diff --git a/tests/entrypoints/openai/completion/test_completion_error.py b/tests/entrypoints/openai/completion/test_completion_error.py index 4056ea49d6..bbbb61f9c9 100644 --- a/tests/entrypoints/openai/completion/test_completion_error.py +++ b/tests/entrypoints/openai/completion/test_completion_error.py @@ -852,6 +852,20 @@ def test_logprobs_below_minus_one_rejected(): ) +@pytest.mark.parametrize("max_tokens", [5, 0]) +def test_beam_search_echo_logprobs_rejected(max_tokens): + """Beam search does not compute prompt logprobs, so echo + logprobs is rejected.""" + with pytest.raises(VLLMValidationError, match="not supported with beam search"): + CompletionRequest( + model=MODEL_NAME, + prompt="Test prompt", + max_tokens=max_tokens, + use_beam_search=True, + echo=True, + logprobs=1, + ) + + class TestCompletionPromptListLimit: """Regression tests for CVE: unbounded prompt list fan-out.""" diff --git a/vllm/entrypoints/openai/completion/protocol.py b/vllm/entrypoints/openai/completion/protocol.py index f26dea2e1f..e99de1e286 100644 --- a/vllm/entrypoints/openai/completion/protocol.py +++ b/vllm/entrypoints/openai/completion/protocol.py @@ -492,6 +492,17 @@ class CompletionRequest(OpenAIBaseModel): parameter="logprob_token_ids", ) + if ( + data.get("use_beam_search") + and data.get("echo") + and data.get("logprobs") is not None + ): + raise VLLMValidationError( + "`echo` with `logprobs` is not supported with beam search " + "because beam search does not compute prompt logprobs.", + parameter="echo", + ) + if ( data.get("logprob_token_ids") and data.get("echo")