From 32985bed7c88f654b11f919ead34d77e846c32e3 Mon Sep 17 00:00:00 2001 From: Benjamin Chislett Date: Wed, 5 Mar 2025 01:30:40 -0500 Subject: [PATCH] [Frontend] Allow return_tokens_as_token_ids to be passed as a request param (#14066) Signed-off-by: Benjamin Chislett --- .../openai/test_return_tokens_as_ids.py | 44 +++++++++++++------ vllm/entrypoints/openai/protocol.py | 12 +++++ vllm/entrypoints/openai/serving_chat.py | 21 +++++---- vllm/entrypoints/openai/serving_completion.py | 12 +++-- 4 files changed, 64 insertions(+), 25 deletions(-) diff --git a/tests/entrypoints/openai/test_return_tokens_as_ids.py b/tests/entrypoints/openai/test_return_tokens_as_ids.py index 9b33eddae2a8..6474858642d7 100644 --- a/tests/entrypoints/openai/test_return_tokens_as_ids.py +++ b/tests/entrypoints/openai/test_return_tokens_as_ids.py @@ -17,18 +17,28 @@ from .test_completion import MODEL_NAME @pytest.fixture(scope="module") -def server_with_return_tokens_as_token_ids_flag( - default_server_args): # noqa: F811 - args_with_flag = default_server_args + ["--return-tokens-as-token-ids"] - with RemoteOpenAIServer(MODEL_NAME, args_with_flag) as remote_server: - yield remote_server +def server_fixture(request, default_server_args): # noqa: F811 + use_server_flag = request.param + if use_server_flag: + args_with_flag = default_server_args + ["--return-tokens-as-token-ids"] + with RemoteOpenAIServer(MODEL_NAME, args_with_flag) as remote_server: + yield (remote_server, True) + else: + with RemoteOpenAIServer(MODEL_NAME, + default_server_args) as remote_server: + yield (remote_server, False) @pytest.mark.asyncio +@pytest.mark.parametrize("server_fixture", [True, False], indirect=True) async def test_completion_return_tokens_as_token_ids_completion( - server_with_return_tokens_as_token_ids_flag): - async with server_with_return_tokens_as_token_ids_flag.get_async_client( - ) as client: + server_fixture): + server, use_server_flag = server_fixture + request_args = {} + if not use_server_flag: + request_args["return_tokens_as_token_ids"] = True + + async with server.get_async_client() as client: completion = await client.completions.create( model=MODEL_NAME, @@ -39,7 +49,8 @@ async def test_completion_return_tokens_as_token_ids_completion( echo=True, temperature=0, max_tokens=10, - logprobs=1) + logprobs=1, + extra_body=request_args) text = completion.choices[0].text token_strs = completion.choices[0].logprobs.tokens @@ -60,10 +71,14 @@ async def test_completion_return_tokens_as_token_ids_completion( @pytest.mark.asyncio -async def test_chat_return_tokens_as_token_ids_completion( - server_with_return_tokens_as_token_ids_flag): - async with server_with_return_tokens_as_token_ids_flag.get_async_client( - ) as client: +@pytest.mark.parametrize("server_fixture", [True, False], indirect=True) +async def test_chat_return_tokens_as_token_ids_completion(server_fixture): + server, use_server_flag = server_fixture + request_args = {} + if not use_server_flag: + request_args["return_tokens_as_token_ids"] = True + + async with server.get_async_client() as client: response = await client.chat.completions.create( model=MODEL_NAME, # Include Unicode characters to test for dividing a single @@ -78,7 +93,8 @@ async def test_chat_return_tokens_as_token_ids_completion( }], temperature=0, max_tokens=8, - logprobs=True) + logprobs=True, + extra_body=request_args) text = response.choices[0].message.content tokenizer = get_tokenizer(tokenizer_name=MODEL_NAME) diff --git a/vllm/entrypoints/openai/protocol.py b/vllm/entrypoints/openai/protocol.py index 14ce71cd3c2e..2c740caf20fb 100644 --- a/vllm/entrypoints/openai/protocol.py +++ b/vllm/entrypoints/openai/protocol.py @@ -369,6 +369,12 @@ class ChatCompletionRequest(OpenAIBaseModel): "arguments. For example: {'qualname': " "'my_module.MyLogitsProcessor', 'args': [1, 2], 'kwargs': " "{'param': 'value'}}.")) + return_tokens_as_token_ids: Optional[bool] = Field( + default=None, + description=( + "If specified with 'logprobs', tokens are represented " + " as strings of the form 'token_id:{token_id}' so that tokens " + "that are not JSON-encodable can be identified.")) # doc: end-chat-completion-extra-params @@ -739,6 +745,12 @@ class CompletionRequest(OpenAIBaseModel): "arguments. For example: {'qualname': " "'my_module.MyLogitsProcessor', 'args': [1, 2], 'kwargs': " "{'param': 'value'}}.")) + return_tokens_as_token_ids: Optional[bool] = Field( + default=None, + description=( + "If specified with 'logprobs', tokens are represented " + " as strings of the form 'token_id:{token_id}' so that tokens " + "that are not JSON-encodable can be identified.")) # doc: end-completion-extra-params