How to do sparsity and quantization-aware-training ?

[Vieeo]

How to implement sparsification and quantization-aware-training , and keep the sparse mask unchanged during quantization.

[SilvesterHsu]

I guess this will be helpful: https://www.nvidia.com/en-us/on-demand/session/gtcspring21-s31552/
Original paper: https://arxiv.org/abs/2104.08378
Blog: https://developer.nvidia.com/blog/sparsity-in-int8-training-workflow-and-best-practices-for-tensorrt-acceleration/

[Vieeo]

I guess this will be helpful: https://www.nvidia.com/en-us/on-demand/session/gtcspring21-s31552/ Original paper: https://arxiv.org/abs/2104.08378 Blog: https://developer.nvidia.com/blog/sparsity-in-int8-training-workflow-and-best-practices-for-tensorrt-acceleration/

Thanks, it can be done with your method. Now, nvidia-modelopt doesn’t seem to keep the sparse mask unchanged during quantization. Do you know the sdk？

[yairb-gm]

I have the same question - Is it possible to freeze sparse mask while doing QAT?
@Vieeo have you find solution / workaround that can be applied?