[recipe, doc] feat: add Qwen-Image Flow-DPPO v1 sync recipe - #28
Open
knlnguyen1802 wants to merge 1 commit into
Open
[recipe, doc] feat: add Qwen-Image Flow-DPPO v1 sync recipe#28knlnguyen1802 wants to merge 1 commit into
knlnguyen1802 wants to merge 1 commit into
Conversation
Add a mechanical v1 sync recipe so Qwen-Image Flow-DPPO can opt into TransferQueue without changing adapters or the loss. This is recipe-only: shared v1 trainer infra already landed in verl-project#296/verl-project#315. Tracking verl-project#389. AI assistance (Cursor) was used for this change. Signed-off-by: knlnguyen1802 <knlnguyen1802@gmail.com> Co-authored-by: Cursor <cursoragent@cursor.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
examples/flowdppo_trainer/qwen_image/run_qwen_image_ocr_lora_v1.sh, the v1 sync counterpart of the Qwen-Image Flow-DPPO OCR LoRA recipe.python3 -m verl_omni.trainer.main_diffusion_v1,trainer.use_v1=true,trainer.v1.trainer_mode=sync, andtrainer.experiment_name=qwen_image_ocr_lora_v1.Tracking: verl-project#389
Not duplicating existing PRs
Duplicate-work checks (2026-09-03):
gh issue view 389 --repo verl-project/verl-omni --comments— Flow-DPPO is still in Need help; comments claim BAGEL ([recipe, tests, doc] feat: add BAGEL FlowGRPO V1 sync recipe verl-project/verl-omni#511) and MixGRPO only.gh pr list --repo verl-project/verl-omni --state open --search "389 in:body"— [4/N][recipe, doc] feat: add Qwen-Image FlowGRPO v1 sync recipe verl-project/verl-omni#489, [recipe, tests, doc] feat: add BAGEL FlowGRPO V1 sync recipe verl-project/verl-omni#511, Add V1 sync recipe for Qwen-Image + MixGRPO verl-project/verl-omni#512, [recipe, doc] feat: add MiniMax H3 FlowGRPO V1 sync recipes verl-project/verl-omni#501, [trainer, tests, doc] feat: async teacher scheduling for diffusion OPD on v1 separate_async verl-project/verl-omni#495; none are Flow-DPPO.gh pr list --repo verl-project/verl-omni --state open --search "flowdppo v1"— empty.gh pr list --repo knlnguyen1802/verl-omni --state open --search "flowdppo v1"— empty.This is not a duplicate of #296 or #466: those landed the shared diffusion v1 trainer and the Wan DanceGRPO recipe. This PR is a per-algorithm recipe that opts Qwen-Image Flow-DPPO into that infra in sync mode. Adapters and
FlowDPPOLossare untouched.Earlier fallbacks were skipped because they were already claimed (BAGEL verl-project#511 / fork #23). Flow-DPPO was the first unclaimed item in the assigned order.
Test plan
bash -n examples/flowdppo_trainer/qwen_image/run_qwen_image_ocr_lora_v1.sh(pass)trainer.use_v1=true,trainer.v1.trainer_mode=sync, experiment name suffix_v1;loss_mode=flow_dppoand remaining Hydra overrides match v0; noseparate_asyncPR_TITLE='[recipe, doc] feat: add Qwen-Image Flow-DPPO v1 sync recipe' python tests/special_sanity/check_pr_title.py(title valid; Windows console only failed to print the checkmark glyph)AI assistance
Cursor (Grok) was used to draft the recipe script, README note, commit, and this PR. A human submitter must review every changed line before merge; this is not a pure code-agent PR.