Cisco TSFM Pipeline #663
Description
We propose adding the Cisco Time Series Foundation Model pipeline into Orion as a pre-trained model. This would introduce a new primitive wrapping Cisco TSFM. Note that Cisco did not provide a PyPI interface for their project, so this is created based on their own code. We did not vendor their code but rather used the necessary parts for prediction tasks. This includes the forward pass of the model and creation of long-term and short-term contexts. For this PR, we do not support batching, but a future PR may include it. The primitive will integrate into Orion’s collection of forecasting-based anomaly detection methods. The primitive will only support univariate inputs and outputs.
We would also introduce a new pipeline which uses the primitive as its prediction step. We can use this pipeline to perform anomaly detection and include it in our benchmark suite.