v0.4.0

What's Changed

• Support DiffusionNFT by @Dinghow in #465
• Fix OOM when prompt is very long. by @foreverlms in #498
• Fix qwen3_vl not supported in rollout. by @foreverlms in #501
• Always init distribution. by @foreverlms in #500
• few fix for nemotron-nano-v3 MoE support by @jcao-ai in #502
• Fix training hang when resuming from ckpt. by @foreverlms in #503
• fix: Resume reproduceable of shuffle case by @lfengad in #504
• support MoE aux-free loadbalancing by @jcao-ai in #506
• Remove a lot of log of loading weight by @foreverlms in #507
• feat: Mesh aware dispatch for avla requirement by @lfengad in #497
• Fix: control of eps by @littlespray in #511
• fix: Add config check and force. by @lfengad in #512
• Fix: step control for sft-trainer by @littlespray in #510
• fix: simple fix for rollout mesh info set by @lfengad in #514
• [vla] support pi05 on b1k by @fwd4 in #515
• Support diffusers-based Cosmos-Predict2.5 by @Dinghow in #499
• fix: resume sft train step missing fix and max_keep -1 support by @lfengad in #517
• Update doc for diffusers-based Cosmos-Predict2.5 by @Dinghow in #518
• Update docs for DDRL by @Dinghow in #521
• fix: pin transformers < 5.0.0 by @lfengad in #519
• Enable no upper limit retry for fetching prompts by default. by @foreverlms in #516
• [vla] add doc by @fwd4 in #520
• feat: support sft multiple replica control by @lfengad in #509
• Make vLLM as an optional dependency by @foreverlms in #525
• Optimize reward service by @Dinghow in #527
• fix: ci flaky 20260130 by @xlu451 in #529
• Add data packer factory mode by @YurongYou in #528
• Update batch_sampler factory API by @YurongYou in #524
• Colocated but separated processes for Policy and Rollout by @foreverlms in #508
• fix: Enhance prompt fetch efficiency by @lfengad in #532
• Fix pickscore by @Dinghow in #536
• feat: support batch group reward calc by @lfengad in #538
• Update the remote reward for DDRL by @Dinghow in #539
• Raise FA not installed error in runtime, not initialized time. by @foreverlms in #540
• Update DDRL configs by @Dinghow in #541
• feat: qwen3-vl user define version with i4 vlm dataloader trainable by @lfengad in #533
• fix: sft-trainer weight upload at final step by @littlespray in #542
• Dynamic enqueue timeout for remote reward by @Dinghow in #545
• Support non-euqal worldsize for policy and rollout in colocated-separated mode by @foreverlms in #544
• Check max_keep across all previous runs; log best ckpt across runs by @YurongYou in #546
• Revert changes in config from_dict function by @YurongYou in #549
• feat: support pi05 sft by @littlespray in #488
• fix: Fix laucnh multi node by @lfengad in #553
• Support moe aux_loss by @kane-vln in #543
• Robust ckpt saving and loading by @YurongYou in #550
• Fix: checkpoint manager init error in datafetcher by @YurongYou in #557
• Support async ckpt delete in checkpoint manager by @YurongYou in #555
• Fix: SFT failed to start if no ckpt to resume by @YurongYou in #556

New Contributors

• @littlespray made their first contribution in #511
• @YurongYou made their first contribution in #528

Full Changelog: v0.3.9...v0.4.0

What's Changed

• Support DiffusionNFT by @Dinghow in #465
• Fix OOM when prompt is very long. by @foreverlms in #498
• Fix qwen3_vl not supported in rollout. by @foreverlms in #501
• Always init distribution. by @foreverlms in #500
• few fix for nemotron-nano-v3 MoE support by @jcao-ai in #502
• Fix training hang when resuming from ckpt. by @foreverlms in #503
• fix: Resume reproduceable of shuffle case by @lfengad in #504
• support MoE aux-free loadbalancing by @jcao-ai in #506
• Remove a lot of log of loading weight by @foreverlms in #507
• feat: Mesh aware dispatch for avla requirement by @lfengad in #497
• Fix: control of eps by @littlespray in #511
• fix: Add config check and force. by @lfengad in #512
• Fix: step control for sft-trainer by @littlespray in #510
• fix: simple fix for rollout mesh info set by @lfengad in #514
• [vla] support pi05 on b1k by @fwd4 in #515
• Support diffusers-based Cosmos-Predict2.5 by @Dinghow in #499
• fix: resume sft train step missing fix and max_keep -1 support by @lfengad in #517
• Update doc for diffusers-based Cosmos-Predict2.5 by @Dinghow in #518
• Update docs for DDRL by @Dinghow in #521
• fix: pin transformers < 5.0.0 by @lfengad in #519
• Enable no upper limit retry for fetching prompts by default. by @foreverlms in #516
• [vla] add doc by @fwd4 in #520
• feat: support sft multiple replica control by @lfengad in #509
• Make vLLM as an optional dependency by @foreverlms in #525
• Optimize reward service by @Dinghow in #527
• fix: ci flaky 20260130 by @xlu451 in #529
• Add data packer factory mode by @YurongYou in #528
• Update batch_sampler factory API by @YurongYou in #524
• Colocated but separated processes for Policy and Rollout by @foreverlms in #508
• fix: Enhance prompt fetch efficiency by @lfengad in #532
• Fix pickscore by @Dinghow in #536
• feat: support batch group reward calc by @lfengad in #538
• Update the remote reward for DDRL by @Dinghow in #539
• Raise FA not installed error in runtime, not initialized time. by @foreverlms in #540
• Update DDRL configs by @Dinghow in #541
• feat: qwen3-vl user define version with i4 vlm dataloader trainable by @lfengad in #533
• fix: sft-trainer weight upload at final step by @littlespray in #542
• Dynamic enqueue timeout for remote reward by @Dinghow in #545
• Support non-euqal worldsize for policy and rollout in colocated-separated mode by @foreverlms in #544
• Check max_keep across all previous runs; log best ckpt across runs by @YurongYou in #546
• Revert changes in config from_dict function by @YurongYou in #549
• feat: support pi05 sft by @littlespray in #488
• fix: Fix laucnh multi node by @lfengad in #553
• Support moe aux_loss by @kane-vln in #543
• Robust ckpt saving and loading by @YurongYou in #550
• Fix: checkpoint manager init error in datafetcher by @YurongYou in #557
• Support async ckpt delete in checkpoint manager by @YurongYou in #555
• Fix: SFT failed to start if no ckpt to resume by @YurongYou in #556
• Update pyproject.toml by @lfengad in #558

New Contributors

• @littlespray made their first contribution in #511
• @YurongYou made their first contribution in #528

Full Changelog: v0.3.9...v0.4.0