Shaswat

Scrollbar customization is done in page.js and delete is done in index.js and task-action.js

I experimented with -inf and this is correct way to mask out attention head. I see the retrieval scores associated with head is 0 so I guess with -inf works!

Hey @HayeonLee , I faced the similar issue. I solved it by using the following code: https://huggingface.co/microsoft/Phi-3.5-mini-instruct/blob/main/modeling_phi3.py You need to add Phi3LongRoPEScaledRotaryEmbedding class from the above code into modeling_phi3.py file....

@nightdessert The kwargs aren't flowing to forward_torch. We had to make few modifications to flow the block_list to forward_torch. ```python # Self Attention if (attn_mode == "flash"): hidden_states, self_attn_weights, present_key_value...