kaiokendev

Results 2 issues of kaiokendev

This PR showcases a proof-of-concept extension that generalizes the idea of using a vectorstore to index large documents to fake a larger (fuzzy/lossy) context window by dumping the prompt into...

Using text-gen webui with Exllama loader gives me different results than with Exllama_HF. Specifically, Exllama_HF gives gibberish with SuperHOT 8K models past 2048 tokens. Even the logits of the two...