v0.1.132026-09-30
What's Changed
* [data] fix: preserve source timing in Qwen-VL video preprocessing by @yeyeyeping in https://github.com/ByteDance-Seed/VeOmni/pull/1165
* [agent] fix: correct stale facts in skills and knowledge docs by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/1135
* [agent, ci] feat: add a test-wiring decision rule by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/1136
* [data, model, config, ci] feat: fold MiniMax parquet loader into generic iterable by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/1156
* [agent] refactor: move the review gate to PR time and drop agent-specific tooling by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/1137
* [agent, docs] refactor: rename the migrate skill to veomni-patchgen-model by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/1145
* [agent, docs, model] fix: split patchgen guidance and guard expert layouts by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/1147
* [agent, docs] refactor: describe the agent tooling generically by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/1148
* [ckpt] fix: do not publish a staged checkpoint when another rank's copy failed by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/1172
* [ckpt] fix: sweep stale staged checkpoints instead of only the current step's by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/1173
* [dist] refactor: inject the async-offload host buffer pool by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/1175
* [BREAKING][config, trainer, dist] refactor: make AcceleratorConfig self-resolving and model-scoped by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/1090
* [BREAKING][dist, trainer] refactor: build ParallelState from AcceleratorConfig by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/1178
* [BREAKING][ckpt, trainer] refactor: extract ModelCheckpointManager and slim checkpoint callbacks by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/1179
* [BREAKING][model, ops, misc] chore: upgrade transformers to 5.16.1 by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/1171
* [BREAKING][ckpt, lora, docs] feat: split the checkpoint step directory by state owner by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/1181
* [model, parallel] feat: add Qwen3.8 Flash Next(Qwen4-Exp) SFT support by @Lihui-Gu in https://github.com/ByteDance-Seed/VeOmni/pull/1151
* [parallel, ckpt] fix: register ExtraParallel specs for aliased parameter names by @JorgenWan in https://github.com/ByteDance-Seed/VeOmni/pull/1190
* [ckpt] fix: keep the staged-checkpoint copy off the training backend by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/1176
* [ckpt] feat: let the HF export plan exclude caller-named keys by @JorgenWan in https://github.com/ByteDance-Seed/VeOmni/pull/1193
* [BREAKING][trainer, model] refactor: extract model-bound logic into VeOmniModelRuntime by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/1125
* [ckpt, trainer] fix: persist device and condition-model RNG across DiT resume by @Feng0w0 in https://github.com/ByteDance-Seed/VeOmni/pull/1189
* [model] feat: support mtp training for Qwen3.5 dense and moe model by @qcm1 in https://github.com/ByteDance-Seed/VeOmni/pull/1088
* [trainer] fix: honor DiT microbatch size by @NancyFyong in https://github.com/ByteDance-Seed/VeOmni/pull/1201
[ops, perf, model] refactor: tighten fused-MoE grouped-GEMM maxM launch bound to token count by @zhangxin81 in https://github.com/ByteDance-Seed/VeOmni/pull/1025
* [agent, docs] feat: add reusable documentation audit skill by @ruchu127 in https://github.com/ByteDance-Seed/VeOmni/pull/1157
* [dist, config, docs] feat: add node-local low-precision ReduceScatter communication by @ZHAO-YIFENG in https://github.com/ByteDance-Seed/VeOmni/pull/1167
* [omni, model, ops, ci] feat: add OmniConfig, OmniModel, graph FSM and processing primitives by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/1192
* [ops, model] feat: add Hugging Face Hub FlashAttention backends by @NancyFyong in https://github.com/ByteDance-Seed/VeOmni/pull/1214
* [model, trainer] feat: pack MiniMax H3 offline training by @NancyFyong in https://github.com/ByteDance-Seed/VeOmni/pull/1204
* [data, model] fix: use sampled video timing in Qwen Omni by @yeyeyeping in https://github.com/ByteDance-Seed/VeOmni/pull/1221
* [dist, ci, docs] fix: release reentrant checkpoint input gradients by @ZHAO-YIFENG in https://github.com/ByteDance-Seed/VeOmni/pull/1169
* [omni, model, trainer, data] feat: add seedomni runtime, trainer and media data by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/1235
* [dist] fix: place Ulysses inside CP on device mesh by @Zx55 in https://github.com/ByteDance-Seed/VeOmni/pull/1257
New Contributors
* @NancyFyong made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/1201
* @ruchu127 made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/1157
* @ZHAO-YIFENG made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/1167
Full Changelog: https://github.com/ByteDance-Seed/VeOmni/compare/v0.1.12...v0.1.13
v0.1.122026-09-09
Highlights
DeepSeek-V4, end to end** — training support (#840), TileLang kernels (#912), Liger kernel path (#946), MFU accounting (#944), Ulysses sequence parallel (#949) and context parallel for attention, compressors and indexer (#1111, #1131), the Lightning Indexer KL objective (#1134) with its TileLang teacher-distribution kernel (#1109), FlashMLA cuDNN DSA backend (#850), fp8/fp4 quantized weight export (#962), QAT fake-quant training (#1089), and NPU support (#974).
