turn off dit_layerwise_offload for wan on rocm by zyzshishui · Pull Request #17569 · sgl-project/sglang

zyzshishui · 2026-01-22T08:49:15Z

Motivation

This pr is related to #16499. On MI350/MI355X, Wan models run faster with dit_layerwise_offload off.

Modifications

Turn off dit_layerwise_offload on rocm platform by default

Accuracy Tests

Benchmarking and Profiling

MI350

Model	1 GPU (on / off)	Speedup	8 GPUs (on / off)	Speedup
Wan2.2-TI2V-5B-Diffusers	53.52s / 34.17s	1.57×	32.10s / 7.69s	4.17×
Wan2.2-T2V-A14B-Diffusers	546.77s / 525.65s	1.04×	117.83s / 100.49s	1.17×

MI355X

Model	1 GPU (on / off)	Speedup	8 GPUs (on / off)	Speedup
Wan2.2-TI2V-5B-Diffusers	48.03s / 29.15s	1.65×	28.35s / 6.85s	4.14×
Wan2.2-T2V-A14B-Diffusers	457.21s / 444.15s	1.03×	101.26s / 82.88s	1.22x

Repro commands

Wan2.2-T2V-A14B (single GPU)

PROMPT="A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."

COMMON_ARGS="--log-level=info \
  --prompt=\"$PROMPT\" --negative-prompt=\" \" --720p \
  --num-inference-steps=40 --num-frames=81 --guidance-scale=5.0 --seed=42 \
  --save-output --warmup --enable-torch-compile true \
  --dit-cpu-offload false --text-encoder-cpu-offload false \
  --image-encoder-cpu-offload false --vae-cpu-offload false"

# on
sglang generate --model-path Wan-AI/Wan2.2-T2V-A14B-Diffusers $COMMON_ARGS --dit-layerwise-offload true
# off
sglang generate --model-path Wan-AI/Wan2.2-T2V-A14B-Diffusers $COMMON_ARGS --dit-layerwise-offload false

Wan2.2-T2V-A14B (8 GPUs)

PROMPT="A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."

COMMON_ARGS="--log-level=info \
  --prompt=\"$PROMPT\" --negative-prompt=\" \" --720p \
  --num-inference-steps=40 --num-frames=81 --guidance-scale=5.0 --seed=42 \
  --save-output --warmup --enable-torch-compile true \
  --num-gpus=8 --enable-cfg-parallel --ulysses-degree=4 \
  --dit-cpu-offload false --text-encoder-cpu-offload false \
  --image-encoder-cpu-offload false --vae-cpu-offload false"

# on
sglang generate --model-path Wan-AI/Wan2.2-T2V-A14B-Diffusers $COMMON_ARGS --dit-layerwise-offload true
# off
sglang generate --model-path Wan-AI/Wan2.2-T2V-A14B-Diffusers $COMMON_ARGS --dit-layerwise-offload false

Wan2.2-TI2V-5B (single GPU / 8 GPUs)

PROMPT="An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
NEG="Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
IMG="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"

COMMON_ARGS="--log-level info --warmup true --enable-torch-compile true \
  --prompt \"$PROMPT\" --negative-prompt \"$NEG\" --image-path $IMG \
  --num-frames 81 --num-inference-steps 50 --guidance-scale 5.0 --seed 42 \
  --save-output --720p \
  --dit-cpu-offload false --vae-cpu-offload false --text-encoder-cpu-offload false"

# single GPU: on/off
sglang generate --model-path Wan-AI/Wan2.2-TI2V-5B-Diffusers $COMMON_ARGS --dit-layerwise-offload true
sglang generate --model-path Wan-AI/Wan2.2-TI2V-5B-Diffusers $COMMON_ARGS --dit-layerwise-offload false

# 8 GPUs: on/off
sglang generate --model-path Wan-AI/Wan2.2-TI2V-5B-Diffusers $COMMON_ARGS \
  --num-gpus=8 --enable-cfg-parallel --ulysses-degree=4 --dit-layerwise-offload true
sglang generate --model-path Wan-AI/Wan2.2-TI2V-5B-Diffusers $COMMON_ARGS \
  --num-gpus=8 --enable-cfg-parallel --ulysses-degree=4 --dit-layerwise-offload false

Checklist

Format your code according to the Format code with pre-commit.
Add unit tests according to the Run and add unit tests.
Update documentation according to Write documentations.
Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
Follow the SGLang code style guidance.

