[BUG] Macbook M2 Apple Sillicon, pytorch operator is not currently implemented for the MPS device.

[kdgutier]

The new Macbooks have Metal Performance Shaders (MPS) backend for GPU training acceleration. PyTorch is yet to use fully the capabilities of the hardware, and some functionalities are not available. A solution is to set the PYTORCH_ENABLE_MPS_FALLBACK=1 environment variable to execute the code on the CPU if Mac GPU's code is not available.

dataset, *_ = TimeSeriesDataset.from_df(df = Y_train_df)
model = NHITS(h=24,
              input_size=24*2,
              max_steps=1,
              windows_batch_size=None, 
              n_freq_downsample=[12,4,1], 
              pooling_mode='MaxPool1d')
model.fit(dataset=dataset)
y_hat = model.predict(dataset=dataset)

NotImplementedError: The operator 'aten::upsample_linear1d.out' is not currently implemented for the MPS device. If you want this op to be added in priority during the prototype phase of this feature, please comment on https://github.com/pytorch/pytorch/issues/77764. As a temporary fix, you can set the environment variable `PYTORCH_ENABLE_MPS_FALLBACK=1` to use the CPU as a fallback for this op. WARNING: this will be slower than running natively on MPS.

The following line in the terminal with the respective neuralforecast environment activated solves the issue.

(neuralforecast) user@path % export PYTORCH_ENABLE_MPS_FALLBACK=1

In a jupyter notebook the same effect is achieved with the following line:

import os
os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"

[kdgutier]

It is important to notice that the numerical stability of MPS operators is not guaranteed. And some normal operations can fail to converge.

It is recommended to use a GPU on a Linux machine and debug on the Macbook M2 machine.