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

Reduce layer norm template size #8209

Merged
merged 10 commits into from
May 16, 2022
Merged

Reduce layer norm template size #8209

merged 10 commits into from
May 16, 2022

Conversation

guo-ran
Copy link
Contributor

@guo-ran guo-ran commented May 13, 2022

减少layer_norm特化，不特化不常出现的col_size

@guo-ran guo-ran requested a review from liujuncheng as a code owner May 13, 2022 10:27
@MARD1NO MARD1NO requested review from oneflow-ci-bot and removed request for oneflow-ci-bot May 14, 2022 02:05
@github-actions
Copy link
Contributor

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/8209/

@github-actions
Copy link
Contributor

Speed stats: 
GPU Name: NVIDIA GeForce GTX 1080 

❌ OneFlow resnet50 time: 129.3ms (= 12931.6ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 142.9ms (= 14287.2ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.10 (= 142.9ms / 129.3ms)

OneFlow resnet50 time: 77.5ms (= 7750.9ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 84.5ms (= 8445.8ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.09 (= 84.5ms / 77.5ms)

OneFlow resnet50 time: 54.8ms (= 10952.2ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 62.6ms (= 12526.4ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.14 (= 62.6ms / 54.8ms)

OneFlow resnet50 time: 43.0ms (= 8602.8ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 40.3ms (= 8068.0ms / 200, input_shape=[2, 3, 224, 224])
❌ Relative speed: 0.94 (= 40.3ms / 43.0ms)

OneFlow resnet50 time: 38.0ms (= 7593.5ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 38.9ms (= 7778.1ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.02 (= 38.9ms / 38.0ms)

OneFlow swin dataloader time: 0.387s (= 77.322s / 200, num_workers=1)
PyTorch swin dataloader time: 0.151s (= 30.100s / 200, num_workers=1)
Relative speed: 0.389 (= 0.151s / 0.387s)

OneFlow swin dataloader time: 0.065s (= 12.985s / 200, num_workers=4)
PyTorch swin dataloader time: 0.042s (= 8.407s / 200, num_workers=4)
Relative speed: 0.647 (= 0.042s / 0.065s)

OneFlow swin dataloader time: 0.037s (= 7.461s / 200, num_workers=8)
PyTorch swin dataloader time: 0.022s (= 4.460s / 200, num_workers=8)
Relative speed: 0.598 (= 0.022s / 0.037s)

❌ OneFlow resnet50 time: 144.6ms (= 14462.7ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 168.6ms (= 16861.0ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.17 (= 168.6ms / 144.6ms)

OneFlow resnet50 time: 97.7ms (= 9765.4ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 113.0ms (= 11303.4ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.16 (= 113.0ms / 97.7ms)

OneFlow resnet50 time: 73.0ms (= 14603.9ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 86.3ms (= 17255.0ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.18 (= 86.3ms / 73.0ms)

OneFlow resnet50 time: 63.8ms (= 12761.0ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 74.1ms (= 14812.3ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
❌ Relative speed: 1.16 (= 74.1ms / 63.8ms)

OneFlow resnet50 time: 57.6ms (= 11525.5ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 69.3ms (= 13868.6ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.20 (= 69.3ms / 57.6ms)

@github-actions
Copy link
Contributor

CI failed when running job: cuda-misc. PR label automerge has been removed

@github-actions
Copy link
Contributor

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/8209/

@github-actions
Copy link
Contributor

Speed stats: 
GPU Name: GeForce GTX 1080 

✔️ OneFlow resnet50 time: 128.6ms (= 12858.1ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 146.6ms (= 14661.2ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.14 (= 146.6ms / 128.6ms)

OneFlow resnet50 time: 79.0ms (= 7898.5ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 85.1ms (= 8512.0ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.08 (= 85.1ms / 79.0ms)

OneFlow resnet50 time: 51.5ms (= 10304.0ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 58.6ms (= 11710.6ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.14 (= 58.6ms / 51.5ms)

OneFlow resnet50 time: 42.0ms (= 8401.8ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 45.5ms (= 9100.0ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.08 (= 45.5ms / 42.0ms)

OneFlow resnet50 time: 37.5ms (= 7494.5ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 39.8ms (= 7960.3ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.06 (= 39.8ms / 37.5ms)

OneFlow swin dataloader time: 0.396s (= 79.200s / 200, num_workers=1)
PyTorch swin dataloader time: 0.152s (= 30.379s / 200, num_workers=1)
Relative speed: 0.384 (= 0.152s / 0.396s)

