@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 98060b001d

Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 15, 12:15 UTC

[Feature][Frontend]: Deprecate --enable-reasoning (#17452)

Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>

Chauncey committed May 1, 2025 at 21:46 UTC 98060b001dfae385c73d2380ad6a38456cbf42c9

16 files changed +49 -91

docs/source/features/reasoning_outputs.md

+5 -8

@@ -21,11 +21,10 @@ vLLM currently supports the following reasoning models:

22 ## Quickstart

24 -To use reasoning models, you need to specify the `--enable-reasoning` and `--reasoning-parser` flags when making a request to the chat completion endpoint. The `--reasoning-parser` flag specifies the reasoning parser to use for extracting reasoning content from the model output.

24 +To use reasoning models, you need to specify the `--reasoning-parser` flags when making a request to the chat completion endpoint. The `--reasoning-parser` flag specifies the reasoning parser to use for extracting reasoning content from the model output.

26 ```bash

27 -vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B \

28 - --enable-reasoning --reasoning-parser deepseek_r1

27 +vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B --reasoning-parser deepseek_r1

28 ```

30 Next, make a request to the model that should return the reasoning content in the response.

@@ -140,8 +139,7 @@ Remember to check whether the `reasoning_content` exists in the response before

139 The reasoning content is also available in the structured output. The structured output engine like `xgrammar` will use the reasoning content to generate structured output. It is only supported in v0 engine now.

140

141 ```bash

143 -VLLM_USE_V1=0 vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B \

144 - --enable-reasoning --reasoning-parser deepseek_r1

142 +VLLM_USE_V1=0 vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B --reasoning-parser deepseek_r1

143 ```

144

145 Please note that the `VLLM_USE_V1` environment variable must be set to `0` to use the v0 engine.

@@ -316,9 +314,8 @@ class DeepSeekReasoner(Reasoner):

314

315 The structured output engine like `xgrammar` will use `end_token_id` to check if the reasoning content is present in the model output and skip the structured output if it is the case.

316

319 -Finally, you can enable reasoning for the model by using the `--enable-reasoning` and `--reasoning-parser` flags.

317 +Finally, you can enable reasoning for the model by using the `--reasoning-parser` flags.

318

319 ```bash

322 -vllm serve <model_tag> \

323 - --enable-reasoning --reasoning-parser example

320 +vllm serve <model_tag> --reasoning-parser example

321 ```

examples/online_serving/openai_chat_completion_structured_outputs_with_reasoning.py

+1 -1

@@ -9,7 +9,7 @@ parser:

10 ```bash

11 vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B \

12 - --enable-reasoning --reasoning-parser deepseek_r1

12 + --reasoning-parser deepseek_r1

13 ```

15 This example demonstrates how to generate chat completions from reasoning models

examples/online_serving/openai_chat_completion_tool_calls_with_reasoning.py

+1 -1

@@ -9,7 +9,7 @@ the reasoning parser and tool calling enabled.

10 ```bash

11 vllm serve Qwen/QwQ-32B \

12 - --enable-reasoning --reasoning-parser deepseek_r1 \

12 + --reasoning-parser deepseek_r1 \

13 --enable-auto-tool-choice --tool-call-parser hermes

15 ```

examples/online_serving/openai_chat_completion_with_reasoning.py

+1 -1

@@ -8,7 +8,7 @@ with the reasoning parser:

9 ```bash

10 vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B \

11 - --enable-reasoning --reasoning-parser deepseek_r1

11 + --reasoning-parser deepseek_r1

12 ```

14 This example demonstrates how to generate chat completions from reasoning models

examples/online_serving/openai_chat_completion_with_reasoning_streaming.py

+1 -1

@@ -8,7 +8,7 @@ parser:

9 ```bash

10 vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B \

11 - --enable-reasoning --reasoning-parser deepseek_r1

11 + --reasoning-parser deepseek_r1

12 ```

14 Unlike openai_chat_completion_with_reasoning.py, this example demonstrates the

tests/entrypoints/openai/test_chat_with_tool_reasoning.py

+3 -3

@@ -13,9 +13,9 @@ MODEL_NAME = "Qwen/QwQ-32B"

