Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

add fuse_layer_norm_grad functor and op #10612

Open
wants to merge 4 commits into
base: master
Choose a base branch
from
Open

add fuse_layer_norm_grad functor and op #10612

wants to merge 4 commits into from
+294 −16

Conversation

crazy-JiangDongHua
Copy link
Contributor

No description provided.

@github-actions GitHub Actions
Copy link
Contributor

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 GitHub Actions
Copy link
Contributor

View latest API docs preview at: https://oneflow-staging.oss-cn-beijing.aliyuncs.com/docs/Oneflow-Inc/oneflow/pr/10612/

@github-actions GitHub Actions
Copy link
Contributor

Speed stats: 
GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.9ms (= 4392.4ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 57.6ms (= 5758.7ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.31 (= 57.6ms / 43.9ms)

OneFlow resnet50 time: 26.4ms (= 2642.9ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 37.1ms (= 3713.9ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.41 (= 37.1ms / 26.4ms)

OneFlow resnet50 time: 18.5ms (= 3693.1ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 35.4ms (= 7083.0ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.92 (= 35.4ms / 18.5ms)

OneFlow resnet50 time: 17.3ms (= 3469.9ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 34.0ms (= 6800.5ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.96 (= 34.0ms / 17.3ms)

OneFlow resnet50 time: 16.8ms (= 3350.3ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 29.9ms (= 5985.4ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.79 (= 29.9ms / 16.8ms)

OneFlow swin dataloader time: 0.200s (= 40.006s / 200, num_workers=1)
PyTorch swin dataloader time: 0.128s (= 25.541s / 200, num_workers=1)
Relative speed: 0.638 (= 0.128s / 0.200s)

OneFlow swin dataloader time: 0.056s (= 11.249s / 200, num_workers=4)
PyTorch swin dataloader time: 0.032s (= 6.491s / 200, num_workers=4)
Relative speed: 0.577 (= 0.032s / 0.056s)

OneFlow swin dataloader time: 0.031s (= 6.203s / 200, num_workers=8)
PyTorch swin dataloader time: 0.017s (= 3.308s / 200, num_workers=8)
Relative speed: 0.533 (= 0.017s / 0.031s)

❌ OneFlow resnet50 time: 49.2ms (= 4919.4ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 65.3ms (= 6533.9ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.33 (= 65.3ms / 49.2ms)

OneFlow resnet50 time: 36.2ms (= 3619.8ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 45.5ms (= 4552.4ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.26 (= 45.5ms / 36.2ms)

OneFlow resnet50 time: 27.9ms (= 5576.9ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 42.1ms (= 8422.8ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.51 (= 42.1ms / 27.9ms)

OneFlow resnet50 time: 25.3ms (= 5063.8ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 38.5ms (= 7693.1ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.52 (= 38.5ms / 25.3ms)

OneFlow resnet50 time: 24.8ms (= 4965.7ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 35.8ms (= 7163.8ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.44 (= 35.8ms / 24.8ms)

@ShawnXuan
Copy link
Collaborator

是否需要一个空的 FuseLayerNormGradCpuKernel 在 oneflow/user/kernels/layer_norm_cpu_kernel.cpp 文件中

@github-actions GitHub Actions
Copy link
Contributor

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.

ShawnXuan
ShawnXuan reviewed Jan 21, 2025
View reviewed changes
oneflow/core/functional/impl/nn_grad_functor.cpp
std::shared_ptr<OpExpr> op_;
};

class FuseLayerNormAffineGradFunctor {
Copy link
Collaborator

@ShawnXuan ShawnXuan Jan 21, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

建议只保留带affine功能的这个

ShawnXuan
ShawnXuan reviewed Jan 21, 2025
View reviewed changes
oneflow/core/functional/functional_api.yaml
signature: "TensorTuple (Tensor dy, Tensor x, Tensor mean, Tensor inv_variance, Int64 begin_norm_axis, Int64 begin_params_axis, Double epsilon) => FuseLayerNormGrad"
bind_python: False

- name: "fuse_layer_norm_affine_grad"
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

fuse_layer_norm_grad fuse_layer_norm_affine_grad 建议只保留一个

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@ShawnXuan ShawnXuan ShawnXuan left review comments

@hjchen2 hjchen2 Awaiting requested review from hjchen2 hjchen2 is a code owner

@jackalcooper jackalcooper Awaiting requested review from jackalcooper jackalcooper is a code owner

At least 1 approving review is required to merge this pull request.

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
3 participants
@crazy-JiangDongHua @ShawnXuan @oneflow-ci-bot