5 ms·
They are almost certainly using some form of sparse attention. If you linearize the attention operation, you can scale up to around 1-10M tokens depending on ha
by lumost 3y ago
They are almost certainly using some form of sparse attention. If you linearize the attention operation, you can scale up to around 1-10M tokens depending on hardware before hitting memory constraints. Linearization works off the assumption that for a subsequence of X tokens out M tokens, where M os much greater than X there are likely only K tokens which are useful for the attention operation.
There are a bunch of techniques to do this, but it's unclear how well any of them scale.
- ein0p 3y agoNot "almost", but certainly. Dense attention is quadratic, not even Google would be able to run it at an acceptable speed. Their model is not recurrent - they did not have the time yet (or resources - believe it or not, Google of 2023-24 is very compute constrained) to train newer SSM or recurrent based models at practical parameter counts. Then there's the fact that those models are far harder to train due to instabilities, which is one of the reasons why you don't yet see FOSS recurrent/SSM models that are SOTA at their size or tokens/sec. With sparse attention, however, long context recall will be far from perfect, and the longer the context the worse the recall. That's better than no recall at all (as in a fully dense attention model which will simply lop off the preceding parts of the conversation), but not by a hell of a lot.
- kiraaa 3y agomaybe they are using ring attention, on top of their 128k model.
- ein0p 3y agoMore likely some clever take on RAG. There’s no way that 1M context is all available at all times. More likely parts of it are retrievable on demand. Hence the retrieval-like use cases you see in the demos. The goal is to find a thing, not to find patterns at a distance
- kiraaa 3y agocould be true, we can only speculate.