| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| README.md | 2026-03-31 | 4.4 kB | |
| v0.4.2 source code.tar.gz | 2026-03-31 | 5.5 MB | |
| v0.4.2 source code.zip | 2026-03-31 | 6.1 MB | |
| Totals: 3 Items | 11.7 MB | 1 | |
What's Changed
- Fix SkippingSampler bug by @YurongYou in https://github.com/nvidia-cosmos/cosmos-rl/pull/621
- feat: tests combined to script by @lfengad in https://github.com/nvidia-cosmos/cosmos-rl/pull/622
- [vla] support cosmos-policy by @fwd4 in https://github.com/nvidia-cosmos/cosmos-rl/pull/617
- fix: lint for docs also (compatible with i4 integration) by @lfengad in https://github.com/nvidia-cosmos/cosmos-rl/pull/623
- Support Qwen3-5 SFT by @kane-vln in https://github.com/nvidia-cosmos/cosmos-rl/pull/607
- Rename check_transformers_version to is_transformers_version_compatible by @kane-vln in https://github.com/nvidia-cosmos/cosmos-rl/pull/624
- Saving ckpt when receiving signals by @foreverlms in https://github.com/nvidia-cosmos/cosmos-rl/pull/601
- Update Wan2pt2 server config & client example by @Dinghow in https://github.com/nvidia-cosmos/cosmos-rl/pull/629
- [vla] support robotwin env setup and test in ci by @fwd4 in https://github.com/nvidia-cosmos/cosmos-rl/pull/627
- Support multi-reward training for diffusion RL by @Dinghow in https://github.com/nvidia-cosmos/cosmos-rl/pull/630
- Support export_safetensors for diffusion models by @Dinghow in https://github.com/nvidia-cosmos/cosmos-rl/pull/626
- feat: gb200 container setup by @lfengad in https://github.com/nvidia-cosmos/cosmos-rl/pull/633
- feat: Slurm more option for mount by @lfengad in https://github.com/nvidia-cosmos/cosmos-rl/pull/632
- [vla] fix pi05 compatibility issues on libero by @littlespray in https://github.com/nvidia-cosmos/cosmos-rl/pull/625
- Fix: pack visual_pos_masks for qwen3_vl_moe when seq_pack enabled by @kane-vln in https://github.com/nvidia-cosmos/cosmos-rl/pull/635
- Enable mixed precision training for diffusion RL by @Dinghow in https://github.com/nvidia-cosmos/cosmos-rl/pull/634
- fix: docs and check for profiler by @lfengad in https://github.com/nvidia-cosmos/cosmos-rl/pull/637
- fix: RL part resume epoch setting with tests added by @lfengad in https://github.com/nvidia-cosmos/cosmos-rl/pull/636
- Support Sequence Packing for Qwen3.5 by @kane-vln in https://github.com/nvidia-cosmos/cosmos-rl/pull/639
- fix: RL version compatible basically runnable for vllm 0.17 by @lfengad in https://github.com/nvidia-cosmos/cosmos-rl/pull/641
- Add tutorial for diffusion SFT & RL by @Dinghow in https://github.com/nvidia-cosmos/cosmos-rl/pull/644
- Add e2e test for diffusion RL by @Dinghow in https://github.com/nvidia-cosmos/cosmos-rl/pull/643
- Sync changes from nemotron branch by @jcao-ai in https://github.com/nvidia-cosmos/cosmos-rl/pull/645
- fix: slurm stability ehancement by @lfengad in https://github.com/nvidia-cosmos/cosmos-rl/pull/646
- fix: support _StridedShard DTensor placements for weight sync by @kane-vln in https://github.com/nvidia-cosmos/cosmos-rl/pull/650
- Add support for FA3 from internal flash_attn_3_nv. Also fix the flash_attn_varlen_func for FA3 by @yufanhuangNV in https://github.com/nvidia-cosmos/cosmos-rl/pull/648
- Set attention implementation to flash_attention_2 by default for HFModel by @kane-vln in https://github.com/nvidia-cosmos/cosmos-rl/pull/651
- fix: compatible with flash-attn-3 tuple output by @lfengad in https://github.com/nvidia-cosmos/cosmos-rl/pull/653
- Support validation for remote reward by @Dinghow in https://github.com/nvidia-cosmos/cosmos-rl/pull/655
- Fix SFT checkpointing barrier for multi-replica by @Dinghow in https://github.com/nvidia-cosmos/cosmos-rl/pull/657
- [vla] support maniskill env by @fwd4 in https://github.com/nvidia-cosmos/cosmos-rl/pull/656
- Fix: relax import assert by @lfengad in https://github.com/nvidia-cosmos/cosmos-rl/pull/658
- feat: Hook for ckpt handling by @lfengad in https://github.com/nvidia-cosmos/cosmos-rl/pull/659
- Sync changes from dev/nemotron by @jcao-ai in https://github.com/nvidia-cosmos/cosmos-rl/pull/660
- Support batched remote reward computation by @Dinghow in https://github.com/nvidia-cosmos/cosmos-rl/pull/661
- Add compute_default_rope_parameters for default RoPE when using Transformers ≥ 5.0 by @kane-vln in https://github.com/nvidia-cosmos/cosmos-rl/pull/654
- Support GRPO for Qwen3.5 by @kane-vln in https://github.com/nvidia-cosmos/cosmos-rl/pull/647
- Sync changes from VLM by @jcao-ai in https://github.com/nvidia-cosmos/cosmos-rl/pull/662
Full Changelog: https://github.com/nvidia-cosmos/cosmos-rl/compare/v0.4.1...v0.4.2