13 @pytest.fixture(scope="module")

14 def server(): # noqa: F811

15 args = [

16 - "--max-model-len", "8192", "--enforce-eager", "--enable-reasoning",

17 - "--reasoning-parser", "deepseek_r1", "--enable-auto-tool-choice",

18 - "--tool-call-parser", "hermes"

16 + "--max-model-len", "8192", "--enforce-eager", "--reasoning-parser",

17 + "deepseek_r1", "--enable-auto-tool-choice", "--tool-call-parser",

18 + "hermes"

19 ]

21 with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:

tests/entrypoints/openai/test_cli_args.py

+3 -11

@@ -122,31 +122,23 @@ def test_enable_auto_choice_fails_with_enable_reasoning(serve_parser):

122 """Ensure validation fails if reasoning is enabled with auto tool choice"""

123 args = serve_parser.parse_args(args=[

124 "--enable-auto-tool-choice",

125 - "--enable-reasoning",

125 + "--reasoning-parser",

126 + "deepseek_r1",

127 ])

128 with pytest.raises(TypeError):

129 validate_parsed_serve_args(args)

130

131

131 -def test_enable_reasoning_passes_with_reasoning_parser(serve_parser):

132 +def test_passes_with_reasoning_parser(serve_parser):

133 """Ensure validation passes if reasoning is enabled

134 with a reasoning parser"""

135 args = serve_parser.parse_args(args=[

135 - "--enable-reasoning",

136 "--reasoning-parser",

137 "deepseek_r1",

138 ])

139 validate_parsed_serve_args(args)

140

141

142 -def test_enable_reasoning_fails_without_reasoning_parser(serve_parser):

143 - """Ensure validation fails if reasoning is enabled

144 - without a reasoning parser"""

145 - args = serve_parser.parse_args(args=["--enable-reasoning"])

146 - with pytest.raises(TypeError):

147 - validate_parsed_serve_args(args)

148 -

149 -

142 def test_chat_template_validation_for_happy_paths(serve_parser):

143 """Ensure validation passes if the chat template exists"""

144 args = serve_parser.parse_args(

vllm/config.py

+2 -3

@@ -3225,10 +3225,9 @@ class DecodingConfig:

3225 in the JSON schema. This is only supported for the `guidance` backend and

3226 is used to better align its behaviour with `outlines` and `xgrammar`."""

3227

3228 - reasoning_backend: Optional[str] = None

3228 + reasoning_backend: str = ""

3229 """Select the reasoning parser depending on the model that you're using.

3230 - This is used to parse the reasoning content into OpenAI API format.

3231 - Required for `--enable-reasoning`."""

3230 + This is used to parse the reasoning content into OpenAI API format."""

3231

3232 def compute_hash(self) -> str:

3233 """

vllm/engine/arg_utils.py

+12 -5

@@ -365,8 +365,9 @@ class EngineArgs:

365 calculate_kv_scales: bool = CacheConfig.calculate_kv_scales

366

367 additional_config: Optional[Dict[str, Any]] = None

368 - enable_reasoning: Optional[bool] = None

369 - reasoning_parser: Optional[str] = DecodingConfig.reasoning_backend

368 + enable_reasoning: Optional[bool] = None # DEPRECATED

369 + reasoning_parser: str = DecodingConfig.reasoning_backend

370 +

371 use_tqdm_on_load: bool = LoadConfig.use_tqdm_on_load

372

373 def __post_init__(self):

@@ -798,8 +799,15 @@ class EngineArgs:

799 "--enable-reasoning",

800 action="store_true",

801 default=False,

801 - help="Whether to enable reasoning_content for the model. "

802 - "If enabled, the model will be able to generate reasoning content."

802 + help=

803 + "[DEPRECATED] " \

804 + "The --enable-reasoning flag is deprecated as of v0.8.6. "

805 + "Use --reasoning-parser to specify " \

806 + "the reasoning parser backend instead. "

807 + "This flag (--enable-reasoning) will be " \

808 + "removed in v0.10.0. "

809 + "When --reasoning-parser is specified, " \

810 + "reasoning mode is automatically enabled."

