[Bug] Wan2.1-14B-I2V VSA acceleration is not satisfactory

[qinziqiao]

Describe the bug

I attempted to use VSA in Wan2.1-14B-I2V-720P with a high sparsification rate of 0.95. However, during the inference process, H20 inference only accelerated by 3 times, while H100 inference accelerated by 2 times. Is this normal?

Reproduction

models: Wan2.1-14B-I2V-720P

Environment

pytorch==2.8.0 vsa==0.03

[Lyxien]

Which script are you using? As far as I remember, training DMD + VSA on Wan2.1-14B-I2V-720P isn’t supported at the moment.

[qinziqiao]

@Lyxien hi, I did not use DMD training. Instead, we replaced Wan finetuning with VSA for speed measurement.