FAQs
In the following section, we address common questions and potential issues you might encounter. If your question is not answered here, please feel free to reach out to us by opening an issue on Github.
What is TiRex?
TiRex is a 35M parameter pre-trained time series forecasting model. It is built upon the xLSTM architecture and was developed by NXAI. For a comprehensive overview, please see our dedicated Introduction section.
How many parameters does TiRex have?
TiRex has 35 million (35M) parameters.
Why are my TiRex forecasts running slowly?
If you are experiencing slow forecasts, please check the following:
- Enable model compilation when loading the model.
- Use hardware acceleration devices like CUDA (for NVIDIA GPUs) or MPS (for Apple Silicon) whenever possible.
- For maximum speed, consider using our custom backends.
- Trade some accuracy for speed with
full_rollout=Trueordynamic_padding=True. See the forecasting workflow.
Example results for batch size 5, context length 128, and prediction length 64 using the torch backend (50-call average; lower MASE and CRPS are better):
| Option | CPU | GPU | MASE (GiftEval) | CRPS (GiftEval) |
|---|---|---|---|---|
| Default | 416 ms | 233 ms | 0.724 | 0.499 |
full_rollout | 223 ms (1.9×) | 116 ms (2.0×) | 0.753 (+4%) | 0.519 (+4%) |
dynamic_padding | 44 ms (9×) | 30 ms (8×) | 0.734 (+1.5%) | 0.508 (+2%) |
dynamic_padding helps less near the 2,048-step context limit, while full_rollout loses more accuracy at longer prediction lengths. See PR #34 for details.
For detailed instructions on optimization, please refer to the API. If you are interested in optimizing TiRex for dedicated hardware platforms (especially for edge or embedded use cases), please get in touch: contact@nx-ai.com
What is the maximum prediction length (context length) of TiRex?
Currently, TiRex supports a maximum context length of up to 2,048.
We are actively working on a version that supports longer contexts. In the meantime, if you need to forecast longer horizons, you may try downsampling your time series data.
To stay updated on the progress of this feature, please follow this Github issue.
Which data input formats does the framework support?
Out of the box, we support NumPy arrays and PyTorch tensors. To use other popular formats, you can install the respective extras:
- GluonTS datasets: Install with
tirex-ts[gluonts] - Hugging Face datasets: Install with
tirex-ts[hfdataset]
How can I finetune TiRex on my own data?
Fine-tuning is offered as a commercial service by NXAI. Please get in touch with us directly at contact@nx-ai.com to discuss your needs.