Thanks to visit codestin.com
Credit goes to github.com

Skip to content

adaptation_of_mlu_device_based_on_refactor_xccl_primitive#10647

Merged
fpzh2011 merged 7 commits into
masterfrom
adaptation_of_mlu_device_based_on_refactor_xccl_primitive
Jul 31, 2025
Merged

adaptation_of_mlu_device_based_on_refactor_xccl_primitive#10647
fpzh2011 merged 7 commits into
masterfrom
adaptation_of_mlu_device_based_on_refactor_xccl_primitive

Conversation

@Flowingsun007
Copy link
Copy Markdown
Contributor

@Flowingsun007 Flowingsun007 commented Jun 16, 2025

No description provided.

@ShawnXuan ShawnXuan requested a review from oneflow-ci-bot June 18, 2025 06:19
@Flowingsun007 Flowingsun007 marked this pull request as ready for review June 24, 2025 03:09
@Flowingsun007
support num_accumulation_steps
dfcf553
@Flowingsun007
Merge branch 'adaptation_of_mlu_device_based_on_refactor_xccl_primiti…
367bc9d 
…ve' of github.com:Oneflow-Inc/oneflow into adaptation_of_mlu_device_based_on_refactor_xccl_primitive
@github-actions
Copy link
Copy Markdown
Contributor

github-actions Bot commented Jul 1, 2025

View latest API docs preview at: https://oneflow-staging.oss-cn-beijing.aliyuncs.com/docs/Oneflow-Inc/oneflow/pr/10647/

@github-actions
Copy link
Copy Markdown
Contributor

github-actions Bot commented Jul 1, 2025

Speed stats: 
GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.7ms (= 4367.7ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 57.6ms (= 5755.3ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.32 (= 57.6ms / 43.7ms)

OneFlow resnet50 time: 26.4ms (= 2635.9ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 38.2ms (= 3817.1ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.45 (= 38.2ms / 26.4ms)

OneFlow resnet50 time: 18.8ms (= 3767.9ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 35.7ms (= 7147.5ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.90 (= 35.7ms / 18.8ms)

OneFlow resnet50 time: 17.5ms (= 3504.6ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 33.3ms (= 6667.9ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.90 (= 33.3ms / 17.5ms)

OneFlow resnet50 time: 17.6ms (= 3518.2ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 28.9ms (= 5776.1ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.64 (= 28.9ms / 17.6ms)

OneFlow swin dataloader time: 0.212s (= 42.469s / 200, num_workers=1)
PyTorch swin dataloader time: 0.129s (= 25.776s / 200, num_workers=1)
Relative speed: 0.607 (= 0.129s / 0.212s)

OneFlow swin dataloader time: 0.054s (= 10.877s / 200, num_workers=4)
PyTorch swin dataloader time: 0.032s (= 6.423s / 200, num_workers=4)
Relative speed: 0.591 (= 0.032s / 0.054s)

OneFlow swin dataloader time: 0.031s (= 6.268s / 200, num_workers=8)
PyTorch swin dataloader time: 0.017s (= 3.321s / 200, num_workers=8)
Relative speed: 0.530 (= 0.017s / 0.031s)

❌ OneFlow resnet50 time: 49.3ms (= 4932.3ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 65.8ms (= 6580.3ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.33 (= 65.8ms / 49.3ms)

OneFlow resnet50 time: 37.0ms (= 3702.9ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 48.2ms (= 4818.4ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.30 (= 48.2ms / 37.0ms)

OneFlow resnet50 time: 29.6ms (= 5927.8ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 40.3ms (= 8057.5ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.36 (= 40.3ms / 29.6ms)

OneFlow resnet50 time: 26.5ms (= 5293.4ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 41.3ms (= 8251.4ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.56 (= 41.3ms / 26.5ms)

OneFlow resnet50 time: 25.6ms (= 5116.3ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 35.9ms (= 7179.4ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.40 (= 35.9ms / 25.6ms)

@github-actions
Copy link
Copy Markdown
Contributor

github-actions Bot commented Jul 1, 2025

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.

