Question about GPU setup for dLDM stage-2 training

Hi, thanks for releasing VideoWorld 2!

I am trying to reproduce the dLDM training pipeline, and kept the stage-2 setting same as your released config:

- `predict2_video2world_ctrl_ldm_fsdp_2b_480p_16fps`
- `train_architecture="base"`
- `finetune_base_model=True`
- `use_cross_embedding=True`
- `batch_size=1` per GPU
- `num_frames=93`
- `video_size=(480, 832)`
- `fsdp_shard_size=8`
- 8 GPUs total

From the logs, stage 2 is training roughly:
- DiT trainable parameters: `2.20B`
- dLDM trainable parameters: `0.14B`

So this is effectively joint training of:
- full DiT
- full control branch
- dLDM

I can confirm that FSDP is enabled (`Distributed parallelism mode: fsdp`, `Using FSDP with shard size 8`), but I still run into OOM issues on 8 GPUs with 48 GB VRAM each.

Because of this, I would like to ask:

1. What exact GPU setup did you use for dLDM stage-2 training?
- number of GPUs
- VRAM per GPU
- GPU model

2. For the released training recipe, was stage 2 expected to fit on 8 x 48 GB GPUs?

3. Did you use any additional memory-saving settings beyond the released config?
For example:
- activation checkpointing changes
- different batch size / grad accumulation
- context parallelism
- different FSDP settings
- partial freezing instead of full base-model finetuning

4. If stage 2 in your internal training used a different setup than the public config, could you share the practical hardware recommendation for reproducing it?

Any clarification would be very helpful for reproduction. Thanks.