test: add Krea NFT async OCR e2e - #243
Draft
zhihengy wants to merge 3 commits into
Draft
Conversation
zhihengy
marked this pull request as ready for review
September 11, 2026 18:11
zhihengy
marked this pull request as draft
September 11, 2026 18:23
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Adds the Krea-2-Raw DiffusionNFT OCR e2e and its four-rollout metric standard, stacked on #232 (
feat/async-training). The canonical async recipe uses two FSDP training GPUs and two separate rollout GPUs; four rollouts cover the initial EMA batches, updated rollouts, and training against lagged EMA references. Registered instage-c-5-gpu-h200with the existinge2elabel and strict metrics registry. The test launchesscripts/run_diffusion_nft_krea2_async.py --reward ocr --num-rollout 4.Recorded locally on four H200 GPUs, then independently rerun through the normal e2e check: all 10 metric series / 40 values matched exactly, both commands exited 0. Each rollout contains 64 samples; the recipe retains 1024px, 10 denoising steps, BF16, and LoRA rank 32. Pre-commit and registry discovery passed; GitHub pre-commit passed too.
The local run used cached model snapshot
4ad9f4b627a647fad78b3dfeebb09f2654aeb494, overriding only its location viaMILES_SCRIPT_EXTRA_ARGS. Environment: PyTorch 2.11.0+cu129, CUDA 12.9, NCCL 2.28.9, FA3 3.0.0b1, Python 3.11.15, Ray 2.53.0, PaddleOCR 2.9.1, diffusersf53d552036a0d1bd5570782a39cd40cfabf112bc.CI runner verification remains pending: the runner recording attempt failed before training with Hugging Face 403 / GatedRepoError. The runner needs Krea access or a pre-provisioned matching snapshot; cross-host agreement is not yet verified. PR CI also requires the repository's trusted approval or a maintainer-applied
run-ci-*label.The unchanged standard was also validated against the consolidated async + NCCL implementation in #232 (
34209bedfbc500abedf043529bb8b40e2eb48b17) with SGLang #38540 (f25d5bba56a20bb80f606bc58424d1fc845f327b): all four full rollouts completed and all 40 values matched exactly. The CI override below selects the NCCL-capable runtime required by the updated base.The test now uses the standalone async launcher from #232. CI discovery and mocked registry-to-CLI dispatch passed, as did pre-commit. Its training arguments match the previously validated async launcher; the metric standard is unchanged. GPU training was not rerun for this launcher extraction.
ci-sglang-pr: #38540