811 )

812

813 return parser

@@ -1088,7 +1096,6 @@ class EngineArgs:

1096 disable_additional_properties=\

1097 self.guided_decoding_disable_additional_properties,

1098 reasoning_backend=self.reasoning_parser

1091 - if self.enable_reasoning else None,

1099 )

1100

1101 observability_config = ObservabilityConfig(

vllm/engine/llm_engine.py

+1 -1

@@ -2096,7 +2096,7 @@ class LLMEngine:

2096 guided_decoding.backend = guided_decoding.backend or \

2097 self.decoding_config.backend

2098

2099 - if self.decoding_config.reasoning_backend is not None:

2099 + if self.decoding_config.reasoning_backend:

2100 logger.debug("Building with reasoning backend %s",

2101 self.decoding_config.reasoning_backend)

2102

vllm/entrypoints/openai/api_server.py

+1 -2

@@ -967,7 +967,6 @@ async def init_app_state(

967 return_tokens_as_token_ids=args.return_tokens_as_token_ids,

968 enable_auto_tools=args.enable_auto_tool_choice,

969 tool_parser=args.tool_call_parser,

970 - enable_reasoning=args.enable_reasoning,

970 reasoning_parser=args.reasoning_parser,

971 enable_prompt_tokens_details=args.enable_prompt_tokens_details,

972 ) if model_config.runner_type == "generate" else None

@@ -1053,7 +1052,7 @@ async def run_server(args, **uvicorn_kwargs) -> None:

1052 f"(chose from {{ {','.join(valid_tool_parses)} }})")

1053

1054 valid_reasoning_parses = ReasoningParserManager.reasoning_parsers.keys()

1056 - if args.enable_reasoning \

1055 + if args.reasoning_parser \

1056 and args.reasoning_parser not in valid_reasoning_parses:

1057 raise KeyError(

1058 f"invalid reasoning parser: {args.reasoning_parser} "

vllm/entrypoints/openai/cli_args.py

-5

@@ -284,11 +284,6 @@ def validate_parsed_serve_args(args: argparse.Namespace):

284 raise TypeError("Error: --enable-auto-tool-choice requires "

285 "--tool-call-parser")

286

287 - # Enable reasoning needs a reasoning parser to be valid

288 - if args.enable_reasoning and not args.reasoning_parser:

289 - raise TypeError("Error: --enable-reasoning requires "

290 - "--reasoning-parser")

291 -

287

288 def create_parser_for_docs() -> FlexibleArgumentParser:

289 parser_for_docs = FlexibleArgumentParser(

vllm/entrypoints/openai/serving_chat.py

+14 -45

@@ -58,8 +58,7 @@ class OpenAIServingChat(OpenAIServing):

58 chat_template: Optional[str],

59 chat_template_content_format: ChatTemplateContentFormatOption,

60 return_tokens_as_token_ids: bool = False,

61 - enable_reasoning: bool = False,

62 - reasoning_parser: Optional[str] = None,

61 + reasoning_parser: str = "",

62 enable_auto_tools: bool = False,

63 tool_parser: Optional[str] = None,

64 enable_prompt_tokens_details: bool = False,

@@ -82,18 +81,17 @@ class OpenAIServingChat(OpenAIServing):

81 " the parallel_tool_calls client option is preset for "

82 "compatibility reasons, it will be ignored.")

