Issues

• Sharded loader: every (rank, worker) draws the same within-shard shuffle stream, so a global batch can contain duplicate samples

  dreamzero0/dreamzero#91 · Filter by author dineshmane opened on Sep 10, 2026

• LoRA checkpoints are ~95 GB because DeepSpeed saves the frozen backbone (exclude_frozen_parameters is never passed: the model is not a PeftModel)

  dreamzero0/dreamzero#90 · Filter by author dineshmane opened on Sep 10, 2026

• Training forward crashes under torch.no_grad() (DiT block output tuple not unpacked)

  dreamzero0/dreamzero#85 · Filter by author jeqcho opened on Aug 7, 2026

• load_lora composes post-trained LoRA checkpoints on raw Wan instead of the DreamZero-AgiBot base they were trained on

  dreamzero0/dreamzero#82 · Filter by author dineshmane opened on Jul 24, 2026
  1

• Dataloader crashes on trailing incomplete batch (np.stack shape mismatch)

  dreamzero0/dreamzero#80 · Filter by author mekala-backup opened on Jul 10, 2026
  1

• About scripts/train/droid_training_wan22.sh

  dreamzero0/dreamzero#79 · Filter by author tyt951021 opened on Jul 7, 2026

• Question: mapping between DreamZero-DROID-Data episodes and raw DROID 1.0.1 trajectories / frame indices

  dreamzero0/dreamzero#75 · Filter by author Candy-Crusher opened on Jun 3, 2026

• The code implementation of Causal Attention does not match the paper description

  dreamzero0/dreamzero#74 · Filter by author bugmakerwww opened on Jun 1, 2026

• Wan2.2 TI2V training

  dreamzero0/dreamzero#73 · Filter by author HSID opened on May 28, 2026

• Question on 1-Step Inference Logic in DreamZero-Flash

  dreamzero0/dreamzero#72 · Filter by author 2460707925 opened on May 24, 2026

• Wan2.2 backbone inference question.

  dreamzero0/dreamzero#70 · Filter by author Tengbo-Yu opened on May 19, 2026

• Timeline for releasing the Agibot pre-training data

  dreamzero0/dreamzero#67 · Filter by author jiaming-zhou opened on Apr 28, 2026
  1

• Unexpected Video Generation in PolaRiS Simulation

  dreamzero0/dreamzero#66 · Filter by author XinyaChen21 opened on Apr 25, 2026

• cuda version

  dreamzero0/dreamzero#65 · Filter by author zhirui-gao opened on Apr 23, 2026
  1

• Any timeline for official closed-loop benchmark support beyond DROID sim?

  dreamzero0/dreamzero#64 · Filter by author yangjian253 opened on Apr 21, 2026

• State token attention mismatch between training and inference

  dreamzero0/dreamzero#61 · Filter by author boqian-li opened on Apr 16, 2026

• Question on expected behavior in the Droid simulation

  dreamzero0/dreamzero#59 · Filter by author FFYY361 opened on Apr 15, 2026

• Training steps / schedule for new-embodiment (YAM) fine-tuning in the paper vs. released scripts

  dreamzero0/dreamzero#58 · Filter by author ShuaiZhou302 opened on Apr 15, 2026
  1

• How long will it take for finetune?

  dreamzero0/dreamzero#56 · Filter by author YandanYang opened on Apr 14, 2026
  3

• The treatment of states seems inconsistent during training and autoregressive inference？

  dreamzero0/dreamzero#55 · Filter by author jby1993 opened on Apr 9, 2026

• Will the YAM play data used for new-embodiment adaptation be released ?

  dreamzero0/dreamzero#54 · Filter by author Laryderr opened on Apr 7, 2026

• Is the API service still active?

  dreamzero0/dreamzero#52 · Filter by author DingLihe opened on Apr 1, 2026

• Question regarding the specific benchmarks used on the MolmoSpaces leaderboard

  dreamzero0/dreamzero#50 · Filter by author Undefinedx-Zhang opened on Apr 1, 2026

• Actual RAM usage during model loading

  dreamzero0/dreamzero#48 · Filter by author JinnnK opened on Mar 30, 2026
  3

• About how to identify the instruction is done

  dreamzero0/dreamzero#44 · Filter by author 929362264 opened on Mar 24, 2026
  2