[trainer] feat: support sequence parallel

[zhtmike]

What does this PR do?

• support sequence parallel for Diffusers SDPA attn backend
• Require diffusers >= 0.38

Memory:

Val Reward:

Performance:

Add concise overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: ...
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, veomni, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward, fully_async, one_step_off
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)
• If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.

[Copilot]

Pull request overview

Adds Ulysses Sequence Parallel (SP) support for Diffusers-based training in the FSDP engine, including runtime monkey-patches to work around current upstream Diffusers limitations.

Changes:

• Enable Ulysses SP device mesh initialization and Diffusers context-parallel enablement in DiffusersFSDPEngine.
• Introduce Diffusers monkey-patches (mesh shape, attention mask shape, attention backward) plus a helper to fix hook mesh wiring.
• Add distributed equivalence tests (fwd, fwd+bwd, and FSDP-wrapped) and an example training script enabling ulysses_sequence_parallel_size=2.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
verl/workers/engine/fsdp/diffusers_impl.py	Turns on Ulysses SP path in the Diffusers FSDP engine, pads seq-lens for SP alignment, and wires Diffusers context-parallel config/hooks.
verl/models/diffusers/monkey_patch.py	Adds monkey-patches and mesh-fix utilities needed for Diffusers Ulysses SP to work.
verl/models/diffusers/__init__.py	Exposes the monkey-patch entrypoint as a package API.
tests/models/test_diffusers_ulysses.py	Adds multi-GPU distributed tests checking SP vs non-SP forward/backward equivalence (plus FSDP-wrapped case).
examples/flowgrpo_trainer/run_flowgrpo_sp2.sh	Provides a runnable example config for training with ulysses_sequence_parallel_size=2.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.

[Copilot]

Pull request overview

Adds Ulysses sequence parallel (SP) support for Diffusers-based training, including device-mesh setup and input sequence alignment, plus accompanying distributed tests and an example launch script.

Changes:

• Enable Ulysses SP in DiffusersFSDPEngine by creating a (dp, ring, ulysses) device mesh, wiring the SP process group, and padding text embeddings to SP-aligned lengths.
• Update vLLM-Omni async server to use OmniEngineArgs for engine arg construction.
• Add distributed tests for Diffusers Ulysses SP forward/backward equivalence and an example script enabling ulysses_sequence_parallel_size=2.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.

File	Description
verl/workers/rollout/vllm_rollout/vllm_omni_async_server.py	Switches vLLM-Omni engine args type used to build the async engine config.
verl/workers/engine/fsdp/diffusers_impl.py	Implements Diffusers Ulysses SP wiring (mesh + group), enables parallelism, and pads text inputs to SP-friendly lengths.
tests/models/test_diffusers_ulysses.py	Adds distributed tests validating SP vs non-SP forward/backward behavior (including an FSDP-wrapped case).
examples/flowgrpo_trainer/run_flowgrpo_sp2.sh	Example FlowGRPO training invocation with Diffusers Ulysses SP size set to 2.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Same issue as above: ContextParallelConfig(..., mesh=ulysses_device_mesh) is not compatible with current released diffusers APIs and is likely to raise TypeError. It would be better to share a single helper to build the config in a version-tolerant way so all tests stay in sync with the production engine behavior.

[Copilot]

Same issue as above: ContextParallelConfig(..., mesh=ulysses_device_mesh) may not be a valid constructor call in released diffusers versions and can fail at import/runtime. Consider adding a small compatibility helper (or reusing the engine’s logic) so the test works across supported diffusers versions.

[Copilot]

ContextParallelConfig construction here passes mesh=..., but in current diffusers releases the public constructor does not accept a mesh kwarg (the mesh is handled internally / via different API). This is likely to raise a TypeError at runtime. To keep compatibility, consider using a version/feature-gated path (e.g., try constructing without mesh and set the mesh via the supported attribute/method, or rely on enable_parallelism() to create the mesh when possible).

Suggested change

	module.enable_parallelism(
	config=ContextParallelConfig(ulysses_degree=sp_size, mesh=self.ulysses_device_mesh)
	)
	cp_config = ContextParallelConfig(ulysses_degree=sp_size)
	# Some diffusers versions do not accept `mesh` in the constructor;
	# set it via attribute when available for compatibility.
	if hasattr(cp_config, "mesh"):
	cp_config.mesh = self.ulysses_device_mesh
	module.enable_parallelism(config=cp_config)

[Copilot]

ContextParallelConfig(ulysses_degree=..., mesh=...) assumes the constructor accepts a mesh kwarg, which is not true for current diffusers releases. This will make the test error before it can run. Consider version/feature-gating this and constructing the config using the API supported by released diffusers (or mirroring whatever compatibility shim is used in the engine).

[niehen6174]

I wasn’t able to find your contact email, so I’m reaching out here.

I’m interested in learning more about the technical details and discussing contributions.

My email is 1639206518@qq.com. I’d greatly appreciate it if you could reach out. I’d be happy to connect and discuss further.