Validate fp16 dynamic loss scaling parameters are positive

deepspeed/runtime/precision_config.py

@@ -130,6 +130,26 @@ def _validate_loss_scale(cls, v):
             raise ValueError("fp16.loss_scale must be >= 0 (0 enables dynamic loss scaling)")
         return v
 
+    @field_validator("loss_scale_window", "min_loss_scale", mode="before")
+    @classmethod
+    def _validate_positive_dynamic_scale_param(cls, v, info):
+        # Both parameters drive dynamic loss scaling and must be strictly positive.
+        # loss_scale_window is used as `stable_interval % scale_window` in
+        # DynamicLossScaler.update_scale, so a value of 0 raises ZeroDivisionError,
+        # and min_loss_scale is the loss-scale floor, which collapses if <= 0.
+        name = info.field_name
+        if isinstance(v, bool):
+            raise ValueError(f"fp16.{name} must be a number, not bool")
+        try:
+            number = float(v)
+        except (TypeError, ValueError):
+            raise ValueError(f"fp16.{name} must be a number")
+        if not math.isfinite(number):
+            raise ValueError(f"fp16.{name} must be a finite number (not inf/-inf/nan)")
+        if number <= 0:
+            raise ValueError(f"fp16.{name} must be > 0")
+        return v
+
     initial_scale_power: int = 16
     """
     For dynamic loss scaling, set initial loss scale to 2^{initial_scale_power}.

tests/unit/runtime/test_precision_config_dynamic_scale.py

@@ -0,0 +1,31 @@
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+import pytest
+from pydantic import ValidationError
+
+from deepspeed.runtime.precision_config import DeepSpeedFP16Config
+
+
+@pytest.mark.parametrize("field", ["loss_scale_window", "min_loss_scale"])
+@pytest.mark.parametrize("value", [0, -1, float("inf"), float("nan"), True])
+def test_fp16_dynamic_scale_rejects_invalid_values(field, value):
+    with pytest.raises(ValidationError):
+        DeepSpeedFP16Config(**{field: value})
+
+
+@pytest.mark.parametrize("field", ["loss_scale_window", "min_loss_scale"])
+@pytest.mark.parametrize("value", [1, 1000, "2"])
+def test_fp16_dynamic_scale_accepts_valid_values(field, value):
+    cfg = DeepSpeedFP16Config(**{field: value})
+    assert getattr(cfg, field) > 0
+
+
+@pytest.mark.parametrize("field", ["loss_scale_window", "min_loss_scale"])
+@pytest.mark.parametrize("value", [[], {}])
+def test_fp16_dynamic_scale_invalid_type_has_clear_error(field, value):
+    with pytest.raises(ValidationError) as excinfo:
+        DeepSpeedFP16Config(**{field: value})
+    assert "must be a number" in str(excinfo.value)