New models** — MiniMax H3 (#1075) with Ulysses SP (#1120), LTX-2.3 (#858, #859), GPT-OSS (#854) plus a 120B EP4 LoRA recipe (#921), and an expert-parallel plan for Qwen3.5-MoE causal LM (#1106).
LoRA** — native PEFT-free LoRA stack with MoE-LoRA and expert parallelism (#758), VLM trainer support (#1018), Qwen3.5-MoE (#1094), EP-sharded streaming load for PEFT models (#915), and faster eager MoE-LoRA routing (#1154, #1155).
Parallelism and performance** — per-module local parallel state via context manager (#893), FSDP2 torch.compile (#881, #1009), MindSpeed-style async activation offload (#965), shared async Ulysses backward helpers (#1113), FlexAttention with Ulysses (#945), MagiAttention on SM90 with Ulysses (#1021, now an optional extra in #1160), HSDP for expert parallel (#867), optional train-step sync (#938), and reused loss denominators (#941).
Muon** — Gram Newton-Schulz backends (#953), FSDP2 all-to-all optimization (#973), head-group split (#980) with qualified module names (#1108), and Qwen3.5-MoE Muon on Ascend (#1092).
Ascend / NPU — fused npurms_norm (#978), AscendC chunk gated delta rule (#924), Triton clamped SwiGLU (#1104), Qwen3-Omni-MoE with fused RoPE (#910), CANN 9.1.0 images (#1128), and GDN Dockerfiles (#1084).
More accelerators** — AMD ROCm (#890) and Cambricon MLU (#903).
Data and trainer** — effective-token dynamic batching (#833), map-style datasets in worker-side dynamic batching (#853), dataloader scheduling knobs (#940), model aux metrics reported without folding them into the loss (#1110), channel loss logging (#892), checkpoint early-stop (#926), and a pending async save awaited at train end (#1162).
Toolchain** — transformers 5.9.0 is the default install (#774), and CUDA 13.0 + PyTorch 2.11 images (#827).
Breaking Changes
* Per-module local parallel state moves to a context manager (#893)
* DeepSeek-V4 kernel naming unified (#934)
* SeedOmni V1 model stack removed (#1082)
* ChunkMBS removed (#1119)
* Head-split Muon is selected by qualified module names instead of a DSA indexer flag (#1108)
What's Changed
* [ckpt, trainer] fix: avoid HBM OOM during DCP save under MoE / VLM training by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/791
* [model] feat: register ForTokenClassification in model registries by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/795
* [parallel] fix stride error in npu when dsp+recompute by @yicheng-gong in https://github.com/ByteDance-Seed/VeOmni/pull/793
[ckpt, data] refactor: remove fillmissingoptimizer_states, lazy-import multimodal deps by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/800
* [ckpt] refactor: deduplicate ExtraParallel dim preprocessing by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/803
[ckpt] refactor: centralize async DCP save lifecycle in waitforpendingsave() by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/802
* [parallel] fix: keep root's auto-no-reshard so fused-linear kernels can backward by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/806
* [model, docs] refactor: make patchgen reusable as a library by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/807
* [ci, docker] chore: bump uv from 0.9.8 to 0.11.16 by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/782
* [model, ci, docs] feat: bump transformers to 5.9.0 by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/774
* [config, ci, docs] chore: relax uv pin from ==0.11.16 to a 0.9.8-0.11 range by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/808
* [model, ci, docs, agent] refactor: ship patchgen as a sibling standalone package by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/809
* [data, model] feat: Implemented dynamic batch logic optimization by @zhihaofang1017 in https://github.com/ByteDance-Seed/VeOmni/pull/789
[ckpt, model] fix: rename per-expert HF weightmap keys for fused MoE HF export by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/799
* [docs, docker] feat: Update A3 Docker image to 9.0.0 and NPU model support table by @cls1206 in https://github.com/ByteDance-Seed/VeOmni/pull/817
[model] fix: derive hybrid-attn layer counts from layertypes for transformers v5 by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/819
* [docker] feat: update docker with template by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/814
* [docs] chore: drop stale offline MoE merge step from Qwen3 / Qwen3-MoE / Qwen3-Omni recipes by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/813
* [data, omni, docs] feat: Qwen3-Omni recipe for offline-extracted audio-enabled video by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/812
* [ops, ci] feat: extend NPU kernel test coverage and add CPU-runnable registry/eager sanity tests by @Kirrito-k423 in https://github.com/ByteDance-Seed/VeOmni/pull/820
[data] feat: align StatefulDataLoader snapshot frequency to savesteps by @Fazziekey in https://github.com/ByteDance-Seed/VeOmni/pull/822
* [docs] fix: Ascend documents fix by @dzy00897185 in https://github.com/ByteDance-Seed/VeOmni/pull/830
* [model] feat: support sp on Qwen-Image by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/815
* [perf] fix: enable ProfilerWithMem with NPU by @yanghw116 in https://github.com/ByteDance-Seed/VeOmni/pull/831
