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 support for variable sequence length in fmha #9991

Merged
merged 11 commits into from
Mar 16, 2023
Merged

Add support for variable sequence length in fmha #9991

merged 11 commits into from
Mar 16, 2023

Conversation

liujuncheng
Copy link
Collaborator

No description provided.

jackalcooper
jackalcooper reviewed Mar 16, 2023
View reviewed changes
oneflow/core/functional/functional_api.yaml Outdated
Copy link
Collaborator

Choose a reason for hiding this comment

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

这个接口加个readme，中文也可以

jackalcooper
jackalcooper reviewed Mar 16, 2023
View reviewed changes
oneflow/user/ops/fused_attention_ops.cpp
} else {
UNIMPLEMENTED_THEN_RETURN();
}
} else if (shape.NumAxes() == 3) {
if (layout == "BM(HK)" || layout == "MB(HK)" || layout == "BM(H2K)" || layout == "MB(H2K)"
Copy link
Collaborator

@jackalcooper jackalcooper Mar 16, 2023 •
edited
Loading

Choose a reason for hiding this comment

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

把这些字符串改为一个ENUM或者一系列static变量？这样万一出错了不需要肉眼看的很辛苦

@github-actions
Copy link
Contributor

Speed stats:

ofhwei
ofhwei reviewed Mar 16, 2023
View reviewed changes
oneflow/core/functional/impl/fused_attention_functor.cpp Outdated
UNIMPLEMENTED_THEN_RETURN() << name
<< "_layout should be '(BM)(HK)', '(BM)(H2K)', or '(BM)(H3K)' "
"when the number of dimensions of "
<< name << " tensor is 3.";
Copy link
Contributor

Choose a reason for hiding this comment

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

这里为：tensor is 3. -> tensor is 2.

oneflow/core/functional/impl/fused_attention_functor.cpp
*b = JUST(batch_size);
*m = JUST(seq_len);
*h = shape.At(1);
*k = shape.At(2);
} else {
UNIMPLEMENTED_THEN_RETURN()
<< name
Copy link
Contributor

Choose a reason for hiding this comment

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

这里漏了"(BM)HK"

@github-actions
Copy link
Contributor

Speed stats: 
GPU Name: GeForce GTX 1080 

❌ OneFlow resnet50 time: 141.1ms (= 14109.6ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 141.7ms (= 14172.9ms / 100, input_shape=[16, 3, 224, 224])
❌ Relative speed: 1.00 (= 141.7ms / 141.1ms)

OneFlow resnet50 time: 80.6ms (= 8063.1ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 84.2ms (= 8423.8ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.04 (= 84.2ms / 80.6ms)

OneFlow resnet50 time: 49.3ms (= 9852.8ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 57.9ms (= 11574.1ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.17 (= 57.9ms / 49.3ms)

OneFlow resnet50 time: 32.4ms (= 6487.3ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 44.3ms (= 8850.8ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.36 (= 44.3ms / 32.4ms)

OneFlow resnet50 time: 25.7ms (= 5133.5ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 34.0ms (= 6796.5ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.32 (= 34.0ms / 25.7ms)

OneFlow swin dataloader time: 0.239s (= 47.864s / 200, num_workers=1)
PyTorch swin dataloader time: 0.147s (= 29.363s / 200, num_workers=1)
Relative speed: 0.613 (= 0.147s / 0.239s)

OneFlow swin dataloader time: 0.070s (= 13.964s / 200, num_workers=4)
PyTorch swin dataloader time: 0.041s (= 8.203s / 200, num_workers=4)
Relative speed: 0.587 (= 0.041s / 0.070s)

OneFlow swin dataloader time: 0.046s (= 9.200s / 200, num_workers=8)
PyTorch swin dataloader time: 0.022s (= 4.473s / 200, num_workers=8)
Relative speed: 0.486 (= 0.022s / 0.046s)

❌ OneFlow resnet50 time: 152.2ms (= 15217.7ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 161.6ms (= 16160.1ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
❌ Relative speed: 1.06 (= 161.6ms / 152.2ms)

OneFlow resnet50 time: 90.9ms (= 9094.3ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 103.2ms (= 10317.0ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.13 (= 103.2ms / 90.9ms)

OneFlow resnet50 time: 58.9ms (= 11775.3ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 79.1ms (= 15823.5ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.34 (= 79.1ms / 58.9ms)

OneFlow resnet50 time: 41.5ms (= 8290.7ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 78.9ms (= 15783.6ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.90 (= 78.9ms / 41.5ms)

OneFlow resnet50 time: 36.8ms (= 7367.2ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 75.2ms (= 15045.4ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 2.04 (= 75.2ms / 36.8ms)

@github-actions
Copy link
Contributor

Speed stats: 
GPU Name: GeForce GTX 1080 

❌ OneFlow resnet50 time: 141.0ms (= 14098.3ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 143.9ms (= 14394.2ms / 100, input_shape=[16, 3, 224, 224])
❌ Relative speed: 1.02 (= 143.9ms / 141.0ms)

OneFlow resnet50 time: 80.6ms (= 8058.4ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 84.0ms (= 8396.9ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.04 (= 84.0ms / 80.6ms)

OneFlow resnet50 time: 48.9ms (= 9771.3ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 58.7ms (= 11735.3ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.20 (= 58.7ms / 48.9ms)