85 - self.enable_reasoning: bool = enable_reasoning

84 self.reasoning_parser: Optional[Callable[[AnyTokenizer],

85 ReasoningParser]] = None

88 - if self.enable_reasoning:

86 + if reasoning_parser:

87 try:

88 self.reasoning_parser = (

89 ReasoningParserManager.get_reasoning_parser(

90 reasoning_parser))

91 + assert self.reasoning_parser is not None

92 except Exception as e:

94 - raise TypeError("Error: --enable-reasoning requires "

95 - f"reasoning_parser:'{reasoning_parser}' "

96 - "which has not been registered") from e

93 + raise TypeError(

94 + f"{reasoning_parser=} has not been registered") from e

95 self.tool_parser: Optional[Callable[[AnyTokenizer], ToolParser]] = None

96 if self.enable_auto_tools:

97 try:

@@ -423,15 +421,12 @@ class OpenAIServingChat(OpenAIServing):

421 not tool_choice_function_name

422 and self._should_stream_with_auto_tool_parsing(request))

423

426 - should_stream_with_reasoning_parsing = (

427 - self._should_stream_with_reasoning_parsing(request))

428 -

424 all_previous_token_ids: Optional[list[list[int]]]

425 function_name_returned: Optional[list[bool]] = None

426

427 # Only one of these will be used, thus previous_texts and

428 # all_previous_token_ids will not be used twice in the same iteration.

434 - if tool_choice_auto or should_stream_with_reasoning_parsing:

429 + if tool_choice_auto or self.reasoning_parser:

430 # These are only required in "auto" tool choice case

431 previous_texts = [""] * num_choices

432 all_previous_token_ids = [[]] * num_choices

@@ -446,12 +441,7 @@ class OpenAIServingChat(OpenAIServing):

441 previous_texts, all_previous_token_ids = None, None

442

443 try:

449 - # There is no need to check if the reasoning_parser is None

450 - # because the should_stream_with_reasoning_parsing check

451 - # already ensures that the reasoning_parser is not None.

452 - # but the pre-commit hook requires it.

453 - if should_stream_with_reasoning_parsing and \

454 - self.reasoning_parser is not None:

444 + if self.reasoning_parser:

445 reasoning_parser = self.reasoning_parser(tokenizer)

446 except RuntimeError as e:

447 logger.exception("Error in reasoning parser creation.")

@@ -459,7 +449,6 @@ class OpenAIServingChat(OpenAIServing):

449 yield f"data: {data}\n\n"

450 yield "data: [DONE]\n\n"

451 return

462 -

452 # Prepare the tool parser if it's needed

453 try:

454 if tool_choice_auto and self.tool_parser:

@@ -592,7 +581,7 @@ class OpenAIServingChat(OpenAIServing):

581 delta_message: Optional[DeltaMessage]

582

583 # just update previous_texts and previous_token_ids

595 - if tool_choice_auto or should_stream_with_reasoning_parsing:

584 + if tool_choice_auto or self.reasoning_parser:

585 assert previous_texts is not None

586 assert all_previous_token_ids is not None

587 previous_text = previous_texts[i]

@@ -603,7 +592,7 @@ class OpenAIServingChat(OpenAIServing):

592

593 # handle streaming deltas for tools with named tool_choice

594 if tool_choice_function_name:

606 - if (self.enable_reasoning

595 + if (self.reasoning_parser

596 and not reasoning_parser.is_reasoning_end(

597 previous_token_ids)):

598 assert reasoning_parser is not None

@@ -630,7 +619,7 @@ class OpenAIServingChat(OpenAIServing):

619 current_text = ""

620 else:

621 # Just to add remaining `content`

633 - if self.enable_reasoning:

622 + if self.reasoning_parser:

623 delta_text = previous_text + delta_text

624 current_text = ""

625

@@ -660,7 +649,7 @@ class OpenAIServingChat(OpenAIServing):

649

650 # handle streaming deltas for tools with "auto" tool choice

651 # and reasoning parser

663 - elif tool_choice_auto and self.enable_reasoning:

652 + elif tool_choice_auto and self.reasoning_parser:

653 assert tool_parser is not None

654 assert reasoning_parser is not None

655 assert added_content_delta_arr is not None

@@ -728,8 +717,7 @@ class OpenAIServingChat(OpenAIServing):

717 delta_token_ids=output.token_ids,

718 request=request))

719 # when only reasoning

731 - elif self.enable_reasoning:

732 - assert reasoning_parser is not None

720 + elif self.reasoning_parser:

721 delta_message = (reasoning_parser.