OneFlow swin dataloader time: 0.064s (= 12.806s / 200, num_workers=4)
PyTorch swin dataloader time: 0.040s (= 8.088s / 200, num_workers=4)
Relative speed: 0.632 (= 0.040s / 0.064s)

OneFlow swin dataloader time: 0.036s (= 7.126s / 200, num_workers=8)
PyTorch swin dataloader time: 0.022s (= 4.403s / 200, num_workers=8)
Relative speed: 0.618 (= 0.022s / 0.036s)

❌ OneFlow resnet50 time: 137.7ms (= 13771.2ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 160.2ms (= 16015.5ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.16 (= 160.2ms / 137.7ms)

OneFlow resnet50 time: 86.2ms (= 8615.4ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 100.9ms (= 10085.5ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.17 (= 100.9ms / 86.2ms)

OneFlow resnet50 time: 60.7ms (= 12130.8ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 79.3ms (= 15865.5ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.31 (= 79.3ms / 60.7ms)

OneFlow resnet50 time: 50.0ms (= 10005.3ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 69.7ms (= 13945.4ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.39 (= 69.7ms / 50.0ms)

OneFlow resnet50 time: 48.7ms (= 9735.3ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 67.8ms (= 13555.1ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.39 (= 67.8ms / 48.7ms)

@github-actions
Copy link
Contributor

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/8209/

@github-actions
Copy link
Contributor

CI failed when running job: cpu-misc. PR label automerge has been removed

@github-actions
Copy link
Contributor

Speed stats:

@github-actions
Copy link
Contributor

CI failed when running job: cuda-speed-test. PR label automerge has been removed

@github-actions
Copy link
Contributor

Speed stats: 
GPU Name: NVIDIA GeForce GTX 1080 

❌ OneFlow resnet50 time: 129.9ms (= 12988.6ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 144.3ms (= 14431.6ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.11 (= 144.3ms / 129.9ms)

OneFlow resnet50 time: 77.0ms (= 7704.0ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 84.3ms (= 8428.8ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.09 (= 84.3ms / 77.0ms)

OneFlow resnet50 time: 51.8ms (= 10357.6ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 53.7ms (= 10737.9ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.04 (= 53.7ms / 51.8ms)

OneFlow resnet50 time: 41.6ms (= 8314.2ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 43.4ms (= 8687.6ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.04 (= 43.4ms / 41.6ms)

OneFlow resnet50 time: 37.8ms (= 7552.6ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 41.1ms (= 8226.5ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.09 (= 41.1ms / 37.8ms)

OneFlow swin dataloader time: 0.241s (= 48.190s / 200, num_workers=1)
PyTorch swin dataloader time: 0.149s (= 29.822s / 200, num_workers=1)
Relative speed: 0.619 (= 0.149s / 0.241s)

OneFlow swin dataloader time: 0.064s (= 12.728s / 200, num_workers=4)
PyTorch swin dataloader time: 0.039s (= 7.846s / 200, num_workers=4)
Relative speed: 0.616 (= 0.039s / 0.064s)

OneFlow swin dataloader time: 0.039s (= 7.809s / 200, num_workers=8)
PyTorch swin dataloader time: 0.023s (= 4.551s / 200, num_workers=8)
Relative speed: 0.583 (= 0.023s / 0.039s)

❌ OneFlow resnet50 time: 147.1ms (= 14710.6ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 169.4ms (= 16939.8ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.15 (= 169.4ms / 147.1ms)

OneFlow resnet50 time: 96.2ms (= 9617.2ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 112.8ms (= 11282.4ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.17 (= 112.8ms / 96.2ms)

OneFlow resnet50 time: 74.0ms (= 14795.6ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 88.9ms (= 17779.2ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.20 (= 88.9ms / 74.0ms)

OneFlow resnet50 time: 61.5ms (= 12301.7ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 74.7ms (= 14932.5ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.21 (= 74.7ms / 61.5ms)

OneFlow resnet50 time: 54.8ms (= 10958.9ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 68.2ms (= 13640.9ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.24 (= 68.2ms / 54.8ms)

@github-actions
Copy link
Contributor

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/8209/

@guo-ran guo-ran merged commit 2f68b46 into master May 16, 2022
@guo-ran guo-ran deleted the dev_reduce_layer_norm_template_size branch May 16, 2022 02:56
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@liujuncheng liujuncheng liujuncheng approved these changes

@MARD1NO MARD1NO MARD1NO approved these changes

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

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

Successfully merging this pull request may close these issues.
3 participants
@guo-ran @liujuncheng @MARD1NO