmodel =<\/span> Interleaved1F1B(<\/span>modules,<\/span> chunks=<\/span>32<\/span>)<\/span>\r\n\r\nfor<\/span> x,<\/span> y in<\/span> dataloader:<\/span>\r\n    losses,<\/span> outputs =<\/span> model.<\/span>forward_backward(<\/span>x,<\/span> criterion=<\/span>loss_fn,<\/span> labels=<\/span>(<\/span>y,<\/span>)<\/span>,<\/span> return_outputs=<\/span>True<\/span>)<\/span>\r\n    if<\/span> rank ==<\/span> world_size -<\/span> 1<\/span>:<\/span>\r\n        loss =<\/span> losses.<\/span>sum<\/span>(<\/span>)<\/span>.<\/span>item(<\/span>)<\/span>  # losses: torch.Size([32])<\/span>\r\n\r\n# eval<\/span>\r\nwith<\/span> torch.<\/span>no_grad(<\/span>)<\/span>:<\/span>\r\n    out =<\/span> model(<\/span>x)<\/span>\r\n    if<\/span> rank ==<\/span> world_size -<\/span> 1<\/span>:<\/span>\r\n        # calculate metrics ...<\/span><\/code><\/pre>\n<\/div>\n
        \u6570\u636e\u5e76\u884c\u548c\u6d41\u6c34\u7ebf\u5e76\u884c\u7ec4\u5408<\/h2>\n
        haiscale \u8fd8\u652f\u6301\u540c\u65f6\u4f7f\u7528 DDP \u548c\u6d41\u6c34\u7ebf\u5e76\u884c\u3002\u6bd4\u5982\u6211\u4eec\u6709 16 \u5757 GPU\uff0c\u6211\u4eec\u53ef\u4ee5\u628a\u8fd9 16 \u5757 GPU \u5212\u5206\u6210\u4e24\u4e2a\u7ec4\uff0c\u6bcf\u4e2a\u7ec4\u6709 8 \u5757 GPU\uff0c\u7136\u540e\u4e24\u4e2a\u7ec4\u4e4b\u95f4\u505a\u6570\u636e\u5e76\u884c\uff0c\u7ec4\u5185\u505a\u6d41\u6c34\u7ebf\u5e76\u884c\u3002<\/p>\n
        \u770b\u5982\u4e0b\u4f7f\u7528\u793a\u4f8b\uff1a<\/p>\n
        \n
        from<\/span> haiscale.<\/span>ddp import<\/span> DistributedDataParallel as<\/span> DDP\r\nfrom<\/span> haiscale.<\/span>pipeline import<\/span> PipeDream,<\/span> partition,<\/span> make_subgroups\r\n\r\ndist.<\/span>init_process_group(<\/span>.<\/span>.<\/span>.<\/span>)<\/span>\r\ntorch.<\/span>cuda.<\/span>set_device(<\/span>local_rank)<\/span>\r\nrank,<\/span> world_size =<\/span> dist.<\/span>get_rank(<\/span>)<\/span>,<\/span> dist.<\/span>get_world_size(<\/span>)<\/span>\r\n\r\ndp_group,<\/span> pp_group =<\/span> make_subgroups(<\/span>pp_size=<\/span>8<\/span>)<\/span>\r\n\r\nmodel =<\/span> nn.<\/span>Sequential(<\/span>.<\/span>.<\/span>.<\/span>)<\/span>\r\nmodel =<\/span> partition(<\/span>model,<\/span> pp_group.<\/span>rank(<\/span>)<\/span>,<\/span> pp_group.<\/span>size(<\/span>)<\/span>)<\/span>\r\n\r\nmodel =<\/span> DDP(<\/span>model.<\/span>cuda(<\/span>)<\/span>,<\/span> process_group=<\/span>dp_group)<\/span>\r\nmodel =<\/span> PipeDream(<\/span>model,<\/span> chunks=<\/span>64<\/span>,<\/span> process_group=<\/span>pp_group)<\/span>\r\n\r\ncriterion =<\/span> nn.<\/span>MSELoss(<\/span>)<\/span>\r\n\r\nfor<\/span> x,<\/span> y in<\/span> dataloader:<\/span>\r\n    model.<\/span>forward_backward(<\/span>x,<\/span> criterion=<\/span>criterion,<\/span> labels=<\/span>(<\/span>y,<\/span>)<\/span>)<\/span><\/code><\/pre>\n<\/div>\n
        CPU Offload<\/h2>\n
        \u9664\u4e86\u4ee5\u4e0a\u7684\u5e76\u884c\u7b56\u7565\uff0c\u5728\u6df1\u5ea6\u5b66\u4e60\u7684\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u5e38\u5e38\u4f1a\u9047\u5230\u663e\u5b58\u4e0d\u8db3\u7684\u95ee\u9898\u3002<\/p>\n
