Feat graph load to new device #10335

strint · 2023-09-22T14:07:54Z

No description provided.

github-actions · 2023-09-22T14:09:45Z

Code got formatted by CI. Please request CI again if you still want to have this PR merged. If the PR is from a forked repo, please download the patch files from the GitHub Actions web page and apply them locally.

github-actions · 2023-09-22T14:46:42Z

View latest API docs preview at: https://oneflow-staging.oss-cn-beijing.aliyuncs.com/docs/Oneflow-Inc/oneflow/pr/10335/

github-actions · 2023-09-22T14:51:10Z

Speed stats:

GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.9ms (= 4389.0ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 57.3ms (= 5725.1ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.30 (= 57.3ms / 43.9ms)

OneFlow resnet50 time: 26.4ms (= 2644.1ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 37.6ms (= 3763.5ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.42 (= 37.6ms / 26.4ms)

OneFlow resnet50 time: 18.7ms (= 3748.9ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 35.8ms (= 7166.9ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.91 (= 35.8ms / 18.7ms)

OneFlow resnet50 time: 18.9ms (= 3774.4ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 33.3ms (= 6651.0ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.76 (= 33.3ms / 18.9ms)

OneFlow resnet50 time: 18.0ms (= 3603.4ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 29.3ms (= 5853.9ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.62 (= 29.3ms / 18.0ms)

OneFlow swin dataloader time: 0.202s (= 40.332s / 200, num_workers=1)
PyTorch swin dataloader time: 0.128s (= 25.674s / 200, num_workers=1)
Relative speed: 0.637 (= 0.128s / 0.202s)

OneFlow swin dataloader time: 0.056s (= 11.177s / 200, num_workers=4)
PyTorch swin dataloader time: 0.033s (= 6.591s / 200, num_workers=4)
Relative speed: 0.590 (= 0.033s / 0.056s)

OneFlow swin dataloader time: 0.031s (= 6.174s / 200, num_workers=8)
PyTorch swin dataloader time: 0.018s (= 3.543s / 200, num_workers=8)
Relative speed: 0.574 (= 0.018s / 0.031s)

❌ OneFlow resnet50 time: 47.6ms (= 4760.4ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 65.4ms (= 6540.6ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.37 (= 65.4ms / 47.6ms)

OneFlow resnet50 time: 32.8ms (= 3283.1ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 43.7ms (= 4370.0ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.33 (= 43.7ms / 32.8ms)

OneFlow resnet50 time: 23.8ms (= 4766.5ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 41.3ms (= 8252.5ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.73 (= 41.3ms / 23.8ms)

OneFlow resnet50 time: 22.0ms (= 4404.8ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 36.5ms (= 7307.8ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.66 (= 36.5ms / 22.0ms)

OneFlow resnet50 time: 21.1ms (= 4214.4ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 36.0ms (= 7200.2ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.71 (= 36.0ms / 21.1ms)

…-Inc/oneflow into feat_graph_to_new_device

github-actions · 2023-09-26T14:19:22Z

Code got formatted by CI. Please request CI again if you still want to have this PR merged. If the PR is from a forked repo, please download the patch files from the GitHub Actions web page and apply them locally.

github-actions · 2023-09-26T14:48:57Z

Speed stats:

github-actions · 2023-09-27T10:03:50Z

Code got formatted by CI. Please request CI again if you still want to have this PR merged. If the PR is from a forked repo, please download the patch files from the GitHub Actions web page and apply them locally.

github-actions · 2023-09-27T10:36:41Z

View latest API docs preview at: https://oneflow-staging.oss-cn-beijing.aliyuncs.com/docs/Oneflow-Inc/oneflow/pr/10335/

github-actions · 2023-09-27T10:41:09Z

Speed stats:

GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.8ms (= 4377.9ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 57.2ms (= 5720.9ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.31 (= 57.2ms / 43.8ms)

OneFlow resnet50 time: 26.2ms (= 2615.1ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 38.1ms (= 3813.8ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.46 (= 38.1ms / 26.2ms)

OneFlow resnet50 time: 18.9ms (= 3782.8ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 36.0ms (= 7191.9ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.90 (= 36.0ms / 18.9ms)

OneFlow resnet50 time: 18.3ms (= 3655.3ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 33.7ms (= 6744.6ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.85 (= 33.7ms / 18.3ms)

OneFlow resnet50 time: 18.0ms (= 3596.9ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 30.0ms (= 5991.6ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.67 (= 30.0ms / 18.0ms)

OneFlow swin dataloader time: 0.202s (= 40.399s / 200, num_workers=1)
PyTorch swin dataloader time: 0.128s (= 25.604s / 200, num_workers=1)
Relative speed: 0.634 (= 0.128s / 0.202s)