* [docker] feat: support cu130+pt211 by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/827
* [model] fix: patched forward docstrings for transformers 5.9 by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/836
* [misc] feat: pin FA4 4.0.0b16 by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/837
* [model] fix: document qwen3 moe cache position by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/838
* [ci] fix: npu aarch64 dependency resolution and uv installation guide by @Crystal-jiang in https://github.com/ByteDance-Seed/VeOmni/pull/832
* [ci] fix: run wan e2e in bf16 by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/835
* [data, trainer] feat: support effective-token dynamic batching by @zhangxin81 in https://github.com/ByteDance-Seed/VeOmni/pull/833
* [trainer] fix: skip NCCL destroy on normal trainer exit by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/849
* [ops, model] feat: add FlashMLA cuDNN DSA backend by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/850
* [data, model] fix: coalesce qwen3.5 pad cu-seqlens for GDN by @YLlllllllllll in https://github.com/ByteDance-Seed/VeOmni/pull/845
* [model] feat: add GPT-OSS VeOmni backend by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/854
* [model] feat: add LTX-2.3 source code by @JimmyMrr in https://github.com/ByteDance-Seed/VeOmni/pull/858
* [model, ci] feat: support deepseek v4 by @zhangxin81 in https://github.com/ByteDance-Seed/VeOmni/pull/840
* [model, ci] fix: GPT-OSS e2e parametrization by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/861
* [config] fix: auto-resolve ops defaults to NPU-compatible by @Crystal-jiang in https://github.com/ByteDance-Seed/VeOmni/pull/863
* [ci, lora] test: add LoRA save/load case to NPU unit tests by @xssty123 in https://github.com/ByteDance-Seed/VeOmni/pull/851
[data] feat: add data-pipeline dry-run script (tasks/datasim.py) by @LiuzcEECS in https://github.com/ByteDance-Seed/VeOmni/pull/865
* [dist] fix: correct FSDP2 grad n
v0.1.112026-05-26
Highlights
* topk fused kernel for OPD
* precompute metadata in dataloader for VLM
* Qwen-Image support with ckpt/lora/inference
* rework moe load balance monitor
* NPU OpSlot fixes
What's Changed
* [model] feat: support qwen-image by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/770
* [model, data, trainer] feat: precompute multimodal forward metadata in dataloader by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/772
* [model, ops] feat: chunked fused-linear top-k forward-KL distillation kernel by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/771
* [omni] chore: add Omni Molde inference script by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/777
[model, data, ci] refactor: collapse multimodal metadata to a single gridthw key by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/778
* [parallel] feat: disable HSDP gradient all-reduce during gradient accumulation by @nono-Sang in https://github.com/ByteDance-Seed/VeOmni/pull/781
[ops, model] refactor: shorten loss-wrapper return to (loss, logits, fusedlinear_aux) by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/780
* [model, ops] refactor: add NPU support and OpSlot guard for Qwen3/VL/MoE, Qwen3.5/MoE by @yanghw116 in https://github.com/ByteDance-Seed/VeOmni/pull/710
* [model, omni] feat: add Qwen-Image lora config by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/784
* [model, ci, agent] feat: wire qwen2-family ViT to the multimodal metadata precompute hook by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/779
* [trainer, ops] feat: rework MoE load-balance monitor (model-agnostic, EP/DP-aware) by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/787
* [ckpt, lora] feat: Save lora ckpt and add omni-infer with lora by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/785
* [model, omni] feat: Update Qwen-Image & Add veomni fsdp state API by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/786
* [model, ops] fix: complete #780 3-tuple loss-wrapper migration by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/790
* [release] chore: bump version to 0.1.11 by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/792
New Contributors
* @yanghw116 made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/710
Full Changelog: https://github.com/ByteDance-Seed/VeOmni/compare/v0.1.10...v0.1.11
v0.1.102026-05-21
What's Changed
* [model] fix: dit + lora related fix by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/735
* [docker] feat: add dockerfile and image for 9.0.0 cann base image by @phdddd in https://github.com/ByteDance-Seed/VeOmni/pull/742
* [parallel, perf] fix: sp gather && optimize: input embeds fusing by @JorgenWan in https://github.com/ByteDance-Seed/VeOmni/pull/681
* [lora] feat: add lora for veomni by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/739
* [optim] feat: support muon by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/744
[model, misc] refactor: drop qwen35(_moe) vision encoder host-syncs by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/752
[model] feat: update seedoss v5 by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/754
