Text Generation
Transformers
Safetensors
falcon
conversational
text-generation-inference

RoPE scaling?

#6
by wolfram - opened

Is extending context size via RoPE scaling supported? I've increased frequence base and frequence scale to get 4K instead of 2K context, but response quality degraded severely.

I'd really like to use the model, but 2K context just isn't enough for longer conversations. Intelligence unfortunately doesn't mean much when its memory is limited to just the last few messages.

Sign up or log in to comment