Encountering an issue with the TFT/DeepAR/LSTM,etc. model

[sdmishra123]

Description

Hello Team,

I'm encountering an issue while running the Temporal Fusion Transformer (TFT)/ DeepAR/LSTM, etc. except for NBEATS/NHits models from the Neural Forecast package. Despite trying to force the model to run on CPU, the problem persists. Are there any alternative solutions or workarounds that I can explore to resolve this issue?

Best,
Sapna

**nf.fit(df=Y_train_df, val_size=val_size)

NotImplementedError: The operator 'aten::nanmedian.dim_values' is not currently implemented for the MPS device. If you want this op to be considered for addition please comment on pytorch/pytorch#141287 and mention use-case, that resulted in missing op as well as commit hash 2236df1770800ffea5697b11b0bb0d910b2e59e1. 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.**

Use case

No response

[marcopeix]

Hello! This not an issue related to neuralforecast, but related to running PyTorch code on a Mac. As the error suggests, some operations do not run on the MPS backend. So, you need to set PYTORCH_ENABLE_MPS_FALLBACK=1 as an environment variable.

In a notebook, you can run this as the first cell (very important that this cell is at the top of the notebook before anything else): %set_env PYTORCH_ENABLE_MPS_FALLBACK=1.

If you run a Python script, you can also do PYTORCH_ENABLE_MPS_FALLBACK=1 python main.py.

You can also do this before running your script:

export PYTORCH_ENABLE_MPS_FALLBACK=1

Or set it inside your Python script:

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

I hope this helps!

[sdmishra123]

Oh, I see. I was initializing ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' just before fitting the model and so the error was't fixed. Thanks! Appreciate it!