From c599b5971faf7ed8bff4bd1b14713293c20082d6 Mon Sep 17 00:00:00 2001 From: Chester Date: Tue, 13 Oct 2026 01:05:55 +0800 Subject: [PATCH] [Bugfix][Frontend] Return all logprobs when completions requests logprobs=-1 (#55717) Signed-off-by: Chester <42577861+superdiaodiao@users.noreply.github.com> --- .../completion/test_completion_error.py | 40 +++++++++++++++++++ vllm/entrypoints/openai/completion/serving.py | 4 +- 2 files changed, 43 insertions(+), 1 deletion(-) diff --git a/tests/entrypoints/openai/completion/test_completion_error.py b/tests/entrypoints/openai/completion/test_completion_error.py index bbbb61f9c9..17934026ef 100644 --- a/tests/entrypoints/openai/completion/test_completion_error.py +++ b/tests/entrypoints/openai/completion/test_completion_error.py @@ -16,6 +16,7 @@ from vllm.entrypoints.openai.models.serving import OpenAIServingModels from vllm.entrypoints.scale_out.render.serving import ServingRender from vllm.entrypoints.serve.engine.protocol import ErrorResponse from vllm.exceptions import GenerationError, VLLMValidationError +from vllm.logprobs import Logprob from vllm.outputs import CompletionOutput, RequestOutput from vllm.renderers.deepseek_v4 import DeepseekV4Renderer from vllm.renderers.hf import HfRenderer @@ -841,6 +842,45 @@ def test_logprobs_minus_one_allowed(): assert request.logprobs == -1 +def _minimal_serving_completion() -> OpenAIServingCompletion: + """A serving object with only what `_create_completion_logprobs` reads.""" + serving = OpenAIServingCompletion.__new__(OpenAIServingCompletion) + serving.return_tokens_as_token_ids = False + return serving + + +@pytest.mark.parametrize( + ("num_output_top_logprobs", "expected"), + [(-1, 3), (1, 2), (2, 3)], +) +def test_completion_logprobs_honors_requested_count( + num_output_top_logprobs: int, expected: int +): + """`logprobs=-1` asks for every logprob, so none may be filtered out. + + The validator accepts -1 and the sampler expands it to the vocabulary, so + dropping the candidates here would return an empty `top_logprobs` for every + position while still charging for the full computation. + """ + tokenizer = MagicMock() + tokenizer.decode = lambda token_ids: f"tok{token_ids[0]}" + step_logprobs = { + 10: Logprob(logprob=-0.1, decoded_token="a"), + 11: Logprob(logprob=-0.2, decoded_token="b"), + 12: Logprob(logprob=-0.3, decoded_token="c"), + } + + logprobs = _minimal_serving_completion()._create_completion_logprobs( + token_ids=[10], + top_logprobs=[step_logprobs], + num_output_top_logprobs=num_output_top_logprobs, + tokenizer=tokenizer, + ) + + assert logprobs.top_logprobs is not None + assert len(logprobs.top_logprobs[0]) == expected + + def test_logprobs_below_minus_one_rejected(): """Values more negative than -1 stay invalid.""" with pytest.raises(Exception, match="must be a positive value or -1"): diff --git a/vllm/entrypoints/openai/completion/serving.py b/vllm/entrypoints/openai/completion/serving.py index 7f7bdac075..e5ba20c52f 100644 --- a/vllm/entrypoints/openai/completion/serving.py +++ b/vllm/entrypoints/openai/completion/serving.py @@ -722,7 +722,9 @@ class OpenAIServingCompletion(GenerateBaseServing): return_as_token_id=should_return_as_token_id, ): max(top_lp[1].logprob, -9999.0) for i, top_lp in enumerate(step_top_logprobs.items()) - if logprob_token_ids or num_output_top_logprobs >= i + if logprob_token_ids + or num_output_top_logprobs == -1 + or num_output_top_logprobs >= i } )