OneFlow swin dataloader time: 0.053s (= 10.641s / 200, num_workers=4)
PyTorch swin dataloader time: 0.033s (= 6.613s / 200, num_workers=4)
Relative speed: 0.621 (= 0.033s / 0.053s)

OneFlow swin dataloader time: 0.030s (= 6.048s / 200, num_workers=8)
PyTorch swin dataloader time: 0.017s (= 3.408s / 200, num_workers=8)
Relative speed: 0.563 (= 0.017s / 0.030s)

❌ OneFlow resnet50 time: 47.5ms (= 4751.5ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 64.3ms (= 6432.2ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.35 (= 64.3ms / 47.5ms)

OneFlow resnet50 time: 31.7ms (= 3169.6ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 45.6ms (= 4557.3ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.44 (= 45.6ms / 31.7ms)

OneFlow resnet50 time: 24.2ms (= 4834.7ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 41.3ms (= 8260.4ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.71 (= 41.3ms / 24.2ms)

OneFlow resnet50 time: 22.1ms (= 4423.7ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 36.9ms (= 7372.4ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.67 (= 36.9ms / 22.1ms)

OneFlow resnet50 time: 20.7ms (= 4132.0ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 34.3ms (= 6867.4ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.66 (= 34.3ms / 20.7ms)

github-actions · 2023-09-28T07:16:56Z

Code got formatted by CI. Please request CI again if you still want to have this PR merged. If the PR is from a forked repo, please download the patch files from the GitHub Actions web page and apply them locally.

github-actions · 2023-09-28T07:46:34Z

Speed stats:

github-actions · 2023-09-30T05:15:05Z

Code got formatted by CI. Please request CI again if you still want to have this PR merged. If the PR is from a forked repo, please download the patch files from the GitHub Actions web page and apply them locally.

github-actions · 2023-09-30T05:44:22Z

Speed stats:

github-actions · 2023-10-04T02:00:57Z

Speed stats:

GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.2ms (= 4325.0ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 57.3ms (= 5731.7ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.33 (= 57.3ms / 43.2ms)

OneFlow resnet50 time: 26.1ms (= 2614.4ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 38.0ms (= 3799.5ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.45 (= 38.0ms / 26.1ms)

OneFlow resnet50 time: 20.4ms (= 4089.0ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 36.1ms (= 7210.8ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.76 (= 36.1ms / 20.4ms)

OneFlow resnet50 time: 18.1ms (= 3615.9ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 33.2ms (= 6648.9ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.84 (= 33.2ms / 18.1ms)

OneFlow resnet50 time: 17.7ms (= 3541.5ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 29.6ms (= 5921.5ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.67 (= 29.6ms / 17.7ms)

OneFlow swin dataloader time: 0.200s (= 40.085s / 200, num_workers=1)
PyTorch swin dataloader time: 0.128s (= 25.563s / 200, num_workers=1)
Relative speed: 0.638 (= 0.128s / 0.200s)

OneFlow swin dataloader time: 0.053s (= 10.673s / 200, num_workers=4)
PyTorch swin dataloader time: 0.033s (= 6.596s / 200, num_workers=4)
Relative speed: 0.618 (= 0.033s / 0.053s)

OneFlow swin dataloader time: 0.032s (= 6.398s / 200, num_workers=8)
PyTorch swin dataloader time: 0.016s (= 3.300s / 200, num_workers=8)
Relative speed: 0.516 (= 0.016s / 0.032s)

❌ OneFlow resnet50 time: 47.5ms (= 4753.1ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 64.9ms (= 6489.4ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.37 (= 64.9ms / 47.5ms)

OneFlow resnet50 time: 31.6ms (= 3163.2ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 45.7ms (= 4568.7ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.44 (= 45.7ms / 31.6ms)

OneFlow resnet50 time: 23.9ms (= 4771.3ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 42.6ms (= 8516.7ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.78 (= 42.6ms / 23.9ms)

OneFlow resnet50 time: 22.0ms (= 4404.8ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 36.4ms (= 7273.6ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.65 (= 36.4ms / 22.0ms)

OneFlow resnet50 time: 20.9ms (= 4176.5ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 34.5ms (= 6895.4ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.65 (= 34.5ms / 20.9ms)

github-actions · 2023-10-04T05:03:09Z

Speed stats:

github-actions · 2023-10-04T05:21:31Z

Code got formatted by CI. Please request CI again if you still want to have this PR merged. If the PR is from a forked repo, please download the patch files from the GitHub Actions web page and apply them locally.

