Lior Khermosh

Results 6 comments of Lior Khermosh

This was also the discussion in the subWG. it is a general problem not just for BERT. there is preprocessing of the datasets. In BERT it creates the split, so...

https://github.com/mlcommons/inference_policies/pull/264

I see no problem to send token ID. I wouldn't specify the other optimizations. I also don't think it is in the category of compression. I think we should try...

it makes sense to add the 3 inputs I will clarify the text of the rule. @nv-jinhosuh What do you mean "texts can be kept in the SUT node."? and...

To capture the past discussion in the WG regarding compression - it was targeting schemes that will reduce the networking data rate, allowing vendors to tradeoff compute and network bandwidth....

This topic was discussed in the Network subWG. it was agreed to allow input of either token IDs, Input Masks and segment IDs or just the token IDs and deduce...