D-RTRL Examples#
The examples/drtrl/ series demonstrates the parameter-dimensional
diagonal RTRL implementation across several model and operator families.
01-basics-integrator.py introduces the integrator workflow. API: D-RTRL.
02-batching-vmap.py uses per-sample
vmapexecution. API: compile.03-batching-batched.py uses batched primitives. API: compile.
04-vjp-single-step.py demonstrates a single-step VJP. API: D-RTRL.
05-vjp-multi-step.py demonstrates a multi-step VJP. API: D-RTRL.
07-operator-lora.py covers a low-rank recurrent operator. API: LoRA.
08-operator-conv.py covers a convolutional ETP operator. API: Conv1d.
09-classification-mnist.py provides an MNIST classification task. API: LSTMCell.
10-char-lm-generation.py provides a character-language-model task. API: MiniGRU.
11-knob-fast-solve.py examines the
fast_solveimplementation option. API: D-RTRL.
Read the D-RTRL examples README and the D-RTRL tutorial alongside the scripts. D-RTRL uses the same ETP/compiler workflow across model architectures whose trainable paths are expressible with registered ETP operators. This interface-level generality does not remove the diagonal hidden-Jacobian approximation or establish general gradient equivalence with BPTT.