[#17158][fix] Reject NaN top_p, min_p and temperature in SamplingParams

tensorrt_llm/sampling_params.py

@@ -372,13 +372,18 @@ def _validate(self):
         For instance, while the greedy decoding with n > 1 is capable in the
         Executor class of C++ runtime, the LLM API disallows such combination.
         """
-        if self.top_p is not None and (self.top_p < 0 or self.top_p > 1):
+        # These bounds are written as negated range checks rather than as
+        # `value < low or value > high`, so that NaN is rejected too: every
+        # comparison against NaN is False, which lets it slip through the
+        # positive form. The top_p_decay / top_p_min checks below already use
+        # this form.
+        if self.top_p is not None and not 0 <= self.top_p <= 1:
             raise ValueError(f"require 0 <= top_p <= 1, got top_p={self.top_p}")
         if self.top_k is not None and self.top_k < 0:
             raise ValueError(f"require top_k >= 0, got top_k={self.top_k}")
-        if self.min_p is not None and (self.min_p < 0 or self.min_p > 1):
+        if self.min_p is not None and not 0 <= self.min_p <= 1:
             raise ValueError(f"require 0 <= min_p <= 1, got min_p={self.min_p}")
-        if self.temperature is not None and self.temperature < 0:
+        if self.temperature is not None and not self.temperature >= 0:
             raise ValueError(f"require temperature >= 0, got temperature={self.temperature}")
 
         # Top-p decay param ranges mirror the hard checks in the

tests/integration/test_lists/test-db/l0_cpu.yml

@@ -41,6 +41,7 @@ l0_cpu:
   - unittest/llmapi/test_mpi_session.py ISOLATION
   - unittest/llmapi/test_reasoning_parser.py
   - unittest/llmapi/test_request_priority.py
+  - unittest/llmapi/test_sampling_params.py
   - unittest/llmapi/test_serialization.py
   - unittest/llmapi/test_utils.py
   - unittest/others/test_http_utils_fail_fast.py

tests/unittest/llmapi/test_sampling_params.py

@@ -117,6 +117,44 @@ def test_completion_logprobs_assignment_revalidates():
         request.to_sampling_params(backend="pytorch")
 
 
+@pytest.mark.parametrize("field", ["top_p", "min_p", "temperature"])
+def test_sampling_params_rejects_nan(field):
+    with pytest.raises(ValueError, match=field):
+        SamplingParams(**{field: float("nan")})
+
+
+@pytest.mark.parametrize(
+    ("field", "value"),
+    [
+        ("top_p", -0.1),
+        ("top_p", 1.1),
+        ("min_p", -0.1),
+        ("min_p", 1.1),
+        ("temperature", -1.0),
+    ],
+)
+def test_sampling_params_rejects_out_of_range(field, value):
+    with pytest.raises(ValueError, match=field):
+        SamplingParams(**{field: value})
+
+
+@pytest.mark.parametrize(
+    ("field", "value"),
+    [
+        ("top_p", 0.0),
+        ("top_p", 0.9),
+        ("top_p", 1.0),
+        ("min_p", 0.0),
+        ("min_p", 0.5),
+        ("min_p", 1.0),
+        ("temperature", 0.0),
+        ("temperature", 1.0),
+    ],
+)
+def test_sampling_params_accepts_in_range_values(field, value):
+    assert getattr(SamplingParams(**{field: value}), field) == value
+
+
 @pytest.mark.parametrize("value", [None, -1])
 def test_thinking_token_budget_unlimited_values(value):
     assert SamplingParams(thinking_token_budget=value).thinking_token_budget is None