How to Run Inference via the CLI
================================

.. warning::
   The information below is incomplete and requires extra work. Use it at your own risk!

.. note::

   - On Apple Silicon (MPS), autocast is disabled by PyTorch; prefer ``--dtype float16``.

Even though it is recommended to use accelerated hardware (GPUs), CPU executions are also possible. It is assumed that
there is enough of the RAM available (>=32GB).

``--device auto`` chooses CUDA → MPS → CPU. Autocast is used on CUDA; for MPS, prefer ``--dtype float16``.

CLI offers a unified interface that can be combined in a set of command in a single bash script, for example.
This is the recommended way to execute inference with the provided models.

We provide a few examples on how to use YAML-based config files alongside the CLI. They can be found under
``/src/emmi_inference/examples/configs``. These configs should work for three model architectures (**Transolver**,
**UPT** and **AB-UPT**). The checkpoints to be acquired/produced independently.

There are two different datasets used: **ShapeNetCar** is used by **Transolver** and **UPT**, whilst **AB-UPT** uses
**DrivAerML**. This comes with different configs when it comes to the data part, as well as the collator. Please refer
to the example files for more details.

A working **Transolver** example is given below:

.. code-block:: bash

   emmi-infer run \
       /src/emmi_inference/examples/configs/example_config_transolver.yaml \
       transolver \
       "/data/transolver/e8ek00ze/checkpoints/transolver cp=latest model.th" \
       --device cpu

The list of model types is available within the ``ModelRegistry`` under ``/src/emmi_inference/models/registry.py``.

The following commands will run **UPT** and **AB-UPT** respectively:

.. code-block:: bash

   emmi-infer run \
       /src/emmi_inference/examples/configs/example_config_upt.yaml \
       upt \
       "/data/upt/to5u4s5i/checkpoints/upt cp=best_model.loss.test.total model.th" \
       --device cpu

.. code-block:: bash

   emmi-infer run \
       /src/emmi_inference/examples/configs/example_config_abupt.yaml \
       abupt \
       /data/abupt/checkpoints/ab-upt-drivaerml-tutorial.th \
       --device cpu

.. note::

   Make sure to update the paths for your datasets in the config files!

Other Links
-------------

:doc:`Inference API Reference <../../autoapi/emmi_inference/index>`