[contrib/ASP] Question about Sparsity performance use case

[sakaia]

I am trying to run ASP toy_problem.py. It seems nothing changes. Is there any method for seeing performance gain?

I am comparing train_loop/arg.num_xxx_steps for dense and sparse. It seems few percentage changes.

Another document for sparsity says, 50% performance gain on BERT (on MLPerf). But toy_problem.py seems no effect for sparsity. Of course BERT uses TensorRT for MLPerf, I understand software interface is different.

References

• How Sparsity Adds Umph to AI Inference

[calclavia]

Seems like the acceleration doesn't work from this code base - since there's no kernel implemented: https://github.com/NVIDIA/apex/blob/master/apex/contrib/sparsity/sparse_masklib.py#L57

[rg321]

Are you using Ampere architecuture? I suppose sparsity works for Ampere architectures for now