Merge branch 'lp/ray-distributed' into main

Author: ludgerpaehler

docs/RayClusterSetup.md

@@ -0,0 +1,45 @@
+# Ray Cluster Setup
+
+## Manual Ray Cluster
+
+Especially for a manual cluster we need to consider a few basic principles:
+
+- All machines need to have **the same** basic setup with the package pre-installed, and all dependencies present
+
+We can then initialize the headnode
+
+```bash
+ray start --head
+```
+
+and subsequently attach all the worker nodes
+
+```bash
+ray start --address=IP_OF_THE_HEAD_NODE
+```
+
+> If a fast interconnect is available e.g. higher-bandwidth ethernet, Omnipath, Infiniband, etc. consider using the custom IPs of that interconnect to force the usage of the faster connections.
+
+With the cluster now initialized we can use it in any of our scripts With
+
+```python
+import ray
+ray.init(address=IP_OF_THE_HEAD_NODE)
+```
+
+
+## Ray Cluster on Kubernetes
+
+Ray can easily be installed on Kubernetes with [Helm](https://helm.sh) using the Ray Kubernetes Operator, the detailed documentation for which can be found in [Ray's documentation](https://docs.ray.io/en/latest/cluster/kubernetes.html).
+
+> On clouds like AWS, Azure, and GCP this is entirely straightforward and automated away.
+
+## Ray Cluster on Slurm
+
+Ray on Slurm is currently still experimental in its support, but the progress and current approach can be viewed in [Ray's documentation](https://docs.ray.io/en/latest/cluster/slurm.html). The biggest difference here coming from the different way Ray and Slurm treat 
+
+1. Ports binding
+2. IP binding
+
+But there exist example [launch script](https://docs.ray.io/en/latest/cluster/examples/slurm-launch.html#slurm-launch), and [templated launch scripts](https://docs.ray.io/en/latest/cluster/examples/slurm-template.html#slurm-template).
+

hydrogym/core.py

@@ -196,6 +196,27 @@ def render(self, **kwargs):
         raise NotImplementedError
 
 
+
+@ray.remote
+class EvaluationActor:
+    """To remotely evaluate Firedrake solutions."""
+
+    def __init__(self, firedrake_instance: "Firedrake_instance", index: int, seeds: Union[ActorSeeds, tuple], state:dict):
+        """
+        Initialize a remote runner for a Firedrake problem instance.
+
+        Args:
+            firedrake_instance: The Firedrake problem instance to be run.
+            index: The index of the actor in question.
+            seed: An integer to be used as the seed.
+            state: The state dictionary to be loaded.
+        """
+       
+       # Add whole class and the two following to have the evaluation logic.
+       # Pipe Firedrake problem into Evotorch's harness, then simmer down the logic needed to have this 
+       #   code here running.
+
+
 class CallbackBase:
     def __init__(self, interval: int = 1):
         """