Issues

• [RFC] Feat: Enable Intel GPU Support

  docenhancement
  ByteDance-Seed/VeOmni#1458 · Filter by author cheehook opened on Nov 12, 2026

• [trainer, dist] fix: skip a non-finite gradient step and coordinate the final export across ranks

  bugcifixmisc
  ByteDance-Seed/VeOmni#1457 · Filter by author Coach257 opened on Nov 12, 2026

• [omni, dist, parallel] feat: support skipping inter-module SP all-gather on conversation_list (skip_sp_gather_between_modules)

  cienhancementroadmap
  ByteDance-Seed/VeOmni#1454 · Filter by author Coach257 opened on Nov 12, 2026

• [RFC] MiniMax-H3-Image support: inference baseline and packed batch training

  ByteDance-Seed/VeOmni#1445 · Filter by author NancyFyong opened on Nov 11, 2026
  1

• [omni, model] feat: add DeepSeek-V4.1-Flash audio module in SeedOmni

  ByteDance-Seed/VeOmni#1440 · Filter by author Coach257 opened on Nov 11, 2026

• [omni, model, ops, perf] feat: add DeepInfer low-precision Triton/SM100 and CUDAGraph inference for DeepSeek-V4.1-Flash

  ByteDance-Seed/VeOmni#1439 · Filter by author Coach257 opened on Nov 11, 2026

• [omni, model, config, docs] feat: support DeepSeek-V4.1-Flash in SeedOmni

  doc
  ByteDance-Seed/VeOmni#1438 · Filter by author Coach257 opened on Nov 11, 2026
  1

• [data, omni, task] feat: add CoTA and ESC-50 audio SFT dataset preparation and SeedOmni preprocessors

  ByteDance-Seed/VeOmni#1436 · Filter by author Coach257 opened on Nov 11, 2026

• [ckpt, dist, omni] feat: support extra_parallel_only, rowwise companion streaming load, and deferred HF shard merge

  ByteDance-Seed/VeOmni#1435 · Filter by author Coach257 opened on Nov 11, 2026
  1

• [dist, optim, trainer] perf: FSDP2 prefetch grouping, async offload sync elimination, Muon zero-comm and foreach grad norm

  ByteDance-Seed/VeOmni#1434 · Filter by author Coach257 opened on Nov 11, 2026

• [ops, model] feat: add DeepGEMM FP8 linear/swiglu_mlp and FP4 kv_round ops and wire into DeepSeek-V4

  ByteDance-Seed/VeOmni#1433 · Filter by author Coach257 opened on Nov 11, 2026

• [omni, model, ops, dist, ckpt] chore: DeepSeek-V4.1-Flash (deepseek_v41) upstreaming plan

  ckptmiscroadmap
  ByteDance-Seed/VeOmni#1432 · Filter by author Coach257 opened on Nov 11, 2026
  1

• [omni, model, config, docs] feat: support Qwen3-Omni in SeedOmni V2

  ByteDance-Seed/VeOmni#1431 · Filter by author Coach257 opened on Nov 11, 2026

• [ops, model] support a pass-through act_fn on the quantizing MoE / SwiGLU rows

  enhancement
  ByteDance-Seed/VeOmni#1430 · Filter by author Coach257 opened on Nov 10, 2026

• [model, dist] The aux-loss-free router bias e_score_correction_bias is never updated

  ByteDance-Seed/VeOmni#1420 · Filter by author arun-elr opened on Nov 7, 2026

• [RFC] Integrating HyperParallel into VeOmni

  ByteDance-Seed/VeOmni#1413 · Filter by author Cui-yshoho opened on Nov 5, 2026
  3

• [Ascend]: 2026 Q4 NPU Roadmap

  ascendroadmap
  ByteDance-Seed/VeOmni#1405 · Filter by author Feng0w0 opened on Nov 3, 2026

• [RFC] Supports the training of JEV decision models

  ByteDance-Seed/VeOmni#1399 · Filter by author Lihui-Gu opened on Nov 2, 2026

• [Tracking] Low-precision training: QAT, FP8 compute, and communication

  ByteDance-Seed/VeOmni#1366 · Filter by author FoolPlayer opened on Oct 26, 2026

• [Roadmap] VeOmni 2026 Q4 roadmap

  roadmap
  ByteDance-Seed/VeOmni#1365 · Filter by author FoolPlayer opened on Oct 26, 2026

• [ops, perf] DeepSeek-V4 sparse MLA backward: rolled split_store loop puts the dKV accumulator in local memory

  ByteDance-Seed/VeOmni#1361 · Filter by author 0z5a opened on Oct 25, 2026

• [RFC] Low-precision training with FSDP2 and torchao

  ByteDance-Seed/VeOmni#1360 · Filter by author amenomori opened on Oct 25, 2026
  1

• H3 BF16 SP gradient differences: is an opt-in cuDNN backend appropriate?

  ByteDance-Seed/VeOmni#1308 · Filter by author blacklist-twx opened on Oct 15, 2026

• [RFC] Target-centric, shard-on-read checkpoint loading

  ByteDance-Seed/VeOmni#1297 · Filter by author Coach257 opened on Oct 13, 2026
  1

• [ops, perf] Grouped GEMM: fc1 activation runs eager between two GEMM launches

  ByteDance-Seed/VeOmni#1289 · Filter by author wangzhao-11a opened on Oct 11, 2026