Satellite-borne large model speculation decoding method and device based on pre-generated token
The invention discloses a satellite-borne large model speculation decoding method and device based on pre-generated tokens, and the method comprises the steps: 1) carrying out the offline pre-generation of next tokens of all tokens, and obtaining a pre-generated token dictionary pair; and 2) during...
Gespeichert in:
Hauptverfasser: | , , , |
---|---|
Format: | Patent |
Sprache: | chi ; eng |
Schlagworte: | |
Online-Zugang: | Volltext bestellen |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | The invention discloses a satellite-borne large model speculation decoding method and device based on pre-generated tokens, and the method comprises the steps: 1) carrying out the offline pre-generation of next tokens of all tokens, and obtaining a pre-generated token dictionary pair; and 2) during reasoning, verifying the guessed sequence through the guessed token sequence in the step 1) and using a text generation model, when the model generates a token through online reasoning, querying the token according to a token dictionary pair, guessing the generated token, online verifying the correctness of the token guessed by a token pre-generation module, receiving the guessed token after the verification is successful, and performing the step 3). Therefore, the accelerated generation of the token decoding of the large model is realized. According to the method, during online reasoning, token query guessing is carried out, verification is carried out, the number of correct tokens obtained during each reasoning o |
---|