Hardware requiremements for 5B model

[LeeKeyu]

Hi authors, thanks a lot for your contributions.
My question is about the GPU usage of the newly updated YUME 1.5 5B version. I tried to run the inference script with 2*40g GPUs and got OOM. Could you please specify about the hardware requirements for testing and training the 5B model?
Thanks very much!

[stdstu12]

We can successfully infer a 1-minute video on an A100 80G. I suspect this is a precision issue; the current code seems to perform inference in Float32 (FP32). We will optimize this in the future.

[LeeKeyu]

@stdstu12 Thanks for your reply! May I ask how many A100 gpus were used for foundation model training and distillation, respectively?
Also I may have found a bug that when using >= 2 gpus for inference, some images in the jpg/ folder will be skipped and some will be evaluated more than once. Maybe the sampling code needs some modification.

[stdstu12]

We used 40 A100 GPUs. Thank you for pointing out the issue; it has been fixed.

[wine-plum]

We used 40 A100 GPUs. Thank you for pointing out the issue; it has been fixed.

Hi, could you please share the approximate time required for each stage during the training process?