[Bugfix] Raise VLLMValidationError for non-integer logit_bias keys (#46612)

Signed-off-by: muhammadfawaz1 <135441198+muhammadfawaz1@users.noreply.github.com>
Co-authored-by: Mahad Durrani <114791389+mahadrehmann@users.noreply.github.com>

tests/entrypoints/openai/chat_completion/test_chat_logit_bias_validation.py

@@ -76,3 +76,60 @@ async def test_chat_logit_bias_invalid(client):
     assert error.status_code == 400
     assert str(invalid_token_id) in error_message
     assert str(vocab_size) in error_message
+
+
+@pytest.mark.asyncio
+async def test_chat_logit_bias_non_integer_key(client):
+    """Test that a non-integer logit_bias key is rejected with a clean,
+    informative error instead of a raw 'invalid literal for int()' message."""
+    with pytest.raises(openai.BadRequestError) as excinfo:
+        await client.chat.completions.create(
+            model=MODEL_NAME,
+            messages=[{"role": "user", "content": "Testing invalid logit bias key"}],
+            max_tokens=5,
+            logit_bias={"not_a_token_id": 50},
+        )
+
+    error = excinfo.value
+    error_message = str(error)
+
+    assert error.status_code == 400
+    assert "not_a_token_id" in error_message
+    assert "logit_bias" in error_message
+
+
+@pytest.mark.asyncio
+async def test_chat_logit_bias_non_numeric_value(client):
+    """Test that a non-numeric logit_bias value is rejected with a message
+    that names the specific offending token, not just a generic TypeError."""
+    with pytest.raises(openai.BadRequestError) as excinfo:
+        await client.chat.completions.create(
+            model=MODEL_NAME,
+            messages=[{"role": "user", "content": "Testing invalid logit bias value"}],
+            max_tokens=5,
+            logit_bias={"1": "not_a_number"},
+        )
+
+    error = excinfo.value
+    error_message = str(error)
+
+    assert error.status_code == 400
+    assert "logit_bias" in error_message
+
+
+@pytest.mark.asyncio
+async def test_chat_logit_bias_multiple_non_integer_keys(client):
+    """Test that ALL invalid logit_bias keys are reported together,
+    not just the first one encountered."""
+    with pytest.raises(openai.BadRequestError) as excinfo:
+        await client.chat.completions.create(
+            model=MODEL_NAME,
+            messages=[{"role": "user", "content": "Testing multiple bad keys"}],
+            max_tokens=5,
+            logit_bias={"bad1": 50.0, "bad2": 20.0},
+        )
+
+    error_message = str(excinfo.value)
+    assert excinfo.value.status_code == 400
+    assert "bad1" in error_message
+    assert "bad2" in error_message

vllm/sampling_params.py

@@ -385,12 +385,31 @@ class SamplingParams(
         repetition_detection: RepetitionDetectionParams | None = None,
     ) -> "SamplingParams":
         if logit_bias is not None:
-            # Convert token_id to integer
-            # Clamp the bias between -100 and 100 per OpenAI API spec
-            logit_bias = {
-                int(token): min(100.0, max(-100.0, bias))
-                for token, bias in logit_bias.items()
-            }
+            # Fast path uses a dict comprehension; on failure we iterate once
+            # to identify the exact offending entry for the error message.
+            try:
+                logit_bias = {
+                    int(token): min(100.0, max(-100.0, bias))
+                    for token, bias in logit_bias.items()
+                }
+            except (ValueError, TypeError):
+                invalid_keys = []
+                converted_logit_bias = {}
+                for token, bias in logit_bias.items():
+                    try:
+                        token_id = int(token)
+                    except (ValueError, TypeError):
+                        invalid_keys.append(token)
+                        continue
+                    converted_logit_bias[token_id] = min(100.0, max(-100.0, bias))
+                if invalid_keys:
+                    raise VLLMValidationError(
+                        f"logit_bias contains key(s) that cannot be "
+                        f"converted to integer token IDs: {invalid_keys!r}",
+                        parameter="logit_bias",
+                        value=invalid_keys,
+                    ) from None
+                logit_bias = converted_logit_bias
 
         return SamplingParams(
             n=1 if n is None else n,