Google

During the prefilling phase for an input sequence $$\mathbf{x}$$, we generate a sequence of prefixes and their corresponding Key-Value (KV) cache states. This mapping is defined as:

$$\begin{matrix} x_0\ (\mathbf{x}_{<1}) & \Rightarrow & \mathrm{cache}_{<1} \\ x_0 x_1\ (\mathbf{x}_{<2}) & \Rightarrow & \mathrm{cache}_{<2} \\ & ... & \\ x_0 x_1 ... x_{m-1}\ (\mathbf{x}_{<m}) & \Rightarrow & \mathrm{cache}_{<m} \end{matrix}$$

where $$\mathrm{cache}_{<i}$$ denotes the KV cache state for the prefix $$\mathbf{x}_{<i}$$. All these mappings can be stored in the prefix cache for efficient reuse.

Formula for Prefix Cache State Generation

An auto-regressive model is processing the input sequence of tokens: `['The', 'cat', 'sat']`. When the model uses the prefix `['The', 'cat']` to generate the next token, `'sat'`, what is the content of the corresponding Key-Value (KV) cache state that is created at this step?

A system has pre-computed and stored the Key-Value (KV) cache state for the prefix `('Once', 'upon', 'a', 'time')`. Now, a new input sequence `('Once', 'upon', 'a', 'star')` needs to be processed. Based on the principle of how prefix cache states are generated, which pre-computed cache state can be reused to accelerate the processing of the new sequence, and why?

Prefix Cache Reuse Scenario

An auto-regressive model is generating a series of Key-Value (KV) cache states for the input sequence of tokens: `['The', 'quick', 'brown']`. Arrange the following events in the correct chronological order in which they occur during this process.

Learn Before

Related