diff --git a/examples/gold_ball_ptycho_spawn.py b/examples/gold_ball_ptycho_spawn.py
new file mode 100644
index 0000000..ae13c9e
--- /dev/null
+++ b/examples/gold_ball_ptycho_spawn.py
@@ -0,0 +1,91 @@
+import cdtools
+from matplotlib import pyplot as plt
+import torch as t
+
+# If you're noticing that the multi-GPU job is hanging (especially with 100%
+# GPU use across all participating devices), you might want to try disabling
+# the environment variable NCCL_P2P_DISABLE.
+import os
+os.environ['NCCL_P2P_DISABLE'] = str(int(True))
+
+
+# The entire reconstruction script needs to be wrapped in a function
+def reconstruct(rank, world_size):
+
+    # In the multigpu setup, we need to explicitly define the rank and
+    # world_size. The master address and master port should also be
+    # defined if we don't specify the `init_method` parameter.
+    cdtools.tools.multigpu.setup(rank=rank,
+                                 world_size=world_size,
+                                 master_addr='localhost',
+                                 master_port='6666')
+
+    filename = 'example_data/AuBalls_700ms_30nmStep_3_6SS_filter.cxi'
+    dataset = cdtools.datasets.Ptycho2DDataset.from_cxi(filename)
+
+    pad = 10
+    dataset.pad(pad)
+    dataset.inspect()
+    model = cdtools.models.FancyPtycho.from_dataset(
+        dataset,
+        n_modes=3,
+        probe_support_radius=50,
+        propagation_distance=2e-6,
+        units='um',
+        probe_fourier_crop=pad
+    )
+    model.translation_offsets.data += 0.7 * \
+        t.randn_like(model.translation_offsets)
+    model.weights.requires_grad = False
+
+    # We need to manually define the rank parameter for the model, or else
+    # all plots will be duplicated by the number of GPUs used.
+    model.rank = rank
+
+    device = 'cuda'
+    model.to(device=device)
+    dataset.get_as(device=device)
+
+    # Rank and world_size also needs to be explicitly defined here
+    recon = cdtools.reconstructors.AdamReconstructor(model,
+                                                     dataset,
+                                                     rank=rank,
+                                                     world_size=world_size)
+
+    with model.save_on_exception(
+            'example_reconstructions/gold_balls_earlyexit.h5', dataset):
+
+        for loss in recon.optimize(20, lr=0.005, batch_size=50):
+            if rank == 0:
+                print(model.report())
+            if model.epoch % 10 == 0:
+                model.inspect(dataset)
+
+        for loss in recon.optimize(50, lr=0.002, batch_size=100,
+                                   schedule=True):
+            if rank == 0:
+                print(model.report())
+
+            if model.epoch % 10 == 0:
+                model.inspect(dataset)
+
+        for loss in recon.optimize(100, lr=0.001, batch_size=100,
+                                   schedule=True):
+            if rank == 0:
+                print(model.report())
+            if model.epoch % 10 == 0:
+                model.inspect(dataset)
+
+    cdtools.tools.multigpu.cleanup()
+
+    model.tidy_probes()
+    model.save_to_h5('example_reconstructions/gold_balls.h5', dataset)
+    model.inspect(dataset)
+    model.compare(dataset)
+    plt.show()
+
+
+if __name__ == '__main__':
+    # Specify the number of GPUs we want to use, then spawn the multi-GPU job
+    ngpus = 2
+    t.multiprocessing.spawn(reconstruct, args=(ngpus,), nprocs=ngpus)
diff --git a/examples/gold_ball_ptycho_torchrun.py b/examples/gold_ball_ptycho_torchrun.py
new file mode 100644
index 0000000..834cb84
--- /dev/null
+++ b/examples/gold_ball_ptycho_torchrun.py
@@ -0,0 +1,70 @@
+import cdtools
+from matplotlib import pyplot as plt
+import torch as t
+
+# At the beginning of the script we need to setup the multi-GPU job
+# by initializing the process group and sycnronizing the RNG seed
+# across all participating GPUs.
+cdtools.tools.multigpu.setup()
+
+# To avoid redundant print statements, we first grab the GPU "rank"
+# (an ID number between 0 and max number of GPUs minus 1).
+rank = cdtools.tools.multigpu.get_rank()
+
+filename = 'example_data/AuBalls_700ms_30nmStep_3_6SS_filter.cxi'
+dataset = cdtools.datasets.Ptycho2DDataset.from_cxi(filename)
+
+pad = 10
+dataset.pad(pad)
+dataset.inspect()
+model = cdtools.models.FancyPtycho.from_dataset(
+    dataset,
+    n_modes=3,
+    probe_support_radius=50,
+    propagation_distance=2e-6,
+    units='um',
+    probe_fourier_crop=pad
+)
+model.translation_offsets.data += 0.7 * t.randn_like(model.translation_offsets)
+model.weights.requires_grad = False
+device = 'cuda'
+model.to(device=device)
+dataset.get_as(device=device)
+
+recon = cdtools.reconstructors.AdamReconstructor(model, dataset)
+
+with model.save_on_exception(
+        'example_reconstructions/gold_balls_earlyexit.h5', dataset):
+
+    for loss in recon.optimize(20, lr=0.005, batch_size=50):
+        # We ensure that only the GPU with rank of 0 runs print statement.
+        if rank == 0:
+            print(model.report())
+
+        # But we don't need to do rank checking for any plotting- or saving-
+        # related methods; this checking is handled internernally.
+        if model.epoch % 10 == 0:
+            model.inspect(dataset)
+
+    for loss in recon.optimize(50, lr=0.002, batch_size=100):
+        if rank == 0:
+            print(model.report())
+
+        if model.epoch % 10 == 0:
+            model.inspect(dataset)
+
+    for loss in recon.optimize(100, lr=0.001, batch_size=100, schedule=True):
+        if rank == 0:
+            print(model.report())
+        if model.epoch % 10 == 0:
+            model.inspect(dataset)
+
+# After the reconstruction is completed, we need to cleanup things by
+# destroying the process group.
+cdtools.tools.multigpu.cleanup()
+
+model.tidy_probes()
+model.save_to_h5('example_reconstructions/gold_balls.h5', dataset)
+model.inspect(dataset)
+model.compare(dataset)
+plt.show()