Authors:
(1) Suzanna Sia, Johns Hopkins University;
(2) David Mueller;
(3) Kevin Duh.
Speeding up transformer inference is of great interest to the community (Fournier et al., 2023). We highlight the potential of speeding up inference time as a direct consequence of identifying where task recognition occurs in the model and redundancy of self-attention processing. Our results indicate that we can achieve significant speedups in inference by removing the processing of context-tokens all-together after a certain point in the model, with little to no impact on downstream performance.
Then, for a model with nℓ layers, the amount of processing in terms of speed and memory saved is approximately (nℓ − r)/nℓ × (k/k + 1).
Using the example of LLAMA7B (32 layers), we see from Figure 2 that the model is very close to it’s ceiling score after processing the examples at layer 14 (ℓ = 14). If we no longer need to process examples after ℓ = 14, under a prompt size of 5 the savings are approximately 45%.
For instruction-tuned models which are typically deployed in production, even if we assume that no examples are provided, savings can be non-trivial as very long-form instructions are typically provided to the model in an attempt to control it’s behavior (prompt engineering).
This paper is available on arxiv under CC BY 4.0 DEED license.