diff --git a/deepspec/data/cuda_prefetcher.py b/deepspec/data/cuda_prefetcher.py
index 69eac905..37be8ad8 100644
--- a/deepspec/data/cuda_prefetcher.py
+++ b/deepspec/data/cuda_prefetcher.py
@@ -29,6 +29,7 @@ def __iter__(self):
         self._done = False
         self._gpu_batch = None
         self._thread = None
+        self._background_error = None
         # First batch: fetch synchronously so __next__ has something to return.
         self._fetch_and_transfer()
         return self
@@ -43,12 +44,24 @@ def _fetch_and_transfer(self):
         with torch.cuda.stream(self.stream):
             self._gpu_batch = move_batch_to_device(cpu_batch, self.device)
 
+    def _prefetch(self):
+        try:
+            self._fetch_and_transfer()
+        except BaseException as exc:
+            self._background_error = exc
+
     def __next__(self):
         # Join the background thread kicked off in the previous __next__.
         if self._thread is not None:
             self._thread.join()
             self._thread = None
 
+        if self._background_error is not None:
+            error = self._background_error
+            self._background_error = None
+            self._gpu_batch = None
+            raise error
+
         if self._done:
             raise StopIteration
 
@@ -64,7 +77,7 @@ def __next__(self):
 
         # Kick off the next fetch and H2D in a background thread so it
         # overlaps with compute on the batch we are about to return.
-        self._thread = Thread(target=self._fetch_and_transfer, daemon=True)
+        self._thread = Thread(target=self._prefetch, daemon=True)
         self._thread.start()
 
         return batch
diff --git a/tests/test_cuda_prefetcher.py b/tests/test_cuda_prefetcher.py
new file mode 100644
index 00000000..da7388d4
--- /dev/null
+++ b/tests/test_cuda_prefetcher.py
@@ -0,0 +1,116 @@
+import contextlib
+import importlib.util
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+
+class FakeTensor:
+    def __init__(self, value):
+        self.value = value
+        self.dtype = "int64"
+        self.recorded_streams = []
+
+    def to(self, *args, **kwargs):
+        return self
+
+    def record_stream(self, stream):
+        self.recorded_streams.append(stream)
+
+
+class FakeStream:
+    def wait_stream(self, stream):
+        self.waited_for = stream
+
+
+class ScriptedIterator:
+    def __init__(self, events):
+        self._events = iter(events)
+
+    def __iter__(self):
+        return self
+
+    def __next__(self):
+        event = next(self._events)
+        if isinstance(event, BaseException):
+            raise event
+        return event
+
+
+class ResettableLoader:
+    def __init__(self, runs):
+        self._runs = iter(runs)
+
+    def __iter__(self):
+        return ScriptedIterator(next(self._runs))
+
+    def __len__(self):
+        return 1
+
+
+@pytest.fixture
+def prefetcher_module(monkeypatch):
+    current_stream = FakeStream()
+    fake_cuda = types.SimpleNamespace(
+        Stream=lambda device=None: FakeStream(),
+        stream=lambda stream: contextlib.nullcontext(),
+        current_stream=lambda device=None: current_stream,
+    )
+    fake_torch = types.ModuleType("torch")
+    fake_torch.cuda = fake_cuda
+    fake_torch.long = "int64"
+    monkeypatch.setitem(sys.modules, "torch", fake_torch)
+
+    module_path = (
+        Path(__file__).parents[1] / "deepspec" / "data" / "cuda_prefetcher.py"
+    )
+    spec = importlib.util.spec_from_file_location("cuda_prefetcher_under_test", module_path)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def batch(value):
+    return {"input_ids": FakeTensor(value)}
+
+
+def test_normal_batches_are_returned_in_order(prefetcher_module):
+    prefetcher = prefetcher_module.CUDAPrefetcher(
+        ScriptedIterator([batch(1), batch(2)]),
+        "cuda:0",
+    )
+
+    assert [item["input_ids"].value for item in prefetcher] == [1, 2]
+
+
+def test_background_loader_error_is_reraised(prefetcher_module):
+    error = RuntimeError("bad batch")
+    prefetcher = prefetcher_module.CUDAPrefetcher(
+        ScriptedIterator([batch(1), error]),
+        "cuda:0",
+    )
+    iterator = iter(prefetcher)
+
+    assert next(iterator)["input_ids"].value == 1
+    with pytest.raises(RuntimeError, match="bad batch") as raised:
+        next(iterator)
+    assert raised.value is error
+
+
+def test_reiteration_resets_background_error(prefetcher_module):
+    loader = ResettableLoader(
+        [
+            [batch(1), RuntimeError("bad batch")],
+            [batch(2)],
+        ]
+    )
+    prefetcher = prefetcher_module.CUDAPrefetcher(loader, "cuda:0")
+    first = iter(prefetcher)
+
+    assert next(first)["input_ids"].value == 1
+    with pytest.raises(RuntimeError, match="bad batch"):
+        next(first)
+
+    assert [item["input_ids"].value for item in prefetcher] == [2]