Package io.github.kirstenali.deepj.models

Interface CausalLM

All Superinterfaces:

Trainable

All Known Implementing Classes:

DecoderOnlyModel, DeepSeekModel, GPTModel, LlamaModel

public interface CausalLM
extends Trainable

Common contract for decoder-only causal language models (GPT, Llama, DeepSeek, …).

Extends Trainable with the three additional methods that every causal-LM training loop needs:

• forward(int[]) — token ids → logits [seqLen × vocabSize]
• backward(io.github.kirstenali.deepj.tensor.Tensor) — back-propagate gradient of the logits
• gradClipNorm() — global gradient-clipping threshold

Method Summary

Modifier and Type	Method	Description
void	backward(Tensor dLogits)	Back-propagates gradient of the logits through the model.
Tensor	forward(int[] inputIds)	Maps token ids to logits [seqLen × vocabSize].
float	gradClipNorm()	Global gradient-clipping threshold (e.g.

Methods inherited from interface io.github.kirstenali.deepj.training.Trainable

parameters, zeroGrad

Method Details

forward

Tensor forward(int[] inputIds)

Maps token ids to logits [seqLen × vocabSize].

backward

void backward(Tensor dLogits)

Back-propagates gradient of the logits through the model.

gradClipNorm

float gradClipNorm()

Global gradient-clipping threshold (e.g. 1.0).