UserKernel init CUDA Graphs with state

[liujuncheng]

No description provided.

[github-actions]

Speed stats:

GPU Name: GeForce GTX 1080 

OneFlow resnet50 time: 126.5ms (= 6323.5ms / 50, input_shape=[16, 3, 224, 224])

OneFlow resnet50 time: 72.7ms (= 3632.9ms / 50, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 82.1ms (= 4104.6ms / 50, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.13 (= 82.1ms / 72.7ms)

OneFlow resnet50 time: 120.1ms (= 6004.6ms / 50, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 103.0ms (= 5150.9ms / 50, input_shape=[4, 3, 224, 224])
❌ Relative speed: 0.86 (= 103.0ms / 120.1ms)

OneFlow resnet50 time: 104.3ms (= 5215.7ms / 50, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 89.9ms (= 4495.6ms / 50, input_shape=[2, 3, 224, 224])
❌ Relative speed: 0.86 (= 89.9ms / 104.3ms)

OneFlow resnet50 time: 89.6ms (= 4478.9ms / 50, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 79.5ms (= 3974.6ms / 50, input_shape=[1, 3, 224, 224])
❌ Relative speed: 0.89 (= 79.5ms / 89.6ms)

OneFlow resnet50 time: 160.3ms (= 8014.3ms / 50, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 159.9ms (= 7997.1ms / 50, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.00 (= 159.9ms / 160.3ms)

OneFlow resnet50 time: 100.2ms (= 5010.5ms / 50, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 113.5ms (= 5673.7ms / 50, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.13 (= 113.5ms / 100.2ms)

OneFlow resnet50 time: 82.6ms (= 4131.2ms / 50, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 76.5ms (= 3824.2ms / 50, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 0.93 (= 76.5ms / 82.6ms)

OneFlow resnet50 time: 71.6ms (= 3579.0ms / 50, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 68.2ms (= 3408.8ms / 50, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 0.95 (= 68.2ms / 71.6ms)

OneFlow resnet50 time: 74.6ms (= 3730.9ms / 50, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 63.9ms (= 3192.9ms / 50, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 0.86 (= 63.9ms / 74.6ms)