3 ms·
> Speculative decoding is not really worthwhile for sparsely-loaded models. If wonder if you can train a model to optimize this, by trying to make the expert s
by amluto 3mo ago
> Speculative decoding is not really worthwhile for sparsely-loaded models.
If wonder if you can train a model to optimize this, by trying to make the expert selection sticky across a few tokens, without too much quality loss.
Another fun idea might be to try to build a model where the router chooses the expert 1-3 tokens in advance.
- johndough 3mo ago> If wonder if you can train a model to optimize this, by trying to make the expert selection sticky across a few tokens You can! > AFM 3 Core Advanced makes routing decisions per prompt. A lightweight, dense block selects a fixed set of experts during initial processing, periodically reselecting them during generation. https://machinelearning.apple.com/research/introducing-third-generation-of-apple-foundation-models https://machinelearning.apple.com/research/introducing-third...