diff --git a/tests/test_optimizer.py b/tests/test_optimizer.py
new file mode 100644
index 00000000..ae5c1dcd
--- /dev/null
+++ b/tests/test_optimizer.py
@@ -0,0 +1,154 @@
+# Copyright (c) 2026 LightSeek Foundation
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in
+# all copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+from unittest import mock
+
+import pytest
+import torch
+import torch.distributed as dist
+
+from torchspec.training.optimizer import BF16Optimizer
+
+
+pytestmark = pytest.mark.skipif(
+    not torch.cuda.is_available(),
+    reason="fused AdamW found_inf handling requires CUDA",
+)
+
+
+def _make_optimizer() -> BF16Optimizer:
+    model = torch.nn.Linear(
+        2,
+        1,
+        bias=False,
+        device="cuda",
+        dtype=torch.bfloat16,
+    )
+    with torch.no_grad():
+        model.weight.copy_(torch.tensor([[1.0, -2.0]], device="cuda"))
+
+    return BF16Optimizer(
+        model,
+        lr=0.1,
+        weight_decay=0.1,
+        max_grad_norm=10.0,
+        total_steps=10,
+        warmup_ratio=0.0,
+        decay_style="constant",
+    )
+
+
+def _set_grad(optimizer: BF16Optimizer, first_value: float) -> None:
+    optimizer.model_params[0].grad = torch.tensor(
+        [[first_value, 0.25]],
+        device="cuda",
+        dtype=torch.bfloat16,
+    )
+
+
+def _initialize_adam_state(optimizer: BF16Optimizer) -> None:
+    _set_grad(optimizer, 0.5)
+    grad_norm = optimizer.step()
+    assert torch.isfinite(grad_norm).item()
+
+
+def _snapshot(optimizer: BF16Optimizer):
+    model_params = [p.detach().clone() for p in optimizer.model_params]
+    master_params = [p.detach().clone() for p in optimizer.fp32_params]
+    states = []
+    for master_param in optimizer.fp32_params:
+        states.append(
+            {
+                key: value.detach().clone() if isinstance(value, torch.Tensor) else value
+                for key, value in optimizer.optimizer.state[master_param].items()
+            }
+        )
+    return model_params, master_params, states
+
+
+def _assert_snapshot_equal(optimizer: BF16Optimizer, snapshot) -> None:
+    model_params, master_params, states = snapshot
+    for actual, expected in zip(optimizer.model_params, model_params):
+        torch.testing.assert_close(actual, expected, rtol=0, atol=0)
+    for actual, expected in zip(optimizer.fp32_params, master_params):
+        torch.testing.assert_close(actual, expected, rtol=0, atol=0)
+    for master_param, expected_state in zip(optimizer.fp32_params, states):
+        actual_state = optimizer.optimizer.state[master_param]
+        assert actual_state.keys() == expected_state.keys()
+        for key, expected in expected_state.items():
+            actual = actual_state[key]
+            if isinstance(expected, torch.Tensor):
+                torch.testing.assert_close(actual, expected, rtol=0, atol=0)
+            else:
+                assert actual == expected
+
+
+def test_finite_gradient_updates_parameters() -> None:
+    optimizer = _make_optimizer()
+    master_before = optimizer.fp32_params[0].detach().clone()
+
+    _set_grad(optimizer, 0.5)
+    grad_norm = optimizer.step()
+
+    assert torch.isfinite(grad_norm).item()
+    assert optimizer.optimizer.found_inf.item() == 0.0
+    assert not torch.equal(optimizer.fp32_params[0], master_before)
+
+
+@pytest.mark.parametrize(
+    "bad_value",
+    [float("nan"), float("inf"), float("-inf")],
+    ids=["nan", "positive_inf", "negative_inf"],
+)
+def test_nonfinite_gradient_skips_entire_adamw_update(bad_value: float) -> None:
+    optimizer = _make_optimizer()
+    _initialize_adam_state(optimizer)
+    before = _snapshot(optimizer)
+
+    _set_grad(optimizer, bad_value)
+    grad_norm = optimizer.step()
+
+    assert not torch.isfinite(grad_norm).item()
+    assert optimizer.optimizer.found_inf.item() == 1.0
+    _assert_snapshot_equal(optimizer, before)
+    assert all(param.grad is None for param in optimizer.model_params)
+
+
+def test_remote_nonfinite_signal_skips_local_finite_update() -> None:
+    optimizer = _make_optimizer()
+    _initialize_adam_state(optimizer)
+    before = _snapshot(optimizer)
+    _set_grad(optimizer, 0.5)
+
+    def mark_remote_nonfinite(found_inf: torch.Tensor, op) -> None:
+        assert op == dist.ReduceOp.MAX
+        found_inf.fill_(1.0)
+
+    with (
+        mock.patch.object(dist, "is_initialized", return_value=True),
+        mock.patch.object(dist, "get_world_size", return_value=2),
+        mock.patch.object(dist, "all_reduce", side_effect=mark_remote_nonfinite) as all_reduce,
+    ):
+        grad_norm = optimizer.step()
+
+    assert torch.isfinite(grad_norm).item()
+    assert optimizer.optimizer.found_inf.item() == 1.0
+    all_reduce.assert_called_once()
+    _assert_snapshot_equal(optimizer, before)
diff --git a/torchspec/training/optimizer.py b/torchspec/training/optimizer.py
index deb9ba93..cff24ef8 100644
--- a/torchspec/training/optimizer.py
+++ b/torchspec/training/optimizer.py
@@ -19,6 +19,7 @@
 # SOFTWARE.
 
 import torch
+import torch.distributed as dist
 
 from torchspec.training.lr_scheduler import LRSchedulerWithWarmup
 from torchspec.utils.logging import print_on_rank0
@@ -51,6 +52,10 @@ def __init__(
             weight_decay=weight_decay,
             fused=True,
         )
+        if not getattr(self.optimizer, "_step_supports_amp_scaling", False):
+            raise RuntimeError(
+                "BF16Optimizer requires fused AdamW with device-side found_inf support"
+            )
         self.scheduler = LRSchedulerWithWarmup(
             self.optimizer,
             max_lr=lr,
@@ -85,6 +90,16 @@ def step(self, closure=None):
                 torch._foreach_copy_(grad_destinations, grad_sources)
 
         grad_norm = torch.nn.utils.clip_grad_norm_(self.fp32_params, self.max_grad_norm)
+
+        # Fused AdamW consumes this device scalar through the same path used by
+        # GradScaler. A nonzero value skips parameter, moment, weight-decay, and
+        # optimizer-step updates without synchronizing the CUDA scalar to Python.
+        found_inf = (~torch.isfinite(grad_norm)).to(dtype=torch.float32)
+        if dist.is_available() and dist.is_initialized() and dist.get_world_size() > 1:
+            # A sharded rank may be the only one that observes a nonfinite
+            # gradient. All training ranks must make the same update decision.
+            dist.all_reduce(found_inf, op=dist.ReduceOp.MAX)
+        self.optimizer.found_inf = found_inf
         self.optimizer.step()
 
         self.optimizer.zero_grad()