v0.4.1

What's Changed

• Upstream core change sync from vlm project by @jcao-ai in #554
• Fix potential teacher model weight loading error. by @foreverlms in #560
• Fix checkpoint offload and loading to CPU by @YurongYou in #562
• Support uncentralized mode for colocated training by @Dinghow in #561
• fix: Fix HF model tie case handling by @lfengad in #565
• Support set primary adapter in multi-lora by @Dinghow in #566
• fix upload issue for sft trainer by @littlespray in #564
• feat: save safetensor async by @lfengad in #569
• fix: Refine custom loader usage by @lfengad in #571
• Add micro_batch & mini_batch for Diffusion RL by @Dinghow in #570
• Support Load-Balanced Dynamic Batching by @kane-vln in #522
• slurm launcher - improve the discovery of cosmos_rl path by @vinjn in #573
• Fix WFM remote reward async mode by @Dinghow in #574
• [vla] support robotwin by @fwd4 in #575
• More metrics for Diffusion RL by @Dinghow in #572
• fix: try simplify for large scale communication by @lfengad in #576
• Feat/refactor pi05 by @littlespray in #563
• fix: more option for reduce on cpu by @lfengad in #579
• Fix DiffusionNFT beta config by @Dinghow in #581
• Fix the image/video range for reward computing by @Dinghow in #580
• Feat: support comet dataset for pi05 by @littlespray in #567
• Remove unused metrics for Diffusion RL by @Dinghow in #582
• Update controller to support datapacker factory mode by @YurongYou in #583
• feat: dataloader broadcast for non dp ranks by @lfengad in #585
• fix: resume issue regarding validation by @lfengad in #587
• fix: separate_model_parts in HF-Model by @jcao-ai in #588
• Fixes for multi-lr optimizer and model resuming by @jcao-ai in #590
• Remove dp_replicate from dp_cp_tp mesh dimension names by @jcao-ai in #591
• Added --slurm-job-time argument in dispatch_job.py by @vinjn in #589
• fix moe weight loading by @jcao-ai in #593
• fix moe weight loading by @jcao-ai in #594
• fix: accurate weight version control & cumem of nccl default off by @lfengad in #592
• Reuse max_num_steps for load_balanced_max_steps by @kane-vln in #595
• fix: check ckpt loaded info for extra info validation by @lfengad in #596
• add qwen3-vl merger supprot by @jcao-ai in #598
• fix: Validation hang when n_generation > 1 by @lfengad in #597
• fix: Ignore version control when policy scaling by @lfengad in #599
• Add WAN2.2 support for reward service by @yufanhuangNV in #602
• fix: untable issue for nccl group usage by @lfengad in #603
• Support SANA for DiffusionNFT by @Dinghow in #586
• Fix reward media_type & WFM local reward service issue by @Dinghow in #604
• Fuse TP/EP with DP for non-moe layer by @kane-vln in #600
• Support batched GenEval reward by @Dinghow in #605
• Add Qwen3VL merged visual forward and FSDP pure-text dummy pass by @Xuanmeng-Zhang in #606
• Support UnifiedReward by @Dinghow in #608
• Update slurm launch script by @YurongYou in #578
• Add LR scheduler for diffusion RL by @Dinghow in #610
• Correct DDRL preset configs & Add docs for parallelism by @Dinghow in #613
• Fix the media_type for video latency benchmark by @Dinghow in #612
• Fix bugs on batch sampler by @YurongYou in #609
• Add example for SANA post-training by @Dinghow in #611
• fix: RL batch sampler epoch set by @lfengad in #614
• fix: Automatically import models in policy.model by @lfengad in #615
• Sync updates from VLM branch by @jcao-ai in #616
• Fix the configuration page import error by @Dinghow in #618
• feat: on-policy distill slurm supoort and doc by @lfengad in #619
• fix: upgrade lint ruff to 0.12.7 in pre-commit for compatible with i4 by @lfengad in #620

New Contributors

• @Xuanmeng-Zhang made their first contribution in #606

Full Changelog: v0.4.0...v0.4.1