722 extract_reasoning_content_streaming(

723 previous_text,

@@ -744,7 +732,7 @@ class OpenAIServingChat(OpenAIServing):

732 delta_message = DeltaMessage(content=delta_text)

733

734 # update the previous values for the next iteration

747 - if tool_choice_auto or should_stream_with_reasoning_parsing:

735 + if tool_choice_auto or self.reasoning_parser:

736 assert previous_texts is not None

737 assert all_previous_token_ids is not None

738 previous_texts[i] = current_text

@@ -931,17 +919,9 @@ class OpenAIServingChat(OpenAIServing):

919 )

920 else:

921 logprobs = None

934 -

935 - should_stream_with_reasoning_parsing = (

936 - self._should_stream_with_reasoning_parsing(request))

937 -

938 - # In the OpenAI API the finish_reason is "tools_called"

939 - # if the tool choice is auto and the model produced a tool

940 - # call. The same is not true for named function calls

922 auto_tools_called = False

923

943 - if should_stream_with_reasoning_parsing and \

944 - self.reasoning_parser is not None:

924 + if self.reasoning_parser:

925 try:

926 reasoning_parser = self.reasoning_parser(tokenizer)

927 except RuntimeError as e:

@@ -1176,17 +1156,6 @@ class OpenAIServingChat(OpenAIServing):

1156 return (request.tools and self.tool_parser and self.enable_auto_tools

1157 and request.tool_choice in ['auto', None])

1158

1179 - def _should_stream_with_reasoning_parsing(self,

1180 - request: ChatCompletionRequest):

1181 - """

1182 - Utility function to check if streamed tokens should go through the

1183 - reasoning parser that was configured.

1184 -

1185 - We only want to do this IF reasoning is enabled and a reasoning

1186 - parser is configured.

1187 - """

1188 - return self.enable_reasoning and self.reasoning_parser is not None

1189 -

1159 def _should_check_for_unstreamed_tool_arg_tokens(

1160 self,

1161 delta_message: Optional[DeltaMessage],

vllm/model_executor/guided_decoding/__init__.py

+2 -2

@@ -103,7 +103,7 @@ async def get_guided_decoding_logits_processor(

103 reasoning_backend: str | None = None) -> LogitsProcessor | None:

104

105 reasoner = None

106 - if reasoning_backend is not None:

106 + if reasoning_backend:

107 reasoner_class = ReasoningParserManager.get_reasoning_parser(

108 reasoning_backend)

109 reasoner = reasoner_class(tokenizer)

@@ -146,7 +146,7 @@ def get_local_guided_decoding_logits_processor(

146 guided_params = maybe_backend_fallback(guided_params)

147

148 reasoner = None

149 - if reasoning_backend is not None:

149 + if reasoning_backend:

150 reasoner_class = ReasoningParserManager.get_reasoning_parser(

151 reasoning_backend)

152 reasoner = reasoner_class(tokenizer)

vllm/model_executor/guided_decoding/outlines_logits_processors.py

+1 -1

@@ -61,7 +61,7 @@ class BaseLogitsProcessor:

61 """Use the FSM to bias the logits before sampling the next token."""

63 # Skip the structured logits processing if reasoning is not finished.

64 - # reasoner is not None only when `--enable-reasoning` is set.

64 + # reasoner is not None only when `--reasoning-parser` is set.

65 if self._reasoner is not None:

66 if not self._reasoner.is_reasoning_end(input_ids):

67 return scores

vllm/model_executor/guided_decoding/xgrammar_decoding.py

+1 -1

@@ -346,7 +346,7 @@ class XGrammarLogitsProcessor:

346 scores: torch.Tensor) -> torch.Tensor:

347

348 # Skip the structured logits processing if reasoning is not finished.

349 - # reasoner is not None only when `--enable-reasoning` is set.

349 + # reasoner is not None only when `--reasoning-parser` is set.

350 if self.reasoner is not None and \

351 not self.reasoner.is_reasoning_end(

352 input_ids):