Capability aware language model with improved computational efficiency

A system comprising:

• a computer processor;
• a data repository in communication with the computer processor and storing:
• an input,
• a first boundary token and a second boundary token, wherein:
• the first boundary token comprises a first token that a language model uses during execution to indicate that a selected execution phase is finished and that a subsequent execution phase will be initiated, and
• the second boundary token comprises a second token that the language model uses during execution to indicate that the subsequent execution phase is finished,
• a designated pruning weight mask and a successive pruning weight mask, wherein the successive pruning weight mask is different than the designated pruning weight mask, and
• a plurality of tokens;
• the language model in communication with the computer processor and having a first computational efficiency, wherein the language model, when executed by the computer processor, is programmed to:
• process the language model on the input,
• generate the first boundary token during the selected execution phase of the language model, wherein the selected execution phase comprises a first stage of execution of the language model to perform a first task,
• generate the second boundary token during the subsequent execution phase of the language model, wherein the subsequent execution phase is after the selected execution phase, and wherein the subsequent execution phase comprises a second stage of execution of the language model to perform a second task, subsequent to the first task,
• generate, during the selected execution phase with the designated pruning weight mask and during the subsequent execution phase with the successive pruning weight mask, a plurality of tokens, and
• return the plurality of tokens;
• and
• a server controller configured, when executed by the computer processor, to:
• receive the input,
• receive, during processing of the language model, the first boundary token,
• modify, prior to the selected execution phase, the language model with the designated pruning weight mask,
• receive, while processing, the second boundary token, and
• modify, prior to the subsequent execution phase, the language model with the successive pruning weight mask,
• wherein at least one of applying the designated pruning weight mask during the selected execution phase, and applying the successive pruning weight mask during the subsequent execution mask, modifies the language model to a modified language model comprising a final computational efficiency greater than the initial computational efficiency.