github-actions · 2023-10-04T05:29:02Z

View latest API docs preview at: https://oneflow-staging.oss-cn-beijing.aliyuncs.com/docs/Oneflow-Inc/oneflow/pr/10335/

github-actions · 2023-10-04T05:33:43Z

Speed stats:

GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.5ms (= 4349.4ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 57.8ms (= 5777.1ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.33 (= 57.8ms / 43.5ms)

OneFlow resnet50 time: 25.9ms (= 2585.1ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 37.6ms (= 3756.5ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.45 (= 37.6ms / 25.9ms)

OneFlow resnet50 time: 19.4ms (= 3872.5ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 37.3ms (= 7458.7ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.93 (= 37.3ms / 19.4ms)

OneFlow resnet50 time: 18.4ms (= 3678.1ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 31.5ms (= 6305.2ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.71 (= 31.5ms / 18.4ms)

OneFlow resnet50 time: 18.2ms (= 3638.1ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 30.1ms (= 6024.4ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.66 (= 30.1ms / 18.2ms)

OneFlow swin dataloader time: 0.202s (= 40.338s / 200, num_workers=1)
PyTorch swin dataloader time: 0.128s (= 25.624s / 200, num_workers=1)
Relative speed: 0.635 (= 0.128s / 0.202s)

OneFlow swin dataloader time: 0.055s (= 10.956s / 200, num_workers=4)
PyTorch swin dataloader time: 0.032s (= 6.487s / 200, num_workers=4)
Relative speed: 0.592 (= 0.032s / 0.055s)

OneFlow swin dataloader time: 0.031s (= 6.229s / 200, num_workers=8)
PyTorch swin dataloader time: 0.017s (= 3.352s / 200, num_workers=8)
Relative speed: 0.538 (= 0.017s / 0.031s)

❌ OneFlow resnet50 time: 47.8ms (= 4784.7ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 65.2ms (= 6524.7ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.36 (= 65.2ms / 47.8ms)

OneFlow resnet50 time: 31.5ms (= 3146.0ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 43.5ms (= 4353.3ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.38 (= 43.5ms / 31.5ms)

OneFlow resnet50 time: 24.3ms (= 4852.5ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 41.1ms (= 8217.9ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.69 (= 41.1ms / 24.3ms)

OneFlow resnet50 time: 21.6ms (= 4312.9ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 36.6ms (= 7318.2ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.70 (= 36.6ms / 21.6ms)

OneFlow resnet50 time: 20.2ms (= 4031.4ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 34.9ms (= 6972.5ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.73 (= 34.9ms / 20.2ms)

python/oneflow/test/expensive/test_graph_multi_graph_v2.py

python/oneflow/nn/graph/graph.py

github-actions · 2023-10-05T11:10:37Z

Code got formatted by CI. Please request CI again if you still want to have this PR merged. If the PR is from a forked repo, please download the patch files from the GitHub Actions web page and apply them locally.

github-actions · 2023-10-05T11:43:15Z

View latest API docs preview at: https://oneflow-staging.oss-cn-beijing.aliyuncs.com/docs/Oneflow-Inc/oneflow/pr/10335/

github-actions · 2023-10-05T11:47:50Z

Speed stats:

GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.7ms (= 4371.4ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 61.6ms (= 6161.1ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.41 (= 61.6ms / 43.7ms)

OneFlow resnet50 time: 26.1ms (= 2610.2ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 37.6ms (= 3764.3ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.44 (= 37.6ms / 26.1ms)

OneFlow resnet50 time: 19.3ms (= 3856.1ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 34.8ms (= 6957.1ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.80 (= 34.8ms / 19.3ms)

OneFlow resnet50 time: 19.1ms (= 3827.4ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 32.2ms (= 6441.2ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.68 (= 32.2ms / 19.1ms)

OneFlow resnet50 time: 18.6ms (= 3715.5ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 28.2ms (= 5640.4ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.52 (= 28.2ms / 18.6ms)

OneFlow swin dataloader time: 0.201s (= 40.292s / 200, num_workers=1)
PyTorch swin dataloader time: 0.130s (= 25.944s / 200, num_workers=1)
Relative speed: 0.644 (= 0.130s / 0.201s)

OneFlow swin dataloader time: 0.055s (= 11.040s / 200, num_workers=4)
PyTorch swin dataloader time: 0.033s (= 6.566s / 200, num_workers=4)
Relative speed: 0.595 (= 0.033s / 0.055s)

OneFlow swin dataloader time: 0.030s (= 6.066s / 200, num_workers=8)
PyTorch swin dataloader time: 0.017s (= 3.311s / 200, num_workers=8)
Relative speed: 0.546 (= 0.017s / 0.030s)