diff --git a/vllm/entrypoints/openai/serving_chat.py b/vllm/entrypoints/openai/serving_chat.py index fc4dd110712e..98b6679722d3 100644 --- a/vllm/entrypoints/openai/serving_chat.py +++ b/vllm/entrypoints/openai/serving_chat.py @@ -450,6 +450,8 @@ class OpenAIServingChat(OpenAIServing): top_logprobs=output.logprobs, tokenizer=tokenizer, num_output_top_logprobs=request.top_logprobs, + return_as_token_id=request. + return_tokens_as_token_ids, ) else: logprobs = None @@ -705,6 +707,7 @@ class OpenAIServingChat(OpenAIServing): top_logprobs=out_logprobs, num_output_top_logprobs=request.top_logprobs, tokenizer=tokenizer, + return_as_token_id=request.return_tokens_as_token_ids, ) else: logprobs = None @@ -852,13 +855,14 @@ class OpenAIServingChat(OpenAIServing): def _get_top_logprobs( self, logprobs: dict[int, Logprob], top_logprobs: Optional[int], - tokenizer: AnyTokenizer) -> list[ChatCompletionLogProb]: + tokenizer: AnyTokenizer, + should_return_as_token_id: bool) -> list[ChatCompletionLogProb]: return [ ChatCompletionLogProb(token=(token := self._get_decoded_token( p[1], p[0], tokenizer, - return_as_token_id=self.return_tokens_as_token_ids)), + return_as_token_id=should_return_as_token_id)), logprob=max(p[1].logprob, -9999.0), bytes=list( token.encode("utf-8", errors="replace"))) @@ -872,15 +876,18 @@ class OpenAIServingChat(OpenAIServing): top_logprobs: GenericSequence[Optional[dict[int, Logprob]]], tokenizer: AnyTokenizer, num_output_top_logprobs: Optional[int] = None, + return_as_token_id: Optional[bool] = None, ) -> ChatCompletionLogProbs: """Create OpenAI-style logprobs.""" logprobs_content: list[ChatCompletionLogProbsContent] = [] + should_return_as_token_id = return_as_token_id if \ + return_as_token_id is not None else self.return_tokens_as_token_ids for i, token_id in enumerate(token_ids): step_top_logprobs = top_logprobs[i] if step_top_logprobs is None: token = tokenizer.decode(token_id) - if self.return_tokens_as_token_ids: + if should_return_as_token_id: token = f"token_id:{token_id}" logprobs_content.append( @@ -898,16 +905,14 @@ class OpenAIServingChat(OpenAIServing): step_token, token_id, tokenizer, - self.return_tokens_as_token_ids, + should_return_as_token_id, ), logprob=max(step_token.logprob, -9999.0), bytes=None if step_decoded is None else list( step_decoded.encode("utf-8", errors="replace")), top_logprobs=self._get_top_logprobs( - step_top_logprobs, - num_output_top_logprobs, - tokenizer, - ), + step_top_logprobs, num_output_top_logprobs, + tokenizer, should_return_as_token_id), )) return ChatCompletionLogProbs(content=logprobs_content) diff --git a/vllm/entrypoints/openai/serving_completion.py b/vllm/entrypoints/openai/serving_completion.py index f52e1de9840e..edcf1b086bad 100644 --- a/vllm/entrypoints/openai/serving_completion.py +++ b/vllm/entrypoints/openai/serving_completion.py @@ -316,6 +316,8 @@ class OpenAIServingCompletion(OpenAIServing): num_output_top_logprobs=request.logprobs, tokenizer=tokenizer, initial_text_offset=previous_text_lens[i], + return_as_token_id=request. + return_tokens_as_token_ids, ) else: logprobs = None @@ -436,6 +438,7 @@ class OpenAIServingCompletion(OpenAIServing): top_logprobs=out_logprobs, tokenizer=tokenizer, num_output_top_logprobs=request.logprobs, + return_as_token_id=request.return_tokens_as_token_ids, ) else: logprobs = None @@ -477,6 +480,7 @@ class OpenAIServingCompletion(OpenAIServing): num_output_top_logprobs: int, tokenizer: AnyTokenizer, initial_text_offset: int = 0, + return_as_token_id: Optional[bool] = None, ) -> CompletionLogProbs: """Create logprobs for OpenAI Completion API.""" out_text_offset: list[int] = [] @@ -486,11 +490,13 @@ class OpenAIServingCompletion(OpenAIServing): last_token_len = 0 + should_return_as_token_id = return_as_token_id if \ + return_as_token_id is not None else self.return_tokens_as_token_ids for i, token_id in enumerate(token_ids): step_top_logprobs = top_logprobs[i] if step_top_logprobs is None: token = tokenizer.decode(token_id) - if self.return_tokens_as_token_ids: + if should_return_as_token_id: token = f"token_id:{token_id}" out_tokens.append(token) @@ -503,7 +509,7 @@ class OpenAIServingCompletion(OpenAIServing): step_token, token_id, tokenizer, - return_as_token_id=self.return_tokens_as_token_ids, + return_as_token_id=should_return_as_token_id, ) token_logprob = max(step_token.logprob, -9999.0) @@ -520,7 +526,7 @@ class OpenAIServingCompletion(OpenAIServing): self._get_decoded_token(top_lp[1], top_lp[0], tokenizer, - return_as_token_id=self.return_tokens_as_token_ids): + 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 num_output_top_logprobs >= i