Reject repetition_penalty=0 in SamplingParams.verify() (#24874)

Co-authored-by: RulinJuice <265952454+RulinJuice@users.noreply.github.com>

docs/basic_usage/sampling_params.md

@@ -62,7 +62,7 @@ python -m sglang.launch_server --model-path <MODEL> --sampling-defaults openai
 |--------------------|------------------------|------------------------------------------------------------------------------------------------------------------------------------------------|
 | frequency_penalty  | `float = 0.0`          | Penalizes tokens based on their frequency in generation so far. Must be between `-2` and `2` where negative numbers encourage repeatment of tokens and positive number encourages sampling of new tokens. The scaling of penalization grows linearly with each appearance of a token. |
 | presence_penalty   | `float = 0.0`          | Penalizes tokens if they appeared in the generation so far. Must be between `-2` and `2` where negative numbers encourage repeatment of tokens and positive number encourages sampling of new tokens. The scaling of the penalization is constant if a token occurred. |
-| repetition_penalty | `float = 1.0`          | Scales the logits of previously generated tokens to discourage (values > 1) or encourage (values < 1) repetition. Valid range is `[0, 2]`; `1.0` leaves probabilities unchanged. |
+| repetition_penalty | `float = 1.0`          | Scales the logits of previously generated tokens to discourage (values > 1) or encourage (values < 1) repetition. Valid range is `(0, 2]`; `1.0` leaves probabilities unchanged. |
 | min_new_tokens     | `int = 0`              | Forces the model to generate at least `min_new_tokens` until a stop word or EOS token is sampled. Note that this might lead to unintended behavior, for example, if the distribution is highly skewed towards these tokens. |
 
 ### Constrained decoding

python/sglang/srt/sampling/sampling_params.py

@@ -139,10 +139,10 @@ class SamplingParams:
             raise ValueError(
                 "presence_penalty must be in [-2, 2], got " f"{self.presence_penalty}."
             )
-        if not 0.0 <= self.repetition_penalty <= 2.0:
+        if not 0.0 < self.repetition_penalty <= 2.0:
             raise ValueError(
-                "repetition_penalty must be in [0, 2], got "
-                f"{self.repetition_penalty}."
+                "repetition_penalty must be in (0, 2] (1.0 = no penalty), "
+                f"got {self.repetition_penalty}."
             )
         if not 0 <= self.min_new_tokens:
             raise ValueError(

test/registered/unit/sampling/test_sampling_params.py

@@ -178,22 +178,36 @@ class TestSamplingParamsVerify(CustomTestCase):
 
     # --- repetition_penalty ---
     def test_repetition_penalty_negative_raises(self):
-        """Test that verify() rejects negative repetition_penalty (valid range is [0, 2])."""
+        """Test that verify() rejects negative repetition_penalty (valid range is (0, 2])."""
         sp = self._make(repetition_penalty=-0.1)
         with self.assertRaises(ValueError):
             sp.verify(self.VOCAB_SIZE)
 
+    def test_repetition_penalty_zero_raises(self):
+        """Test that verify() rejects repetition_penalty=0.
+
+        A value of 0 makes the sampling kernel divide logits by 0, producing
+        inf/NaN in the probability tensor and crashing every TP rank with a
+        device-side assert.
+        """
+        sp = self._make(repetition_penalty=0.0)
+        with self.assertRaises(ValueError):
+            sp.verify(self.VOCAB_SIZE)
+
     def test_repetition_penalty_above_two_raises(self):
         """Test that verify() rejects repetition_penalty > 2.0."""
         sp = self._make(repetition_penalty=2.1)
         with self.assertRaises(ValueError):
             sp.verify(self.VOCAB_SIZE)
 
-    def test_repetition_penalty_boundaries_valid(self):
-        """Test that boundary values 0.0 and 2.0 are both accepted."""
-        self._make(repetition_penalty=0.0).verify(self.VOCAB_SIZE)
+    def test_repetition_penalty_boundary_two_valid(self):
+        """Test that the upper boundary value 2.0 is accepted."""
         self._make(repetition_penalty=2.0).verify(self.VOCAB_SIZE)
 
+    def test_repetition_penalty_small_positive_valid(self):
+        """Test that a small positive repetition_penalty (e.g. 1e-3) is accepted."""
+        self._make(repetition_penalty=1e-3).verify(self.VOCAB_SIZE)
+
     # --- min_new_tokens / max_new_tokens ---
     def test_negative_min_new_tokens_raises(self):
         """Test that verify() rejects negative min_new_tokens."""