Inquiry About Wan-I2V Training/Inference Performance on A6000 GPUs

[ZhouQianang]

Hi @Artiprocher,

I'd like to consult about the training and inference speeds of Wan-I2V-14B-480P. My setup consists of 4×A6000 (49GB GPUs). After installing Diffsynth-Studio, I ran the example code test and observed the following performance:

wan-1.3B-T2V: ~5 minutes per video generation

wan-14B-I2V-480P:

~50 minutes for 81 frames (bfloat16, 50 iterations)

~37 minutes for 21 frames

My questions:

1. Baseline Validation: Are these inference times normal?

2. Inference Acceleration: Is multi-GPU parallelization supported for inference? (I couldn't find related documentation)

3. Training Acceleration: The current 50min/it training speed is impractical. Are there optimization strategies?

Thank you for your help!

[Artiprocher]

My answers:

1. Yes.
2. Please refer to the usp script.
3. If your GPU memory is sufficient, we recommend disabling gradient checkpointing to achieve faster performance.

[MukundVarmaT]

Hi,
I wanted to perform some full finetuning expts using WAN I2V, is there any scripts that you can point me to for the same.

Thanks