SlimSAM
SlimSAM copied to clipboard
The MACs result doesn't contain q@k and attn@V
The calculation workload of matrix multiplication in the attention modules(qk and attnV) which is nonnegligible, is not included in torch_pruning.utils.count_ops_and_params。 torchprofile can calculate the macs of q@k and k@v.
Hi @Yiman-GO , Thanks for the issue. I will check it with the author of torch pruning.