Hspix
Results
4
comments of
Hspix
> 可以参考huggingface上的讨论：https://huggingface.co/BAAI/bge-m3/discussions/22 当batch越来越大时会不会带来更明显的差异？
Got it.
> The latest `transformers` version has native support for Attention Sinks for Llama, Mistral, Phi and Persimmon :) This support doesn't require `attention_sinks`, and should stay working for future `transformers`...
The fix is pending？