Logits Processor
A hook interface in large model inference frameworks that allows developers to intercept and modify raw logit score vectors during the token decoding stage, enabling capabilities such as constrained decoding and structured output generation.
Timeline (last 90 days)
The Structured Output feature required almost no changes to the core Transformers.js code, because it already supports the Logits Processor mechanism.
A Logits Processor is an interface that can hook into the token generation decoding step and modify the logits vector before softmax sampling.
All Facts (2)
The Structured Output feature required almost no changes to the core Transformers.js code, because it already supports the Logits Processor mechanism.
50%UnverifiedA Logits Processor is an interface that can hook into the token generation decoding step and modify the logits vector before softmax sampling.
50%