add mlipx serve

.gitignore

@@ -154,6 +154,9 @@ dmypy.json
 # Cython debug symbols
 cython_debug/
 
+# hatch-vcs generated version file
+mlipx/_version.py
+
 # PyCharm
 #  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
 #  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore

docs/source/concept.rst

@@ -39,6 +39,7 @@ providing a quick overview of all past evaluations.
    concept/data
    concept/models
    concept/recipes
+   concept/serve
    concept/zntrack
    concept/zndraw
    concept/metrics

docs/source/concept/models.rst

@@ -72,22 +72,63 @@ The :code:`SevenCalc` class, used in the example above, is defined in :code:`src
 
 For more details, refer to the :ref:`custom_nodes` section.
 
+.. _serve-integration:
+
+Serve Integration
+-----------------
+
+The dependencies of MLIPs can be incompatible, making it impossible to run multiple MLIPs in the same environment.
+To address this, ``mlipx`` can circumvent this using ``uv`` (https://docs.astral.sh/uv/) in combination with ZeroMQ (https://zeromq.org/) to automatically serve models from different environments.
+
+To make a model available for serving, you need to make it installable through an extra for ``mlipx``.
+For this, you need to configure a new extra in the ``mlipx`` ``pyproject.toml`` using ``[tool.uv]
+conflicts`` to specify dependencies that cannot be installed in the same environment.
+
+Additionally, you need to include this extra in the model definition.
+
+.. code-block:: python
+
+   import mlipx
+
+   ALL_MODELS["mace-mpa-0"] = mlipx.GenericASECalculator(
+       module="mace.calculators",
+       class_name="mace_mp",
+       device="auto",
+       extra=["mace"],  #  mlipx extras for this model
+   )
+
+**Using served models:**
+
+.. code-block:: python
+
+   # Environment variable (global control)
+   import os
+   os.environ["MLIPX_USE_SERVE"] = "true"
+
+   # Or explicit per-model control
+   calc = model.get_calculator(use_serve=True)
+
+See :ref:`serve` for details on starting brokers, managing workers, and DVC integration.
+
 .. _update-frames-calc:
 
 Updating Dataset Keys
 ---------------------
 
-In some cases, models may need to be defined to convert existing dataset keys into the format :code:`mlipx` expects.
-For example, you may need to provide isolated atom energies or convert data where energies are stored as :code:`atoms.info['DFT_ENERGY']`
-and forces as :code:`atoms.arrays['DFT_FORCES']`.
+In some cases, you may need to convert existing dataset keys into the format :code:`mlipx` expects.
+For example, your dataset may store energies as :code:`atoms.info['DFT_ENERGY']` and forces as :code:`atoms.arrays['DFT_FORCES']`.
 
-Here’s how to define a model for such a scenario:
+Use :code:`mlipx.UpdateFramesCalc` to remap these keys:
 
 .. code-block:: python
 
     import mlipx
 
-    REFERENCE = mlipx.UpdateFramesCalc(
+    # Remap existing dataset keys to mlipx format
+    data_remapper = mlipx.UpdateFramesCalc(
         results_mapping={"energy": "DFT_ENERGY", "forces": "DFT_FORCES"},
         info_mapping={mlipx.abc.ASEKeys.isolated_energies.value: "isol_ene"},
     )
+
+This is typically used in the :code:`metrics` recipe when comparing MLIP predictions against pre-computed reference data (e.g., DFT results).
+The generated :code:`main.py` includes a commented section showing how to configure this.