OneFlow resnet50 time: 32.4ms (= 6475.1ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 45.2ms (= 9038.0ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.40 (= 45.2ms / 32.4ms)

OneFlow resnet50 time: 25.7ms (= 5145.6ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 37.6ms (= 7529.7ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.46 (= 37.6ms / 25.7ms)

OneFlow swin dataloader time: 0.238s (= 47.674s / 200, num_workers=1)
PyTorch swin dataloader time: 0.158s (= 31.624s / 200, num_workers=1)
Relative speed: 0.663 (= 0.158s / 0.238s)

OneFlow swin dataloader time: 0.069s (= 13.786s / 200, num_workers=4)
PyTorch swin dataloader time: 0.042s (= 8.317s / 200, num_workers=4)
Relative speed: 0.603 (= 0.042s / 0.069s)

OneFlow swin dataloader time: 0.041s (= 8.135s / 200, num_workers=8)
PyTorch swin dataloader time: 0.022s (= 4.475s / 200, num_workers=8)
Relative speed: 0.550 (= 0.022s / 0.041s)

❌ OneFlow resnet50 time: 152.3ms (= 15228.7ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 160.8ms (= 16080.2ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
❌ Relative speed: 1.06 (= 160.8ms / 152.3ms)

OneFlow resnet50 time: 90.9ms (= 9088.1ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 101.0ms (= 10100.5ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.11 (= 101.0ms / 90.9ms)

OneFlow resnet50 time: 58.9ms (= 11778.6ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 79.8ms (= 15962.8ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.36 (= 79.8ms / 58.9ms)

OneFlow resnet50 time: 42.7ms (= 8530.6ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 78.7ms (= 15748.8ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.85 (= 78.7ms / 42.7ms)

OneFlow resnet50 time: 37.1ms (= 7429.8ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 67.9ms (= 13572.2ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.83 (= 67.9ms / 37.1ms)

@github-actions
Copy link
Contributor

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/9991/

@liujuncheng liujuncheng enabled auto-merge (squash) March 16, 2023 14:33
@github-actions
Copy link
Contributor

Speed stats: 
GPU Name: GeForce GTX 1080 

❌ OneFlow resnet50 time: 140.7ms (= 14074.6ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 143.4ms (= 14344.7ms / 100, input_shape=[16, 3, 224, 224])
❌ Relative speed: 1.02 (= 143.4ms / 140.7ms)

OneFlow resnet50 time: 80.4ms (= 8041.4ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 85.1ms (= 8505.7ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.06 (= 85.1ms / 80.4ms)

OneFlow resnet50 time: 48.5ms (= 9693.7ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 58.4ms (= 11670.7ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.20 (= 58.4ms / 48.5ms)

OneFlow resnet50 time: 32.2ms (= 6448.3ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 44.6ms (= 8924.0ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.38 (= 44.6ms / 32.2ms)

OneFlow resnet50 time: 25.7ms (= 5138.7ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 40.6ms (= 8113.5ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.58 (= 40.6ms / 25.7ms)

OneFlow swin dataloader time: 0.235s (= 46.960s / 200, num_workers=1)
PyTorch swin dataloader time: 0.150s (= 30.066s / 200, num_workers=1)
Relative speed: 0.640 (= 0.150s / 0.235s)

OneFlow swin dataloader time: 0.067s (= 13.478s / 200, num_workers=4)
PyTorch swin dataloader time: 0.043s (= 8.602s / 200, num_workers=4)
Relative speed: 0.638 (= 0.043s / 0.067s)

OneFlow swin dataloader time: 0.041s (= 8.181s / 200, num_workers=8)
PyTorch swin dataloader time: 0.022s (= 4.497s / 200, num_workers=8)
Relative speed: 0.550 (= 0.022s / 0.041s)

❌ OneFlow resnet50 time: 152.7ms (= 15265.1ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 162.0ms (= 16199.3ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
❌ Relative speed: 1.06 (= 162.0ms / 152.7ms)

OneFlow resnet50 time: 91.0ms (= 9097.4ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 101.4ms (= 10144.2ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.12 (= 101.4ms / 91.0ms)

OneFlow resnet50 time: 58.8ms (= 11761.2ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 87.3ms (= 17469.8ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.49 (= 87.3ms / 58.8ms)

OneFlow resnet50 time: 41.9ms (= 8387.8ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 74.1ms (= 14827.9ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.77 (= 74.1ms / 41.9ms)

OneFlow resnet50 time: 37.4ms (= 7477.5ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 68.0ms (= 13594.7ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.82 (= 68.0ms / 37.4ms)

@github-actions
Copy link
Contributor

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/9991/

@liujuncheng liujuncheng merged commit def0c7e into master Mar 16, 2023
@liujuncheng liujuncheng deleted the dev_att_var_seq_len branch March 16, 2023 16:23
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@jackalcooper jackalcooper jackalcooper approved these changes

@ofhwei ofhwei ofhwei approved these changes

@daquexian daquexian Awaiting requested review from daquexian

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

@BBuf BBuf Awaiting requested review from BBuf

@oneflow-ci-bot oneflow-ci-bot Awaiting requested review from oneflow-ci-bot

Assignees
No one assigned
Labels
automerge feature op
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
3 participants
@liujuncheng @jackalcooper @ofhwei