Question about loading from checkpoint

[wangq326]

Hello，

I am trying to run your example/moirai_forecast_pandas.ipynb script, but when I try to load the checkpoint with the following statement

model = MoiraiForecast(
    module=MoiraiModule.from_pretrained(f"loacalfoler/moirai-1.0-R-small"),
    prediction_length=48,
    context_length=168,
    patch_size=32,
    num_samples=100,
    target_dim=1,
    feat_dynamic_real_dim=0,
    past_feat_dynamic_real_dim=0,
)

I encountered an error AttributeError: type object 'MoiraiModule' has no attribute 'from_pretrained'. I checked the current code, in the module.py script,

class MoiraiModule(
    nn.Module,
    PyTorchModelHubMixin,
    coders={DistributionOutput: (encode_distr_output, decode_distr_output)},
):
    def __init__(...):
        ...
    def forward(...):
        ...

and in the forecast.py script:

class MoiraiForecast(L.LightningModule):
    ...

Since I am not familiar with the L.LightningModule class, I am unsure whether it is appropriate to directly add L.LightningModule to the MoiraiModule inheritance, and whether it might cause other interference?

In addition, if I mimic the previous example script and execute

model = MoiraiForecast.load_from_checkpoint(
    checkpoint_path=f"localfolder/Moirai-1.0-R-small/model.ckpt",
    prediction_length=48,
    context_length=168,
    patch_size=32,
    num_samples=100,
    target_dim=1,
    feat_dynamic_real_dim=0,
    past_feat_dynamic_real_dim=0,
    map_location="cuda:0" if torch.cuda.is_available() else "cpu",
)

then the checkpoint can be loaded successfully.

This issue discussed the relevant content, but the error messages are not consistent.

[gorold]

2 things to do:

1. pull the most recent commit
2. re-install uni2ts module / update the huggingface_hub and safetensors packages