Review Process

Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
Get approvals from CODEOWNERS and other reviewers.
Trigger CI tests with comments or contact authorized users to do so.
- /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
After green CI and required approvals, ask Merge Oncalls to merge.

gemini-code-assist · 2026-01-22T08:49:30Z

Summary of Changes

Hello @zyzshishui, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request implements a crucial performance enhancement for users running Wan models on ROCm-based hardware. By disabling the dit_layerwise_offload feature by default on ROCm, the system can achieve substantial speedups, particularly in multi-GPU configurations. This change streamlines the user experience by automatically applying an optimal configuration, removing the need for manual adjustments to achieve better inference performance.

Highlights

Performance Optimization for ROCm: The dit_layerwise_offload feature is now disabled by default for Wan models when running on ROCm platforms, based on benchmark results showing significant speed improvements.
Platform-Specific Configuration: A new class method, enable_dit_layerwise_offload_for_wan_by_default, has been introduced in the Platform interface, allowing specific platforms like ROCm to override the default behavior for this setting.
Dynamic Default Setting: The server argument checking logic has been updated to dynamically determine the default dit_layerwise_offload setting for Wan models based on the current platform's configuration.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist

Code Review

This pull request disables dit_layerwise_offload for Wan models on ROCm platforms by default to improve performance, which is supported by the provided benchmarks. The implementation is well-structured, introducing a new method in the Platform interface to handle platform-specific defaults and overriding it for ROCm. The logic is correctly applied in the server arguments check. The changes are clear and effective. I have one minor style suggestion to ensure code consistency.

zyzshishui · 2026-01-23T06:08:42Z

/tag-run-ci-label

turn off dit_layerwise_offload for wan on rocm

da5071b

zyzshishui requested review from mickqian and yhyang201 as code owners January 22, 2026 08:49

github-actions Bot added amd diffusion SGLang Diffusion labels Jan 22, 2026

gemini-code-assist Bot reviewed Jan 22, 2026

View reviewed changes

Comment thread python/sglang/multimodal_gen/runtime/platforms/rocm.py Outdated

lint

2666056

Merge branch 'main' into 1

6e991be

github-actions Bot added the run-ci label Jan 23, 2026

yhyang201 approved these changes Jan 23, 2026

View reviewed changes

mickqian approved these changes Jan 23, 2026

View reviewed changes

Fridge003 merged commit 2169025 into sgl-project:main Jan 23, 2026
80 of 87 checks passed

Fridge003 pushed a commit that referenced this pull request Jan 23, 2026

turn off dit_layerwise_offload for wan on rocm (#17569)

b6e4893

zyzshishui deleted the 1 branch January 23, 2026 20:36

Johnsonms pushed a commit to Johnsonms/sglang that referenced this pull request Feb 14, 2026

turn off dit_layerwise_offload for wan on rocm (sgl-project#17569)

4804d9a

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

turn off dit_layerwise_offload for wan on rocm#17569

turn off dit_layerwise_offload for wan on rocm#17569
Fridge003 merged 3 commits intosgl-project:mainfrom
zyzshishui:1

zyzshishui commented Jan 22, 2026

Uh oh!

gemini-code-assist Bot commented Jan 22, 2026

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

Uh oh!

zyzshishui commented Jan 23, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

zyzshishui commented Jan 22, 2026

Motivation

Modifications

Accuracy Tests

Benchmarking and Profiling

MI350

MI355X

Repro commands

Wan2.2-T2V-A14B (single GPU)

Wan2.2-T2V-A14B (8 GPUs)

Wan2.2-TI2V-5B (single GPU / 8 GPUs)

Checklist

Review Process

Uh oh!

gemini-code-assist Bot commented Jan 22, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

zyzshishui commented Jan 23, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants