Inquiry on Training Configuration and Experimental Settings in FramePack

Nobody has claimed this yet.

Dear FramePack Team,

Thanks for your great job. I am deeply impressed by the technical advancements demonstrated in your work. Based on the paper description:
> *"FramePack achieves a batch size of 64 on a single 8×A100-80G node with the 13B HunyuanVideo model at 480p resolution (without using any image-based workaround training that many community LoRAs are built on top of)."*

I have the following questions regarding the implementation and experimental settings:

**Question 1: Usage of DeepSpeed ZeRO Optimization**
The paper mentions training the 13B HunyuanVideo model with a batch size of 64 on a single 8×A100-80G node at 480p resolution. To achieve this, did you employ DeepSpeed ZeRO optimization? If so, could you clarify which stage of ZeRO (e.g., Stage 1, 2, or 3) was used? Additionally, was offloading (e.g., CPU offload) utilized alongside ZeRO? This would help the community better understand the memory optimization strategies for training large video models.

**Question 2: Latent Frame Length Compatibility with Batch Size 64 on 8×A100-80G**
The released configuration specifies a latent frame length of 9, but the paper does not explicitly state whether this length was used in the experiments described (batch size 64 on 8×A100-80G at 480p). Could you clarify:
- **Is the latent frame length of 9 directly compatible with the reported training setup (64 batch size on 8×A100-80G)?**
- If not, what adjustments (e.g., reduced batch size, gradient accumulation, or architectural changes) would be required to use 9-frame sequences with this hardware configuration?
- Does the 9-frame configuration assume additional optimizations (e.g., memory-efficient attention, activation checkpointing) beyond what is described in the paper?

The issue names no files, tests, or entry points. Start by checking the released training configuration and the paper details for DeepSpeed ZeRO, offloading, latent frame length, and memory optimizations. Done means documenting whether the reported 64-sample setup used these settings and what adjustments are needed for nine-frame sequences.

Written by the indexing model from the issue text.