Comment by shubhamgrg04

Isn't condensing rotary embeddings just a clever hack to circumvent the original training limitations? Is it really a sustainable solution for extending context length in language models?