3 ms·
I would expect this only to be true for linear architectures like Mamba or Gated DeltaNet. Transformers and hybrid architectures do not have constant compute co
by cubefox 2mo ago
I would expect this only to be true for linear architectures like Mamba or Gated DeltaNet. Transformers and hybrid architectures do not have constant compute cost per token.
- monkpit 2mo agoPerformant could certainly mean “higher performing” and not “quicker”.