ReleaseByteDance (Doubao/Seed)ByteDance (Doubao/Seed)published Oct 17, 2025seen Jun 26

ByteDance-Seed/VeOmni v0.1.2

ByteDance-Seed/VeOmni

Open original ↗

Captured source

source ↗
GH
GitHub/github.com/ByteDance-Seed/VeOmni
ByteDance-Seed/VeOmni v0.1.2
Source ↗
published Oct 17, 2025seen Jun 26captured 4dhttp 200method plain

v0.1.2

Repository: ByteDance-Seed/VeOmni

Tag: v0.1.2

Published: 2025-10-17T18:08:52Z

Prerelease: yes

Release notes:

What's Changed

  • [misc] shift bytecheckpoint to optional dependency by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/92
  • [misc] revert ckpt default to avoid internal exceptions by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/93
  • [dist] minor fixes by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/94
  • [misc] feat: add GITBUG ISSUE TEMPLETE by @Fazziekey in https://github.com/ByteDance-Seed/VeOmni/pull/95
  • [data] feat: support megatron-energon dataset by @ziqi-wlb in https://github.com/ByteDance-Seed/VeOmni/pull/62
  • [data] add interleaved dataset by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/90
  • fix:remove a failing assertion by @KaijingOfficial in https://github.com/ByteDance-Seed/VeOmni/pull/97
  • [config] clean gitignore by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/99
  • [dist] fix: DCP auto load by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/106
  • [model] fix: Switch qwen3 and seed_oss to veomni defined GradientCheckpointingLayer by @piyifan123 in https://github.com/ByteDance-Seed/VeOmni/pull/109
  • [misc] feat: Add uv support to allow simple uv sync based python package management by @piyifan123 in https://github.com/ByteDance-Seed/VeOmni/pull/110
  • [BREAKING][dist] feat: Unified dcp saving for model and optimizer by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/107
  • [misc] feat: add skip_ulysses flag to bypass Ulysses logic in flash_attention_forward by @Juntian777 in https://github.com/ByteDance-Seed/VeOmni/pull/111
  • [dist] fix: remove unnecessary assert by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/112
  • [misc] feat: option to profile rank0 only or all the ranks by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/113
  • [misc] fix: remove buggy memory timeline export by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/114
  • [config] feat: add allow_cuda_launch_blocking by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/115
  • [ckpt] fix: remove unnecessary path joining for dcp by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/121
  • [ckpt][BREAKING] fix unnecessary wrapping for model and optimizer states by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/122
  • fix: qwen2 vl yaml by @Ziyi-Wang in https://github.com/ByteDance-Seed/VeOmni/pull/127
  • [data] fix :fix data collator for sp with cu_seq_lens_q and max_length_q by @Fazziekey in https://github.com/ByteDance-Seed/VeOmni/pull/126
  • [data] fix: dataset call hdfs api by @Ziyi-Wang in https://github.com/ByteDance-Seed/VeOmni/pull/128
  • [misc] fix: update asomeworks by @Fazziekey in https://github.com/ByteDance-Seed/VeOmni/pull/135
  • [model] fix: remove patch for npu by @heidongxianhua in https://github.com/ByteDance-Seed/VeOmni/pull/134
  • [dist] feat: faster weight loading through broadcasting from rank0 by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/123
  • [data] feat: support correct cu_seqlens handling for SP and non-SP by @Juntian777 in https://github.com/ByteDance-Seed/VeOmni/pull/136
  • [ckpt] fix: rank for get last iteraton for non-dcp path by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/140
  • [model] fix: deepseek-v3 by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/139
  • [model] fix: remove Qwen3-MoE redundant flashattention prep and fix input_ids access bug by @Juntian777 in https://github.com/ByteDance-Seed/VeOmni/pull/141
  • [data] fix: remove hf dependency on prepare_fa_kwargs_from_position_ids by @Juntian777 in https://github.com/ByteDance-Seed/VeOmni/pull/144
  • fix: wan_attnetion_missing_config_issue by @JeffryLee in https://github.com/ByteDance-Seed/VeOmni/pull/133
  • [fsdp] feat: support broadcast large weight by chunk. by @ZZWHU in https://github.com/ByteDance-Seed/VeOmni/pull/142
  • [core] fix: use flash_attention_2 backend by @KKZ20 in https://github.com/ByteDance-Seed/VeOmni/pull/124

New Contributors

  • @ziqi-wlb made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/62
  • @KaijingOfficial made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/97
  • @Juntian777 made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/111
  • @Ziyi-Wang made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/127
  • @heidongxianhua made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/134
  • @JeffryLee made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/133
  • @ZZWHU made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/142

Full Changelog: https://github.com/ByteDance-Seed/VeOmni/compare/v0.1.1...v0.1.2