Skip to content

fix autocast to support global tensor #10605

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
Open
wants to merge 5 commits into
base: master
Choose a base branch
from
Open

fix autocast to support global tensor #10605

wants to merge 5 commits into from

Conversation

@fpzh2011
Copy link
Contributor

oneflow autocast 不支持 global tensor。下面代码直接报错

import oneflow as flow

placement = flow.placement("cuda", ranks=[0])
sbp = flow.sbp.broadcast
a = flow.randn(2, 3).to_global(placement=placement, sbp=sbp)
b = flow.randn(3, 4).to_global(placement=placement, sbp=sbp)
with flow.autocast(device_type="cuda"):
    c = flow.matmul(a, b)

@github-actions
Copy link
Contributor

View latest API docs preview at: https://oneflow-staging.oss-cn-beijing.aliyuncs.com/docs/Oneflow-Inc/oneflow/pr/10605/

@github-actions
Copy link
Contributor

Speed stats: 
GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.2ms (= 4322.0ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 57.3ms (= 5733.8ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.33 (= 57.3ms / 43.2ms)

OneFlow resnet50 time: 26.5ms (= 2646.1ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 37.1ms (= 3710.4ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.40 (= 37.1ms / 26.5ms)

OneFlow resnet50 time: 17.8ms (= 3552.6ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 35.3ms (= 7052.8ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.99 (= 35.3ms / 17.8ms)

OneFlow resnet50 time: 15.5ms (= 3096.9ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 31.4ms (= 6279.5ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 2.03 (= 31.4ms / 15.5ms)

OneFlow resnet50 time: 15.0ms (= 3004.1ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 28.7ms (= 5731.9ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.91 (= 28.7ms / 15.0ms)

OneFlow swin dataloader time: 0.199s (= 39.777s / 200, num_workers=1)
PyTorch swin dataloader time: 0.128s (= 25.690s / 200, num_workers=1)
Relative speed: 0.646 (= 0.128s / 0.199s)

OneFlow swin dataloader time: 0.057s (= 11.333s / 200, num_workers=4)
PyTorch swin dataloader time: 0.033s (= 6.639s / 200, num_workers=4)
Relative speed: 0.586 (= 0.033s / 0.057s)

OneFlow swin dataloader time: 0.038s (= 7.618s / 200, num_workers=8)
PyTorch swin dataloader time: 0.022s (= 4.346s / 200, num_workers=8)
Relative speed: 0.571 (= 0.022s / 0.038s)

❌ OneFlow resnet50 time: 49.3ms (= 4926.9ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 65.5ms (= 6554.8ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.33 (= 65.5ms / 49.3ms)

OneFlow resnet50 time: 36.8ms (= 3684.1ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 46.5ms (= 4647.9ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.26 (= 46.5ms / 36.8ms)

OneFlow resnet50 time: 28.4ms (= 5670.9ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 39.9ms (= 7987.3ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.41 (= 39.9ms / 28.4ms)

OneFlow resnet50 time: 25.7ms (= 5141.0ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 39.6ms (= 7916.6ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.54 (= 39.6ms / 25.7ms)

OneFlow resnet50 time: 24.7ms (= 4945.4ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 37.4ms (= 7477.2ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.51 (= 37.4ms / 24.7ms)

@github-actions
Copy link
Contributor

View latest API docs preview at: https://oneflow-staging.oss-cn-beijing.aliyuncs.com/docs/Oneflow-Inc/oneflow/pr/10605/

@github-actions
Copy link
Contributor

Speed stats: 
GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.2ms (= 4321.6ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 57.5ms (= 5752.2ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.33 (= 57.5ms / 43.2ms)

OneFlow resnet50 time: 26.1ms (= 2614.6ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 38.1ms (= 3806.9ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.46 (= 38.1ms / 26.1ms)

OneFlow resnet50 time: 18.3ms (= 3653.3ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 36.4ms (= 7270.4ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.99 (= 36.4ms / 18.3ms)

OneFlow resnet50 time: 17.3ms (= 3463.7ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 31.6ms (= 6320.1ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.82 (= 31.6ms / 17.3ms)

OneFlow resnet50 time: 17.4ms (= 3488.9ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 29.2ms (= 5837.5ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.67 (= 29.2ms / 17.4ms)

