NVIDIA Dory1.0-4B fails standard setup with 'dory' error

NVIDIA released Dory1.0-4B as a 'loop-type Transformer' that lets users switch the number of times the recurrent block loops between 1 and 4 times, trading inference compute for accuracy without retraining. The model card presents simple instructions: update transformers via pip, or use a special llama.cpp fork for GGUF. The author tried the simplest path first and hit a wall: both transformers and stock llama.cpp could read the files correctly but did not implement the dory architecture.

A local GPU test then confirmed that even the error message's own advice—reinstalling transformers from GitHub main (5.23.0.dev0)—still lacked dory support, so that workaround did not function as of Nov 5, 2026. Only the model card's second workaround, building the pinned fork of llama.cpp, succeeded, and even then, on a simple arithmetic problem, increasing loop count from 1 to 4 did not reduce output tokens and made generation take about 3.2 times longer. The author suggests the paper's reported token-reduction effect likely requires more complex reasoning tasks to appear.