* [config, ci, docker] feat: make transformers v5.2.0 the default install by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/751
[ci, model] test: add v5 loader-path test and glmmoe_dsa coverage by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/727
* [dist, config, docs] feat: drop FSDP1 by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/756
* [model] feat: add MoE router replay hook for RL training frameworks by @hjshi84 in https://github.com/ByteDance-Seed/VeOmni/pull/719
[model, ci] feat: deepseekv3update v5 & clean v4 ci by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/755
[dist] fix: update dtensorfactory to use partial for local split in paralle… by @kahlun in https://github.com/ByteDance-Seed/VeOmni/pull/757
* [docs] fix: fix docs for ascend by @phdddd in https://github.com/ByteDance-Seed/VeOmni/pull/738
* [parallel] feat: add cpu offload for fsdp2 by @nono-Sang in https://github.com/ByteDance-Seed/VeOmni/pull/753
[model] fix: read tieword_embeddings from inner text/decoder config by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/693
* [model] feat: update qwen2.5omni & llama to v5 by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/767
[model, ci] test: add implicit CUDA sync gate for qwen35 generated modeling by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/760
* [docs] fix: fix doc build by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/769
* [BREAKING][ci, model] feat: cleanup v4 by @Coach257 in https://github.com/ByteDance-Seed/VeOmni/pull/768
* [model, perf] fix: remove production-path CPU syncs from Qwen3.5 / Qwen3.5-MoE patches by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/762
* [model, perf] fix: remove production-path CPU syncs from Qwen3-VL / VL-MoE / Omni-MoE patches by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/764
* [release] chore: release v0.1.10 by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/759
Full Changelog: https://github.com/ByteDance-Seed/VeOmni/compare/v0.1.9a5...v0.1.10
v0.1.9a5预发布2026-05-07
What's Changed
* [ci, model] test: add bitwise logits-equal tests for v5 models by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/722
* [ci] refactor: split GPU unit tests into v4/v5 parallel jobs by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/725
* [BREAKING][ops, model] feat: GPU-optimal ops defaults + strict NPU validation by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/716
* [model] fix: fix flops count by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/730
[model] fix: preserve FSDP2 pre-backward hooks for log-prob outputs in qwen35_moe and other models by @deerlu in https://github.com/ByteDance-Seed/VeOmni/pull/731
* [ci] refactor: split GPU e2e tests into v4/v5 parallel jobs by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/733
* [docker] fix: datasets version fix by @phdddd in https://github.com/ByteDance-Seed/VeOmni/pull/691
* [release] chore: release v0.1.9a5 by @deerlu in https://github.com/ByteDance-Seed/VeOmni/pull/736
Full Changelog: https://github.com/ByteDance-Seed/VeOmni/compare/v0.1.9a4...v0.1.9a5
v0.1.9a4预发布2026-05-05
What's Changed
[model] fix: qwen3moe router double-softmax and idempotent initweights wrap by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/715
* [ops, model] feat: dispatch Qwen3.5 linear-attention kernels via OpSlot by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/714
* [ops, model] feat: forward temperature and return entropy from log-probs path by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/720
* [ci, model] test: add bitwise logits-equal tests for transformers v4 models by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/721
* [release] chore: release v0.1.9a4 by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/723
Full Changelog: https://github.com/ByteDance-Seed/VeOmni/compare/v0.1.9a3...v0.1.9a4
v0.1.9a3预发布2026-05-01
What's Changed
* [model, ckpt] feat: runtime checkpoint tensor converter for transformers v4 MoE by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/696
[model] fix: qwen3vl vision dummy_forward sp shape mismatch by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/697
[misc] feat: add epsize property to ParallelState by @deerlu in https://github.com/ByteDance-Seed/VeOmni/pull/706
* [model, ops, trainer] refactor: unify MoE dispatch through OpSlot by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/705
[ops] fix: use reshape for non-contiguous chunks in chunkloss by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/708
* [parallel, ci] feat: support loading large tensor for fsdp2 by @JorgenWan in https://github.com/ByteDance-Seed/VeOmni/pull/688
* [ops, trainer] feat: chunked per-token log-probs for PPO-style RL by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/711
* [release] chore: release v0.1.9a3 by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/709
Full Changelog: https://github.com/ByteDance-Seed/VeOmni/compare/v0.1.9a2...v0.1.9a3
v0.1.9a2预发布2026-04-27
What's Changed
[parallel] feat: add enableforward_prefetch for fsdp2 by @JorgenWan in https://github.com/ByteDance-Seed/VeOmni/pull/660
