Hi everyone, why neuralforecast TFT implemetation does not support passing multipletargets while darts and pytorch-forecasting does? Is there any workarounds

Unfortunately it's not supported. The workaround is to just use it as a univariate model. If you open an issue, we can have a look and keep it on our radar to implement in the future.

Thank you for the answer. Because NeuralForecast’s TFT doesn’t support multi-target, I also experimented with PyTorch-Forecasting’s TFT, using as close to the same hyperparameters as possible. However, my metrics still don’t come close to what I get with NeuralForecast. Are there any significant differences between these two implementations that could explain the performance gap?

I'm not familiar with PT's implementation, unfortunately. But the difference between a univariate and multivariate model in runtime can be massive, especially if you have a lot of timeseries. So running TFT in multivariate mode can be much slower. Does your use case require TFT or is the multivariate part more important? If the latter is the case I'd suggest to use TSMixer or TSMixerx.