# DLRM training
###### tags: `Accelerator`

paper: [link](https://arxiv.org/pdf/2005.04680.pdf)

* In a typical DL framework, a training iteration consist of 3 passes.
1. forward pass which computes **loss** with respect to model parameters.
2. backward pass which computes **gradients** with respect to model parameters.
3. update pass (typically run by an optimizer) which takes learning rate and computed gradients from backward pass and **updates the model parameters**.