❌ OneFlow resnet50 time: 47.6ms (= 4755.7ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 63.6ms (= 6362.2ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.34 (= 63.6ms / 47.6ms)

OneFlow resnet50 time: 31.1ms (= 3109.5ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 45.0ms (= 4502.0ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.45 (= 45.0ms / 31.1ms)

OneFlow resnet50 time: 24.0ms (= 4799.7ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 40.7ms (= 8135.3ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.69 (= 40.7ms / 24.0ms)

OneFlow resnet50 time: 22.5ms (= 4498.4ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 36.7ms (= 7337.2ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.63 (= 36.7ms / 22.5ms)

OneFlow resnet50 time: 21.0ms (= 4208.8ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 35.4ms (= 7081.2ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.68 (= 35.4ms / 21.0ms)

github-actions · 2023-10-06T04:29:49Z

View latest API docs preview at: https://oneflow-staging.oss-cn-beijing.aliyuncs.com/docs/Oneflow-Inc/oneflow/pr/10335/

github-actions · 2023-10-06T04:34:16Z

Speed stats:

GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.8ms (= 4376.3ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 61.6ms (= 6161.1ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.41 (= 61.6ms / 43.8ms)

OneFlow resnet50 time: 26.1ms (= 2605.9ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 38.5ms (= 3851.8ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.48 (= 38.5ms / 26.1ms)

OneFlow resnet50 time: 18.4ms (= 3689.9ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 36.2ms (= 7238.3ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.96 (= 36.2ms / 18.4ms)

OneFlow resnet50 time: 18.7ms (= 3733.8ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 31.5ms (= 6296.1ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.69 (= 31.5ms / 18.7ms)

OneFlow resnet50 time: 17.4ms (= 3487.0ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 29.6ms (= 5922.9ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.70 (= 29.6ms / 17.4ms)

OneFlow swin dataloader time: 0.203s (= 40.516s / 200, num_workers=1)
PyTorch swin dataloader time: 0.129s (= 25.817s / 200, num_workers=1)
Relative speed: 0.637 (= 0.129s / 0.203s)

OneFlow swin dataloader time: 0.054s (= 10.703s / 200, num_workers=4)
PyTorch swin dataloader time: 0.034s (= 6.708s / 200, num_workers=4)
Relative speed: 0.627 (= 0.034s / 0.054s)

OneFlow swin dataloader time: 0.031s (= 6.152s / 200, num_workers=8)
PyTorch swin dataloader time: 0.017s (= 3.432s / 200, num_workers=8)
Relative speed: 0.558 (= 0.017s / 0.031s)

❌ OneFlow resnet50 time: 47.7ms (= 4766.0ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 64.3ms (= 6431.8ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.35 (= 64.3ms / 47.7ms)

OneFlow resnet50 time: 31.1ms (= 3113.4ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 44.1ms (= 4413.3ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.42 (= 44.1ms / 31.1ms)

OneFlow resnet50 time: 23.9ms (= 4783.7ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 41.4ms (= 8287.9ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.73 (= 41.4ms / 23.9ms)

OneFlow resnet50 time: 22.1ms (= 4417.1ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 37.5ms (= 7505.8ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.70 (= 37.5ms / 22.1ms)

OneFlow resnet50 time: 20.8ms (= 4159.3ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 33.9ms (= 6774.6ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.63 (= 33.9ms / 20.8ms)

Support changing runtime_state_dict's device with `runtime_state_dict_to` ```python def warmup_with_load(self, file_path, device=None): state_dict = flow.load(file_path) if device is not None: state_dict = flow.nn.Graph.runtime_state_dict_to(state_dict, device) self.load_runtime_state_dict(state_dict) ``` Depends on PR in oneflow: Oneflow-Inc/oneflow#10335 # Performance check ## save - speed: 5.99~6.08 it/s - mem: - before compile: 7.4G - after compile: 8.6 G - run: 15G ## load - speed: 5.94~6.11 it/s - mem: - before compile: 8.5G - after compile: 8.5G - run: 15G ## load from cuda 0 to cuda 1 - speed: 6.17~6.22 it/s - mem: - before compile: 7.3G - after compile: 8.5G - run: 15G ## load from cuda 0 to cuda 0 and cuda 1 - speed: 6.05~6.12 it/s - mem: - before compile: 7.3G - after compile: 8.5G - run: 15G --------- Co-authored-by: Shenghang Tsai <[email protected]>

This reverts commit dea3f43.