[dist] fix: use srcdatarank=None in loadmodel_weights to skip redundant scatter by @kahlun in https://github.com/ByteDance-Seed/VeOmni/pull/648
[model] feat: add config.addhelper for module-level helpers in patchgen by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/663
* [model] feat: [transformers v5] support qwen3vl for transformer v5 by @yiwzhao in https://github.com/ByteDance-Seed/VeOmni/pull/527
* [ci] fix: add -v for npu ut e2e test by @phdddd in https://github.com/ByteDance-Seed/VeOmni/pull/667
* [ci] chore: trigger gpu/npu ci only when lint check pass by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/650
[model, ci, agent] feat: migrate qwen3vl_moe to transformers v5 by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/666
* [model, ci] test: bitwise-equal HF vs veomni logits; fix bf16 MoE dtype divergences by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/670
[model, ci, agent] feat: migrate qwen3omni_moe to transformers v5 by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/668
[task] fix: fix bug for args.train.accelerator.fsdpconfig.mixed_precision.enable by @UserChen666 in https://github.com/ByteDance-Seed/VeOmni/pull/638
[data] chore: add modal mask into keysto_pad by @JorgenWan in https://github.com/ByteDance-Seed/VeOmni/pull/672
* [docker] feat: add Ascend A3 Dockerfile and docs by @phdddd in https://github.com/ByteDance-Seed/VeOmni/pull/659
* [ops] refactor: replace kernel env vars with args by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/651
[ci] fix: disable pinmemory in test_datasets multisource tests for NPU by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/677
* [model] feat: support npu qwen3p5 and qwen3p5-moe by @yicheng-gong in https://github.com/ByteDance-Seed/VeOmni/pull/641
* [model] feat: guard unmigrated models against transformers v5 by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/675
* [misc] feat: support Python 3.12 by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/683
* [docker] feat: Fix npu workflow by @phdddd in https://github.com/ByteDance-Seed/VeOmni/pull/680
* [misc] chore: update fa2 cp312 wheel URL and revert .python-version by @Luosuu in https://github.com/ByteDance-Seed/VeOmni/pull/684
* [docker] fix: Fix workflow npu docker by @phdddd in https://github.com/ByteDance-Seed/VeOmni/pull/686
* [BREAKING][ops, model] feat: registration-based kernel replacement framework by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/678
[ops, model] refactor: relax moeimplementation Literal to str for extensibility by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/689
* [ci] feat: update ci workflow by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/679
* [docs] fix: add install datasets==2.21.0 to ascend doc by @phdddd in https://github.com/ByteDance-Seed/VeOmni/pull/692
* [release] chore: release v0.1.9a2 by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/694
New Contributors
* @kahlun made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/648
* @UserChen666 made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/638
Full Changelog: https://github.com/ByteDance-Seed/VeOmni/compare/v0.1.9a1...v0.1.9a2
v0.1.9a1预发布2026-04-15
What's Changed
[data, config] feat: add dynamic per-frame videomax_pixels for Qwen3-VL by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/635
* [agent] feat: add profile skill by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/626
* [misc] fix: revert fsdp basic modules, fix it with VL models only by @heidongxianhua in https://github.com/ByteDance-Seed/VeOmni/pull/634
* [docs] fix: list allowed modules and types in PR template by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/645
* [misc] feat: add pypi pub workflow by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/644
[ci] fix: disable pinmemory in multisource dataset test for NPU by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/639
* [ci] fix: fix ci port select error by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/646
* [agent] feat: add skill for migrating models to transformers v5 by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/652
* [misc] chore: update pub with release by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/649
[model, ci] feat: migrate qwen25_vl to transformers v5 by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/653
* [model] refactor: preserve leading comments in patchgen codegen by @TimYangst in https://github.com/ByteDance-Seed/VeOmni/pull/655
* [release] chore: release v0.1.9a1 by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/656
* [misc] fix: fix version check by @FoolPlayer in https://github.com/ByteDance-Seed/VeOmni/pull/658
Full Changelog: https://github.com/ByteDance-Seed/VeOmni/compare/v0.1.8...v0.1.9a1
v0.1.8预发布2026-04-09
This release significantly expands VeOmni’s model coverage, distributed training capabilities, and training workflow support since v0.1.7. Major updates include broader Qwen family support with Qwen3.5 and Transformers v5 integrations, the introduction of Extra Parallel and multiple EP/FSDP/SP improvements, new DPO and DiT/WAN training workflows, stronger MoE and checkpoint compatibility, and much deeper test, CI, and documentation coverage across both GPU and NPU environments.
