Text Generation
Transformers
PyTorch
TensorBoard
Safetensors
bloom
Eval Results
text-generation-inference
Inference Endpoints

What is the length of the prompt considered by BLOOM?

#236
by akratz - opened

See subject line. How long can a prompt be to be considered in its entirety? What happens if it exceeds some length, is the beginning cut off?

BigScience Workshop org
edited Apr 19, 2023

BLOOM was trained with sequences of length 2048, but uses ALiBi position embeddings (https://arxiv.org/abs/2108.12409) -- meaning it can be used with longer sequences.

So ALIBI allows input of what length to be considered? You write that it allows for “longer sequences” but this way I only know it is more than 2048…

Any update on this? @cakiki

BigScience Workshop org

I'd refer you to both the paper I linked to and empirical experimentation to answer that question. The practical limit will likely sooner be your hardware than any theoretical one.

christopher changed discussion status to closed

Sign up or log in to comment