[recipe] feat: add Qwen-Image / MiniMax H3 DiffusionNFT v1 sync recipes - #29
Open
knlnguyen1802 wants to merge 2 commits into
Open
[recipe] feat: add Qwen-Image / MiniMax H3 DiffusionNFT v1 sync recipes#29knlnguyen1802 wants to merge 2 commits into
knlnguyen1802 wants to merge 2 commits into
Conversation
Add mechanical v1 sync launchers so DiffusionNFT can opt into TransferQueue without changing adapters or losses. 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
main_diffusion→main_diffusion_v1,trainer.use_v1=true,trainer.v1.trainer_mode=sync.examples/dancegrpo_trainer/wan22/run_wan22_5b_t2v_hpsv3_v1.sh) and the SD3.5 twinexamples/flowgrpo_trainer/sd35/run_sd35_medium_ocr_lora_v1.sh. GPU layout stays colocated (same as each v0 sibling).*_v1.shopt-in launchers. Adapters, losses, and trainer Python are unchanged.Not a duplicate
Refreshed 2026-09-03:
gh issue view 389 --repo verl-project/verl-omni— DiffusionNFT is still in Need help (unclaimed).gh pr list --repo verl-project/verl-omni --state open --search "389 in:body"gh pr list --repo verl-project/verl-omni --state open --search "v1"gh pr list --repo verl-project/verl-omni --state open --search "DiffusionNFT"No open
verl-project/verl-omniPR owns this (model, algorithm, mode) path. verl-project#501 is MiniMax H3 FlowGRPO, not DiffusionNFT.Test plan
*_v1.shagainst its v0 sibling: only header comment,main_diffusion_v1,trainer.use_v1=true,trainer.v1.trainer_mode=sync, and distinctexperiment_name/WANDB_RUN_ID(*_v1).run_qwen_image_ocr_lora_v1.sh(and MiniMax siblings when checkpoints are available). This is not a flagship default-switch, so v0-vs-v1 curves are not attached.Note on trainer coverage
v1 sync uses
PolicyGradientDiffusionTrainerV1Sync(same entry as verl-project#296 / verl-project#466). v0 DiffusionNFT still goes throughDirectPreferenceRayTrainer(old-adapter copy/EMA). This PR does not add a second trainer. If old-policy refresh is missing on the v1 control plane, that is a follow-up trainer item, not a recipe-fork.AI assistance
AI assistance (Cursor) was used for this change. A human submitter has reviewed every changed line.