Module compat 

The Laya compat graph on the reference kernels: ModernBERT or mmBERT, the type embedding, the two layer decision head, the option scorer and the act head, as DecisionModel.forward in Laya 0.3.7 runs them in FP32.

This is the correctness reference. It allocates its buffers per call and converts the F16 weights to f32 once at load, which costs 1.7 GB for Laya English. The fast paths are checked against it.

Structs

Compat

A compat checkpoint ready to run on the CPU.

Input

One question, laid out as Laya lays it out.

Output

What the model says about one question, before temperature.

Functions

act_features

[top1, top1 - top2, entropy / ln k, k / 255] over the softmax of the logits, with k at least 2, as Laya computes them. A single option has a top2 of 0. No options at all is an error in Laya, and gives zeros here apart from the count.

executor

The compat graph of model on the plan executor with the default bucket table, on threads threads. This is the fast path, and Compat is what it is checked against.

executor_from

executor from the pieces of a model.

executor_with

executor on a backend the caller set up, such as one with_int8.