[Bugfix] Enable step-wise execution by knlnguyen1802 · Pull Request #81 · verl-project/verl-omni

knlnguyen1802 · 2026-05-14T08:18:45Z

What does this PR do?

Temporary fix to enable stepwise execution for vllm-omni

A temporary fix to enable stepwise execution

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, vllm_omni, rollout, trainer, ci, training_utils, recipe, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward, diffusion, omni, tests, docker
- If this PR involves multiple modules, separate them with , like [diffusion, 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][diffusion, fsdp] feat: new rollout scheduler

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: ...

gemini-code-assist

Code Review

This pull request introduces support for pre-tokenized prompt IDs and step-wise execution within the vllm_omni_rollout_adapter.py. Key additions include the _extract_prompt_ids, _tokenize_text_prompt, and prepare_encode methods, which facilitate the initialization of diffusion states from tokenized inputs. Additionally, the async server configuration now enables step_execution and sets a maximum sequence limit. Feedback indicates that since step_execution is enabled, the adapter must also override the per-step execution methods to ensure RL-specific fields are correctly collected. Other suggestions include correcting a class name typo in a ValueError and utilizing the build_img_shapes helper function to reduce code duplication.

gemini-code-assist · 2026-05-14T08:21:29Z

+        ).to(self.device)
+        return tokens.input_ids, tokens.attention_mask
+
+    def prepare_encode(


To fully support step-wise execution for RL rollouts, this class likely needs to override the per-step execution method (e.g., execute_step or step). The current implementation overrides diffuse to collect all_log_probs and all_latents during the denoising loop. If step_execution is enabled in the engine, the engine will bypass diffuse and call the per-step method instead. Without an override that performs similar data collection and state updates (like incrementing state.step_index), these RL-specific fields will be missing from the final output.

gemini-code-assist · 2026-05-14T08:21:29Z

+            raise ValueError(
+                "QwenImagePipelineWithLogProbForTest.prepare_encode requires either "
+                "'prompt_ids' or a text 'prompt' in state.prompts[0]."
+            )


The error message contains a typo in the class name, referring to QwenImagePipelineWithLogProbForTest instead of QwenImagePipelineWithLogProb.

Suggested change

raise ValueError(

"QwenImagePipelineWithLogProbForTest.prepare_encode requires either "

"'prompt_ids' or a text 'prompt' in state.prompts[0]."

)

raise ValueError(

"QwenImagePipelineWithLogProb.prepare_encode requires either "

"'prompt_ids' or a text 'prompt' in state.prompts[0]."

)

gemini-code-assist · 2026-05-14T08:21:30Z

+            None,
+        )
+
+        img_shapes = [[(1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2)]] * batch_size


This logic for building img_shapes is already implemented in the build_img_shapes utility function in common.py. It is better to use the utility to avoid code duplication and ensure consistency.

Suggested change

img_shapes = [[(1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2)]] * batch_size

img_shapes = build_img_shapes(height, width, batch_size, self.vae_scale_factor)

References

Avoid code duplication by reusing existing helper functions for common logic, such as constructing image shapes.

Signed-off-by: knlnguyen1802 <knlnguyen1802@gmail.com>

zhtmike · 2026-05-15T10:29:37Z

I don't really understand why we need to add so many things to make it work..

The main problem is vllm-omni does not official support passing prompt_token_ids as input.
The main logic of pipeline still try to tokenize the prompt.
That why the custom pipeline need to align with new function prepare_encode added by step-wise function.
To make this stable compatible, support from vllm-omni side is better

I see. So vllm-omni does not support truly --skip-tokenizer-init in https://docs.vllm.ai/en/stable/configuration/engine_args/#modelconfig to accept prompt_token_ids.

Can we make a feature request for this?

Yes, I will work on it

FYI @SamitHuang

does it conflict with the change with #66? which rename prompt_ids → prompt_token_ids for vllm-omni 0.20+

fix step batch rollout

SamitHuang · 2026-05-21T03:35:54Z

Consider compatibility for other algorithms beside FlowGRPO. Currently the training on mixgrpo is not converging well (reward mean increase slower than previsou, it seems this PR diables sde window in mixgrpo

SamitHuang · 2026-05-21T03:59:12Z

            engine_args["custom_pipeline_args"] = {"pipeline_class": pipeline_path}

+        engine_args["max_num_seqs"] = 256
+        engine_args["step_execution"] = True


why hard-coded step_execution=True?
so as max_num_seqs

we should make step-wise continuous batching a configurable option, and default is False (since only qwen-image supports for now)

yes, it is a patch PR for quick validation. there will be a formal PR later. @knlnguyen1802

SamitHuang · 2026-05-21T04:06:12Z

Bug: step_execution=True silently bypasses MixGRPO's window-positioning logic

This PR sets engine_args["step_execution"] = True unconditionally in verl_omni/workers/rollout/vllm_rollout/vllm_omni_async_server.py:111, which makes
QwenImageMixGRPOPipelineWithLogProb.forward() unreachable. As a result, _maybe_make_progressive_window never runs, and rollouts within a single training step end up sampling SDE on
different windows — violating MixGRPO's invariant that all rollouts in a batch share one window for advantage estimation.

Signed-off-by: knlnguyen1802 <knlnguyen1802@gmail.com>

knlnguyen1802 requested a review from SamitHuang as a code owner May 14, 2026 08:18

knlnguyen1802 closed this May 14, 2026

knlnguyen1802 reopened this May 14, 2026

gemini-code-assist Bot reviewed May 14, 2026

View reviewed changes

[Bugfix] Enable step-wise execution

3c5b5da

Signed-off-by: knlnguyen1802 <knlnguyen1802@gmail.com>

zhtmike reviewed May 15, 2026

View reviewed changes

zhtmike and others added 3 commits May 18, 2026 09:58

Merge branch 'main' into stepwise_fix

e6712fc

fix step batch rollout

628d43a

Merge pull request #1 from zhtmike/pr81

eefca1b

fix step batch rollout

SamitHuang reviewed May 21, 2026

View reviewed changes

knlnguyen1802 added 2 commits May 21, 2026 15:10

Fix MixGRPO

101ec56

Signed-off-by: knlnguyen1802 <knlnguyen1802@gmail.com>

Rebase with main

a8f4dd7

Signed-off-by: knlnguyen1802 <knlnguyen1802@gmail.com>

	img_shapes = [[(1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2)]] * batch_size
	img_shapes = build_img_shapes(height, width, batch_size, self.vae_scale_factor)

Conversation

knlnguyen1802 commented May 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What does this PR do?

Checklist Before Starting

Test

API and Usage Example

Design & Code Changes

Checklist Before Submitting

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot May 14, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot May 14, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot May 14, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SamitHuang May 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SamitHuang commented May 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

knlnguyen1802 commented May 14, 2026 •

edited

Loading

SamitHuang May 21, 2026 •

edited

Loading