CI: PyTorch Surrogate Example

.github/workflows/dependencies/gcc-openmpi.sh

@@ -32,4 +32,6 @@ python3 -m pip install -U -r src/python/impactx/dashboard/requirements.txt
 python3 -m pip install -U -r examples/requirements.txt
 python3 -m pip install -U -r tests/python/requirements.txt
 
+# extra tests
+python3 -m pip install -U -r examples/requirements_torch_cpu.txt
 python3 -m pip install -U openPMD-validator

.github/workflows/dependencies/gcc.sh

@@ -30,4 +30,6 @@ python3 -m pip install -U -r src/python/impactx/dashboard/requirements.txt
 python3 -m pip install -U -r examples/requirements.txt
 python3 -m pip install -U -r tests/python/requirements.txt
 
+# extra tests
+python3 -m pip install -U -r examples/requirements_torch_cpu.txt
 python3 -m pip install -U openPMD-validator

examples/CMakeLists.txt

@@ -116,6 +116,8 @@ function(add_impactx_test name input is_mpi analysis_script plot_script)
     else()
         set_property(TEST ${name}.run APPEND PROPERTY ENVIRONMENT "OMP_NUM_THREADS=2")
     endif()
+    # special return code for skipped tests (e.g., runtime prerequisite fails)
+    set_tests_properties(${name}.run PROPERTIES SKIP_RETURN_CODE 42)
 
     # analysis and plots
     set(THIS_Python_SCRIPT_EXE)
@@ -131,6 +133,11 @@ function(add_impactx_test name input is_mpi analysis_script plot_script)
 
         # make HDF5 I/O more robust on various filesystems
         set_property(TEST ${name}.analysis APPEND PROPERTY ENVIRONMENT "HDF5_USE_FILE_LOCKING=FALSE")
+
+        # run test failed? Mark this as skipped
+        set_property(TEST ${name}.analysis PROPERTY SKIP_REGULAR_EXPRESSION
+            "Supplied directory is not valid: diags"
+        )
     endif()
     if(plot_script)
         add_test(NAME ${name}.plot
@@ -141,6 +148,11 @@ function(add_impactx_test name input is_mpi analysis_script plot_script)
 
         # make HDF5 I/O more robust on various filesystems
         set_property(TEST ${name}.plot APPEND PROPERTY ENVIRONMENT "HDF5_USE_FILE_LOCKING=FALSE")
+
+        # run test failed? Mark this as skipped
+        set_property(TEST ${name}.plot PROPERTY SKIP_REGULAR_EXPRESSION
+            "ValueError: No objects to concatenate"
+        )
     endif()
 endfunction()
 
@@ -1000,6 +1012,7 @@ add_impactx_test(spectrometer.py
     OFF  # no plot script yet
 )
 
+
 # Chicane with CSR ###########################################################
 #
 if(ImpactX_FFT)
@@ -1097,6 +1110,7 @@ add_impactx_test(linac-segment.py
     OFF  # no plot script yet
 )
 
+
 # Iteration of a linear one-turn map #########################################
 #
 # w/o space charge
@@ -1112,3 +1126,14 @@ add_impactx_test(linear-map.py
     examples/linear_map/analysis_map.py
     OFF  # no plot script yet
 )
+
+
+# PyTorch Surrogate: Staged LPA ##############################################
+#
+add_impactx_test(pytorch_surrogate_model
+    examples/pytorch_surrogate_model/run_ml_surrogate_15_stage.py
+      OFF  # ImpactX MPI-parallel
+    examples/pytorch_surrogate_model/analyze_ml_surrogate_15_stage.py
+    examples/pytorch_surrogate_model/visualize_ml_surrogate_15_stage.py
+)
+label_impactx_test(pytorch_surrogate_model slow)

examples/pytorch_surrogate_model/run_ml_surrogate_15_stage.py

@@ -37,7 +37,7 @@
     print("Warning: Cannot import PyTorch. Skipping test.")
     import sys
 
-    sys.exit(0)
+    sys.exit(42)  # ImpactX special return code for skipped tests
 
 import zipfile
 from urllib import request
@@ -100,18 +100,19 @@ def download_and_unzip(url, data_dir):
 data_url = "https://zenodo.org/records/10810754/files/models.zip?download=1"
 download_and_unzip(data_url, "models.zip")
 
-# It was found that the PyTorch multithreaded defaults interfere with MPI-enabled AMReX
-# when initializing the models: https://github.com/AMReX-Codes/pyamrex/issues/322
+# It was found that the PyTorch multithreaded defaults interfere with AMReX OpenMP
+# when initializing the models or iterating elements:
+# https://github.com/AMReX-Codes/pyamrex/issues/322
+# https://github.com/ECP-WarpX/impactx/issues/773#issuecomment-2585043099
 # So we manually set the number of threads to serial (1).
-if Config.have_mpi:
-    n_threads = torch.get_num_threads()
-    torch.set_num_threads(1)
+# Torch threading is not a problem with GPUs and might work when MPI is disabled.
+# Could also just be a mixing of OpenMP libraries (gomp and llvm omp) when using the
+# pre-build PyTorch pip packages.
+torch.set_num_threads(1)
 model_list = [
     surrogate_model(f"models/beam_stage_{stage_i}_model.pt", device=device)
     for stage_i in range(N_stage)
 ]
-if Config.have_mpi:
-    torch.set_num_threads(n_threads)
 
 pp_amrex = amr.ParmParse("amrex")
 pp_amrex.add("the_arena_init_size", 0)
@@ -328,6 +329,7 @@ def set_lens(self, pc, step, period):
     lpa = LPASurrogateStage(i, model_list[i], L_surrogate, L_stage_period * i)
     lpa.nslice = n_slice
     lpa.ds = L_surrogate
+    lpa.threadsafe = False
     lpa_stages.append(lpa)
 
 monitor = elements.BeamMonitor("monitor")

examples/pytorch_surrogate_model/surrogate_model_definitions.py

@@ -106,9 +106,9 @@ class surrogate_model:
     def __init__(self, model_file, device=None):
         self.device = device
         if device is None:
-            model_dict = torch.load(model_file, map_location="cpu")
+            model_dict = torch.load(model_file, map_location="cpu", weights_only=False)
         else:
-            model_dict = torch.load(model_file, map_location=device)
+            model_dict = torch.load(model_file, map_location=device, weights_only=False)
         self.source_means = torch.tensor(
             model_dict["source_means"], device=self.device, dtype=torch.float64
         )

examples/requirements_torch_cpu.txt

@@ -0,0 +1,6 @@
+# This is for CPU CI tests with extra requirements.
+#
+# For PyTorch, see alternative packages, e.g., for GPU here:
+# https://pytorch.org/get-started/locally/
+--extra-index-url https://download.pytorch.org/whl/cpu
+torch