Reference Rework

.github/workflows/run_pytest.yaml

@@ -48,4 +48,4 @@ jobs:
         run: |
           pip install .[dev]
           pip list
-          pytest --skip-gpu --psiflow-config=.github/threadpool.yaml
+          pytest --skip-gpu 

configs/local_test.yaml

@@ -0,0 +1,32 @@
+---
+parsl_log_level: WARNING
+retries: 0
+make_symlinks: false
+
+ModelEvaluation:
+  gpu: false
+  use_threadpool: false
+  max_simulation_time: 1
+
+ModelTraining:
+  gpu: true
+  use_threadpool: true
+  max_training_time: 1
+  max_workers: 1  # suppress assertion for multigpu training
+
+CP2K:
+  cores_per_worker: 1
+  max_evaluation_time: 0.1
+  container_uri: 'oras://ghcr.io/molmod/cp2k:2024.1'
+
+GPAW:
+  cores_per_worker: 1
+  max_evaluation_time: 0.1
+  container_uri: 'oras://ghcr.io/molmod/gpaw:24.1'
+
+ORCA:
+  cores_per_worker: 1
+  max_evaluation_time: 0.1
+
+
+...

psiflow/__init__.py

@@ -33,3 +33,42 @@ def resolve_and_check(path: Path) -> Path:
 load = ExecutionContextLoader.load
 context = ExecutionContextLoader.context
 wait = ExecutionContextLoader.wait
+
+
+# TODO: EXECUTION
+#  - max_runtime is in seconds, max_simulation_time (and others) is in minutes
+#  - ExecutionDefinition gpu argument?
+#  - ExecutionDefinition wrap_in_timeout functionality
+#  - ExecutionDefinition wrap_in_srun functionality? Actually for MD this is more iffy right
+#  - ExecutionDefinition why properties?
+#  - ExecutionDefinition centralize wq_resources
+#  - configuration file with all options
+#  - timeout -s 9 or -s 15?
+#  - executor keys are hardcoded strings..
+#  - define a 'format_env_variables' util
+#  - update GPAW + ORCA + Default containers
+#    include s-dftd3 in modelevaluation + install s-dftd3 with openmp
+#  - what with mem_per_node / mem_per_worker
+#  - always GPU for training?
+#  - currently reference mpi_args have to be tuple according to typeguard..
+#  - cores_per_block has to be specified even when exclusive..?
+#  - can we do something with WQ priority?
+#  - see chatgpt convo for process memory limits and such
+#  - make /tmp for app workdirs an option?
+#  - what is scaling_cores_per_worker in WQ
+#  - can we clean up psiflow_internal slightly?
+#  -
+# TODO: REFERENCE
+#  - reference MPI args not really checked
+#  - mpi flags are very finicky across implementations --> use ENV args?
+#    OMPI_MCA_orte_report_bindings=1 I_MPI_DEBUG=4
+#  - commands ends with 'exit 0' - what if we do not want to exit yet?
+#  - some actual logging?
+#  - safe_compute_dataset functionality?
+#  - ReferenceDummy is a bit sloppy
+#  -
+# TODO: MISC
+#  - think about test efficiency
+#  - some imports take a very long time
+#  - fix serialisation
+#  -

psiflow/data/dataset.py

@@ -284,6 +284,14 @@ def evaluate(
         Returns:
             Dataset: A new Dataset with evaluation results.
         """
+        # TODO: WIP
+        from psiflow.hamiltonians import Hamiltonian
+
+        if not isinstance(computable, Hamiltonian):
+            # avoid extracting and inserting the same quantities
+            return computable.compute_dataset(self)
+
+        # use Hamiltonian.compute method
         if batch_size is not None:
             outputs = computable.compute(self, batch_size=batch_size)
         else: