[Bugfix][Frontend] Honor default penalties in batch chat completions (#60307)

Signed-off-by: sricursion <sricursion@gmail.com>
Co-authored-by: Flora Feng <4florafeng@gmail.com>

tests/entrypoints/openai/test_penalty_default_resolution.py

@@ -2,7 +2,8 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 """Unit tests for presence_penalty / frequency_penalty resolution from
-default_sampling_params in ChatCompletionRequest and CompletionRequest.
+default_sampling_params in ChatCompletionRequest, BatchChatCompletionRequest
+and CompletionRequest.
 
 Regression test for https://github.com/vllm-project/vllm/issues/50767:
 these two penalties defaulted to 0.0 (not None) and to_sampling_params()
@@ -12,6 +13,7 @@ discarded on the /v1/chat/completions and /v1/completions endpoints.
 """
 
 from vllm.entrypoints.openai.chat_completion.protocol import (
+    BatchChatCompletionRequest,
     ChatCompletionRequest,
 )
 from vllm.entrypoints.openai.completion.protocol import (
@@ -29,6 +31,14 @@ def _chat(**kwargs):
     )
 
 
+def _batch(**kwargs):
+    conversation = [{"role": "user", "content": "hi"}]
+    request = BatchChatCompletionRequest(
+        model="test-model", messages=[conversation], **kwargs
+    )
+    return request.to_chat_completion_request(conversation)
+
+
 class TestChatCompletionPenaltyDefaults:
     def test_defaults_applied_when_client_omits(self):
         """Server-default penalties are applied when the client sends none."""
@@ -49,6 +59,22 @@ class TestChatCompletionPenaltyDefaults:
         assert sp.frequency_penalty == 0.0
 
 
+class TestBatchChatCompletionPenaltyDefaults:
+    def test_defaults_applied_when_client_omits(self):
+        """Server-default penalties reach every conversation of a batch."""
+        sp = _batch().to_sampling_params(100, _DEFAULTS)
+        assert sp.presence_penalty == 1.5
+        assert sp.frequency_penalty == 0.5
+
+    def test_client_value_overrides_default(self):
+        """An explicit client penalty, even 0.0, wins over the server default."""
+        sp = _batch(presence_penalty=0.0, frequency_penalty=0.2).to_sampling_params(
+            100, _DEFAULTS
+        )
+        assert sp.presence_penalty == 0.0
+        assert sp.frequency_penalty == 0.2
+
+
 class TestCompletionPenaltyDefaults:
     def test_defaults_applied_when_client_omits(self):
         sp = CompletionRequest(model="test-model", prompt="hi").to_sampling_params(

vllm/entrypoints/openai/chat_completion/protocol.py

@@ -1093,7 +1093,7 @@ class BatchChatCompletionRequest(OpenAIBaseModel):
     model: str | None = None
 
     # Shared sampling / generation fields — mirror ChatCompletionRequest.
-    frequency_penalty: float | None = 0.0
+    frequency_penalty: float | None = None
     logit_bias: dict[str, float] | None = None
     logprobs: bool | None = False
     top_logprobs: TopLogprobsParam = 0
@@ -1108,7 +1108,7 @@ class BatchChatCompletionRequest(OpenAIBaseModel):
     max_tokens: int | None = None
     max_completion_tokens: int | None = None
     n: int | None = 1
-    presence_penalty: float | None = 0.0
+    presence_penalty: float | None = None
     response_format: Any | None = None
     seed: int | None = Field(None, ge=_INT64_MIN, le=_INT64_MAX)
     stop: StopParam = Field(default_factory=list)