[Training] About the stage 1 training details

[Teiboku]

As you said in paper, "Stage 1 uses 89-frame clips (768 × 1280) and paired images; Stage 2 continues training with videos only; Stage 3 adopts the same setting."
I want to know if you use sliding_window to reduce the GPU memory usage in these stage?

[Teiboku]

Because if input is 1280*768*89, the GPU memory will OOM obviously，and I have read your paper about this doesn't find useful information to explain this point.

[zhuang2002]

Hi @Teiboku, thanks for your attention to FlashVSR.
We addressed the OOM issue mainly with memory optimization techniques like DeepSpeed ZeRO, gradient checkpointing, CPU offloading.

[Teiboku]

Thanks for your kindly reply.