Google

In autoregressive models, the generation of each token is causally dependent on all previously generated tokens. This sequential dependency means that the computation for a given token cannot begin until the computation for the preceding token is complete. As a result, there is an inherent delay in predicting subsequent tokens in a sequence; for instance, the prediction of the second output token is delayed until the first has been fully generated.

Latency from Sequential Dependency in Autoregressive Generation

A company's real-time translation service uses an autoregressive model. Users complain about the delay in receiving the full translation for long sentences. The model generates the translated text one word at a time, where each new word can only be determined after the previous one has been generated. The engineering team proposes the following solution to address the delay: 'We will run ten copies of the model in parallel. This will allow us to handle ten different user translation requests at the same time.' Critically evaluate this proposal. Will it solve the specific user complaint about the delay for a single, long translation? Justify your reasoning.

Evaluating a Performance Optimization Strategy

True or False: For an autoregressive text generation model, doubling the number of parallel processing units available for computation will cut the total time required to generate a 100-token sequence in half.

A team is comparing two text generation systems to produce a 10-token sequence.

- **System A** generates tokens one after another. The computation for each token takes 100ms.
- **System B** is a hypothetical system that can compute all 10 tokens simultaneously, with each token's computation also taking 100ms.

Why does System A take approximately 10 times longer than System B to produce the full sequence?

Learn Before

Related