strint added 2 commits September 22, 2023 09:28

fix load and warmup for kwargs

4e60628

fix save and load from non 0 device

905ccaf

strint mentioned this pull request Sep 22, 2023

multi device demo siliconflow/onediff#258

Merged

auto format by CI

bb2f2ac

strint and others added 3 commits September 26, 2023 14:16

rsd to part 1

2cd9919

Merge branch 'feat_graph_to_new_device' of https://github.com/Oneflow…

45cd46c

…-Inc/oneflow into feat_graph_to_new_device

auto format by CI

5957b84

strint added 3 commits September 26, 2023 17:31

draft 2

f00dff8

draft 3

591e2aa

merge

4957da5

strint added graph graph mode feature labels Sep 27, 2023

auto format by CI

60b6ee1

strint and others added 2 commits September 28, 2023 07:14

rm debug and add test

887b9aa

auto format by CI

787de41

strint and others added 2 commits September 30, 2023 05:12

fix actor id

b2602c1

auto format by CI

1098675

strint changed the title ~~Feat graph load to non 0 device~~ Feat graph load to new device Oct 4, 2023

rm debug

7cfc873

strint and others added 2 commits October 4, 2023 05:18

refine test

afb16fc

auto format by CI

7cc7043

jackalcooper reviewed Oct 5, 2023

View reviewed changes

python/oneflow/test/expensive/test_graph_multi_graph_v2.py Outdated Show resolved Hide resolved

jackalcooper reviewed Oct 5, 2023

View reviewed changes

python/oneflow/nn/graph/graph.py Show resolved Hide resolved

strint and others added 3 commits October 5, 2023 10:51

rename with_new

5399268

address review: split funcs

7f8f837

auto format by CI

63a1195

leaves-zwx approved these changes Oct 6, 2023

View reviewed changes

Merge branch 'master' into feat_graph_to_new_device

da1b978

strint added the automerge label Oct 6, 2023

strint merged commit dea3f43 into master Oct 6, 2023

strint deleted the feat_graph_to_new_device branch October 6, 2023 07:28

jackalcooper added a commit that referenced this pull request Oct 18, 2023

Revert "Feat graph load to new device (#10335)"

c7c2b1a

This reverts commit dea3f43.

jackalcooper added a commit that referenced this pull request Dec 20, 2023

Revert "Feat graph load to new device (#10335)"

29b5c13

This reverts commit dea3f43.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Feat graph load to new device #10335

Feat graph load to new device #10335

strint commented Sep 22, 2023

github-actions bot commented Sep 22, 2023

github-actions bot commented Sep 22, 2023

github-actions bot commented Sep 22, 2023

github-actions bot commented Sep 26, 2023

github-actions bot commented Sep 26, 2023

github-actions bot commented Sep 27, 2023

github-actions bot commented Sep 27, 2023

github-actions bot commented Sep 27, 2023

github-actions bot commented Sep 28, 2023

github-actions bot commented Sep 28, 2023

github-actions bot commented Sep 30, 2023

github-actions bot commented Sep 30, 2023

github-actions bot commented Oct 4, 2023

github-actions bot commented Oct 4, 2023

github-actions bot commented Oct 4, 2023

github-actions bot commented Oct 4, 2023

github-actions bot commented Oct 4, 2023

github-actions bot commented Oct 5, 2023

github-actions bot commented Oct 5, 2023

github-actions bot commented Oct 5, 2023

github-actions bot commented Oct 6, 2023

github-actions bot commented Oct 6, 2023

Feat graph load to new device #10335

Feat graph load to new device #10335

Conversation

strint commented Sep 22, 2023

github-actions bot commented Sep 22, 2023

github-actions bot commented Sep 22, 2023

github-actions bot commented Sep 22, 2023

github-actions bot commented Sep 26, 2023

github-actions bot commented Sep 26, 2023

github-actions bot commented Sep 27, 2023

github-actions bot commented Sep 27, 2023

github-actions bot commented Sep 27, 2023

github-actions bot commented Sep 28, 2023

github-actions bot commented Sep 28, 2023

github-actions bot commented Sep 30, 2023

github-actions bot commented Sep 30, 2023

github-actions bot commented Oct 4, 2023

github-actions bot commented Oct 4, 2023

github-actions bot commented Oct 4, 2023

github-actions bot commented Oct 4, 2023

github-actions bot commented Oct 4, 2023

github-actions bot commented Oct 5, 2023

github-actions bot commented Oct 5, 2023

github-actions bot commented Oct 5, 2023

github-actions bot commented Oct 6, 2023

github-actions bot commented Oct 6, 2023