proger / hippogriff

Griffin MQA + Hawk Linear RNN Hybrid
https://pypi.org/project/hippogriff
MIT License
79 stars 5 forks source link

variable sequence length #3

Open ruoyxue opened 3 months ago

ruoyxue commented 3 months ago

It seems that you utilize accelerated_scan.warp.scan to accelerate the RG-LRU in Griffin, but it only supports sequence length to be a power of 2, is there any solution to support variable sequence length?

proger commented 3 months ago

Hi @ruoyxue, I just pad the sequence length to the nearest power of 2 for now. If you need other sizes for efficiency, could you make an issue and describe your use case in https://github.com/proger/accelerated-scan?