Flowingsun007
Flowingsun007 commented Jul 1, 2025
View reviewed changes
Comment thread cmake/oneflow.cmake Outdated
get_directory_property(EXISTING_DEFS COMPILE_DEFINITIONS)

if(NOT "WITH_DEVICES" IN_LIST EXISTING_DEFS)
add_definitions(-DWITH_DEVICES)
Copy link
Copy Markdown
Contributor Author

@Flowingsun007 Flowingsun007 Jul 1, 2025

Choose a reason for hiding this comment

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

当编译npu/xpu/mlu等设备时，DEVICES_ENABLED即为True，会定义一一个WITH_DEVICES的宏表示多设备支持，后续在代码中方便统一管理。

// before:
#if defined(WITH_CUDA) || defined(WITH_NPU) || defined(WITH_MLU) || defined(WITH_XPU) ...

// after: 
#if defined(WITH_CUDA) || defined(WITH_DEVICES)

@github-actions
Copy link
Copy Markdown
Contributor

github-actions Bot commented Jul 1, 2025

View latest API docs preview at: https://oneflow-staging.oss-cn-beijing.aliyuncs.com/docs/Oneflow-Inc/oneflow/pr/10647/

@github-actions
Copy link
Copy Markdown
Contributor

github-actions Bot commented Jul 1, 2025

Speed stats: 
GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.3ms (= 4329.4ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 57.5ms (= 5750.3ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.33 (= 57.5ms / 43.3ms)

OneFlow resnet50 time: 26.6ms (= 2657.4ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 37.1ms (= 3712.1ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.40 (= 37.1ms / 26.6ms)

OneFlow resnet50 time: 18.6ms (= 3712.5ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 35.2ms (= 7037.2ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.90 (= 35.2ms / 18.6ms)

OneFlow resnet50 time: 17.6ms (= 3520.2ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 31.3ms (= 6264.6ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.78 (= 31.3ms / 17.6ms)

OneFlow resnet50 time: 17.4ms (= 3484.8ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 29.6ms (= 5921.1ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.70 (= 29.6ms / 17.4ms)

OneFlow swin dataloader time: 0.199s (= 39.804s / 200, num_workers=1)
PyTorch swin dataloader time: 0.128s (= 25.570s / 200, num_workers=1)
Relative speed: 0.642 (= 0.128s / 0.199s)

OneFlow swin dataloader time: 0.057s (= 11.474s / 200, num_workers=4)
PyTorch swin dataloader time: 0.032s (= 6.410s / 200, num_workers=4)
Relative speed: 0.559 (= 0.032s / 0.057s)

OneFlow swin dataloader time: 0.030s (= 6.022s / 200, num_workers=8)
PyTorch swin dataloader time: 0.017s (= 3.342s / 200, num_workers=8)
Relative speed: 0.555 (= 0.017s / 0.030s)

❌ OneFlow resnet50 time: 49.7ms (= 4974.6ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 66.1ms (= 6610.2ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.33 (= 66.1ms / 49.7ms)

OneFlow resnet50 time: 37.6ms (= 3762.8ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 45.5ms (= 4550.5ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.21 (= 45.5ms / 37.6ms)

OneFlow resnet50 time: 27.9ms (= 5584.7ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 40.3ms (= 8051.5ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.44 (= 40.3ms / 27.9ms)

OneFlow resnet50 time: 25.2ms (= 5034.6ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 38.2ms (= 7643.8ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.52 (= 38.2ms / 25.2ms)

OneFlow resnet50 time: 24.8ms (= 4956.2ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 36.0ms (= 7197.6ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.45 (= 36.0ms / 24.8ms)

@fpzh2011 fpzh2011 merged commit f485d7e into master Jul 31, 2025
20 checks passed
@fpzh2011 fpzh2011 deleted the adaptation_of_mlu_device_based_on_refactor_xccl_primitive branch July 31, 2025 07:54
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@fpzh2011 fpzh2011 fpzh2011 approved these changes

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

@jackalcooper jackalcooper Awaiting requested review from jackalcooper

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

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

Assignees

No one assigned

Labels

enhancement feature system

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@Flowingsun007 @fpzh2011 @ShawnXuan @oneflow-ci-bot