Introducing ><former
Most transformers are rectangles◻️: every layer has the same width
But is that optimal?🤔
We propose variable-width transformers that have different widths across layers, improving loss while cutting compute & KV cache size 🧵
@luke_wood_ml yes, but thats what you do in regular gradient descent. Lets say there are no batches, and you have N training examples, then the loss is 1/N * (per example loss). We don't sum the loss across all the training examples and then use that to update the gradients.
@AlaskaAir have flight credits that were booked for June 2021, and unable to use them due to covid travel ban. Email says expiry on 04/2022 but was told they wouldn't expire. Any help here ?