Highlights
Model Coverage & Training Workflows
- Added Qwen3.5 MoE language support and extended Qwen3.5 ViT coverage, significantly broadening the Qwen3.5 training path (#547, #552, #602).
- Upgraded Qwen2 and Qwen2-VL integrations to the Transformers v5 stack (#526, #543).
- Added DPOTrainer together with Qwen3 DPO configs and examples (#558, #583).
- Added support for DiT / WAN training workflows, including new task entrypoints, configs, and examples (#570).
- Fixed freezevit behavior for Qwen3.5** models (#616).
- Improved compatibility for DeepSeek V3 rollout and FA4-related paths (#609, #582).
- Added FLOPs counting support for qwen35 and qwen35_moe (#561).
- Improved hub-kernel loader compatibility with Transformers v5.3.0+ (#633).
Distributed Training & Parallelism
- Introduced Extra Parallel, a breaking upgrade to VeOmni’s distributed parallel abstraction (#429).
- Added mixedprecision support in fsdpconfig and CPU parameter loading for FSDP1 (#627, #612).
- Extended EP support to merged FC1 and quack GEMM backends, and improved EP/FSDP behavior for expert modules (#588, #577).
- Added a fused Triton kernel for MoE load-balancing loss, improving distributed MoE training efficiency (#560).
- Improved Sequence Parallel stability with a zero-division guard in ReduceLoss and new rollwithsequence_parallel utilities (#618, #608).
Data Pipeline
- Added worker-side multi-source dynamic batching resume, a breaking change for resume behavior in multi-source data loading (#603).
- Added support for non-divisible frame alignment with framefactorremainder and improved video preprocessing behavior (#587, #585).
- Improved data transform robustness for source_name handling and multimodal edge cases (#553, #554, #550).
Checkpointing & Compatibility
- Added runtime checkpoint tensor conversion for Qwen3-MoE Transformers v5 fused expert weights (#589).
- Added a DCP consolidation patch for HDFS FUSE compatibility and fixed duplicate checkpoint saves (#536, #595).
- Preserved tokenizer config in the MoE merge script and fixed checkpoint writer filtering logic (#622, #593).
Testing, CI & Documentation
- Added dummy forward and FSDP equivalence tests to improve distributed correctness coverage (#620).
- Expanded and aligned NPU CI coverage with GPU, including additional unit and end-to-end cases (#566, #567, #597, #623).
- Added stronger testing documentation and consolidated test helpers (#631).
- Updated Ascend TorchCodec installation docs and added a helper installation script (#613).
- Added a CI verifier to ensure patchgen outputs stay in sync (#559).
Breaking Changes
- Extra Parallel is now the main abstraction for expert-style parallelism, with follow-up API/doc updates from epplan to extraparallel_plan (#429, #579).
- Worker-side multi-source dynamic batching resume changes resume semantics and should be validated carefully before upgrading existing jobs (#603).
New Contributors
* @hjshi84 made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/536
* @JorgenWan made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/429
* @deerlu made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/552
* @nono-Sang made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/611
* @xzzWZY made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/595
* @whisylan made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/608
* @cls1206 made their first contribution in https://github.com/ByteDance-Seed/VeOmni/pull/623
Full Changelog: https://github.com/ByteDance-Seed/VeOmni/compare/v0.1.7...v0.1.8