        haiscale.cpu_offload.CPUOffload<\/code>\u00a0\u80fd\u591f\u5e2e\u52a9\u6211\u4eec\u5728\u8bad\u7ec3\u4e2d\u628a\u4e00\u90e8\u5206\u9700\u8981\u4fdd\u5b58\u7684\u4e2d\u95f4\u53d8\u91cf\u79fb\u52a8\u5230 CPU \u5185\u5b58\u4e0a\uff0c\u7136\u540e\u5728\u53cd\u5411\u4f20\u64ad\u65f6\u628a\u9700\u8981\u7528\u5230\u7684 tensor \u4f20\u8f93\u56de GPU \u663e\u5b58\u91cc\uff0c\u4ece\u800c\u8fbe\u5230\u8282\u7701\u663e\u5b58\u7684\u76ee\u7684\u3002<\/p>\n
        haiscale \u91c7\u7528\u5f02\u6b65\u4f20\u8f93\u62f7\u8d1d<\/strong>\u7b56\u7565\uff0c\u80fd\u591f\u628a\u4e00\u90e8\u5206\u7684\u4f20\u8f93\u65f6\u95f4\u548c GPU \u7684\u8ba1\u7b97\u91cd\u53e0\u8d77\u6765\uff0c\u4ece\u800c\u51cf\u5c11\u62f7\u8d1d\u5e26\u6765\u7684\u5f00\u9500\uff0c\u63d0\u5347\u6574\u4f53\u8ba1\u7b97\u6548\u7387\u3002<\/p>\n
        \u4f7f\u7528\u65f6\u9700\u8981\u6307\u5b9a\u00a0offload_ratio<\/code>\u00a0\u53c2\u6570\uff0c\u5176\u4ee3\u8868\u9700\u8981 offload \u7684\u4e2d\u95f4\u53d8\u91cf\u7684\u6bd4\u4f8b\uff0coffload_ratio=1<\/code>\u00a0\u4ee3\u8868\u6240\u6709\u4fdd\u5b58\u7684\u4e2d\u95f4\u53d8\u91cf\u90fd\u4f1a\u88ab\u79fb\u52a8\u5230 CPU \u5185\u5b58\u91cc\u3002<\/p>\n
        \u4ee5\u4e0b\u662f\u4f7f\u7528\u793a\u4f8b\uff1a<\/p>\n
        \n
        from<\/span> haiscale.<\/span>cpu_offload import<\/span> CPUOffload\r\n\r\nmodel =<\/span> MyModel(<\/span>)<\/span>.<\/span>cuda(<\/span>)<\/span>\r\noptimizer =<\/span> torch.<\/span>optim.<\/span>AdamW(<\/span>model.<\/span>parameters(<\/span>)<\/span>,<\/span> lr=<\/span>0.01<\/span>)<\/span>\r\n\r\nfor<\/span> x,<\/span> y in<\/span> dataloader:<\/span>\r\n    optimizer.<\/span>zero_grad(<\/span>)<\/span>\r\n    with<\/span> CPUOffload(<\/span>offload_ratio=<\/span>0.1<\/span>,<\/span> tag=<\/span>\"MyModel\"<\/span>)<\/span>:<\/span>\r\n        output =<\/span> model(<\/span>x)<\/span>\r\n    loss_fn(<\/span>y,<\/span> output)<\/span>.<\/span>backward(<\/span>)<\/span>\r\n    optimizer.<\/span>step(<\/span>)<\/span><\/code><\/pre>\n<\/div>","protected":false},"excerpt":{"rendered":"
        haiscale (Highflyer AI Scale) \u662f\u4e00\u4e2a\u8f7b\u91cf\u7ea7\u7684\u9ad8\u6027\u80fd\u5e76\u884c\u8bad\u7ec3\u5de5\u5177\u5e93\uff0c\u5176\u6574\u5408\u4e86\u5e7b\u65b9  […]<\/p>","protected":false},"author":1,"featured_media":447,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[5],"tags":[],"class_list":["post-446","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-firefly-platform"],"acf":[],"_links":{"self":[{"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/posts\/446","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/comments?post=446"}],"version-history":[{"count":1,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/posts\/446\/revisions"}],"predecessor-version":[{"id":452,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/posts\/446\/revisions\/452"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/media\/447"}],"wp:attachment":[{"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/media?parent=446"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/categories?post=446"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/tags?post=446"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
      haiscale FSDP \u8fd8\u652f\u6301\u4f20\u5165\u00a0auto_wrap_policy<\/code>\u00a0\u53c2\u6570\uff0c\u5177\u4f53\u4f5c\u7528\u53ef\u4ee5\u53c2\u8003\u00a0PyTorch FSDP \u7684\u6587\u6863<\/a>\u4ee5\u53ca\u6211\u4eec\u63d0\u4f9b\u7684\u00a0GPT-2 \u793a\u4f8b<\/a>\u3002<\/p>\n