Evaluation gets nan mae

I want to evaluate the performance of pretrained Graphormer using testing data and I use following code

python evaluate.py \
    --user-dir ../../graphormer \
    --num-workers 16 \
    --ddp-backend=legacy_ddp \
    --dataset-name pcqm4m \
    --dataset-source ogb \
    --task graph_prediction \
    --criterion l1_loss \
    --arch graphormer_base \
    --num-classes 1 \
    --batch-size 64 \
    --pretrained-model-name pcqm4mv1_graphormer_base \
    --load-pretrained-model-output-layer \
    --split test \
    --seed 1

It is similar to code in https://graphormer.readthedocs.io/en/latest/Quick-Start.html#evaluate-pre-trained-models, and the only difference is --split.

However, I get result like this. I don't know why.

Start with evaluate.py and the documented Quick-Start evaluation command, then compare the effects of using --split test instead of the documented split. Run the provided command and trace where the MAE is computed or reported; done means identifying why the test evaluation produces nan and documenting or correcting that behavior.