47 ms·
SparseGPT-for-LLaMA[0] exists. Pruning more than 30% of weights of 30B starts to show significant perplexity losses. 50% is a total disaster, while it was not f
by MacsHeadroom 3y ago
SparseGPT-for-LLaMA[0] exists. Pruning more than 30% of weights of 30B starts to show significant perplexity losses. 50% is a total disaster, while it was not for OPT or BLOOM. So your intuition seems to be good here.
[0] https://github.com/AlpinDale/sparsegpt-for-LLaMA https://github.com/AlpinDale/sparsegpt-for-LLaMA
- generalizations 3y agoHas anyone figured out what the optimal pruning is for 65b? I don't really know what that matrix in your link is saying, but it didn't seem to show optimal pruning.