OneFlow swin dataloader time: 0.200s (= 39.950s / 200, num_workers=1)
PyTorch swin dataloader time: 0.129s (= 25.731s / 200, num_workers=1)
Relative speed: 0.644 (= 0.129s / 0.200s)

OneFlow swin dataloader time: 0.056s (= 11.262s / 200, num_workers=4)
PyTorch swin dataloader time: 0.033s (= 6.504s / 200, num_workers=4)
Relative speed: 0.578 (= 0.033s / 0.056s)

OneFlow swin dataloader time: 0.030s (= 6.088s / 200, num_workers=8)
PyTorch swin dataloader time: 0.016s (= 3.294s / 200, num_workers=8)
Relative speed: 0.541 (= 0.016s / 0.030s)

❌ OneFlow resnet50 time: 49.4ms (= 4941.5ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 66.3ms (= 6634.7ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.34 (= 66.3ms / 49.4ms)

OneFlow resnet50 time: 36.4ms (= 3640.0ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 46.8ms (= 4683.3ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.29 (= 46.8ms / 36.4ms)

OneFlow resnet50 time: 27.8ms (= 5563.4ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 38.6ms (= 7726.3ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.39 (= 38.6ms / 27.8ms)

OneFlow resnet50 time: 25.0ms (= 4994.0ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 38.8ms (= 7756.7ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.55 (= 38.8ms / 25.0ms)

OneFlow resnet50 time: 25.3ms (= 5054.9ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 35.8ms (= 7163.3ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.42 (= 35.8ms / 25.3ms)

@github-actions
Copy link
Contributor

github-actions bot commented Mar 7, 2025

View latest API docs preview at: https://oneflow-staging.oss-cn-beijing.aliyuncs.com/docs/Oneflow-Inc/oneflow/pr/10605/

@github-actions
Copy link
Contributor

github-actions bot commented Mar 7, 2025

Speed stats: 
GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.2ms (= 4322.1ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 57.4ms (= 5740.0ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.33 (= 57.4ms / 43.2ms)

OneFlow resnet50 time: 26.7ms (= 2671.4ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 37.3ms (= 3732.3ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.40 (= 37.3ms / 26.7ms)

OneFlow resnet50 time: 18.6ms (= 3723.7ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 35.6ms (= 7126.1ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.91 (= 35.6ms / 18.6ms)

OneFlow resnet50 time: 17.4ms (= 3472.4ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 30.8ms (= 6162.1ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.77 (= 30.8ms / 17.4ms)

OneFlow resnet50 time: 16.6ms (= 3314.1ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 29.2ms (= 5836.9ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.76 (= 29.2ms / 16.6ms)

OneFlow swin dataloader time: 0.201s (= 40.267s / 200, num_workers=1)
PyTorch swin dataloader time: 0.129s (= 25.742s / 200, num_workers=1)
Relative speed: 0.639 (= 0.129s / 0.201s)

OneFlow swin dataloader time: 0.058s (= 11.592s / 200, num_workers=4)
PyTorch swin dataloader time: 0.032s (= 6.462s / 200, num_workers=4)
Relative speed: 0.557 (= 0.032s / 0.058s)

OneFlow swin dataloader time: 0.031s (= 6.103s / 200, num_workers=8)
PyTorch swin dataloader time: 0.017s (= 3.314s / 200, num_workers=8)
Relative speed: 0.543 (= 0.017s / 0.031s)

❌ OneFlow resnet50 time: 49.3ms (= 4927.0ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 66.2ms (= 6620.2ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.34 (= 66.2ms / 49.3ms)

OneFlow resnet50 time: 36.5ms (= 3647.2ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 46.4ms (= 4644.2ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.27 (= 46.4ms / 36.5ms)

OneFlow resnet50 time: 27.6ms (= 5526.9ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 39.6ms (= 7915.3ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.43 (= 39.6ms / 27.6ms)

OneFlow resnet50 time: 25.2ms (= 5040.0ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 38.8ms (= 7765.9ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.54 (= 38.8ms / 25.2ms)

OneFlow resnet50 time: 25.0ms (= 5000.5ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 36.2ms (= 7240.2ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.45 (= 36.2ms / 25.0ms)

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

Reviewers

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

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

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

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

Assignees

No one assigned

Labels

enhancement global op

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@fpzh2011 @ShawnXuan