I'm not sure you're actually doing the math for these long contexts. A naked transformer generating 1k tokens with 1k prompt is spending all its time doing a bu

I'm not sure you're actually doing the math for these long contexts. A naked transformer generating 1k tokens with 1k prompt is spending all its time doing a bunch of forward passes to generate each token- that's what's driven your intuition. A naked transformer generating 1k tokens with 1M prompt is spending all its time generating the embeddings for the prompt (filling the kv cache), and then the iterating generation at the end is a tiny fraction of the compute even if you have to run it 1k times