[Refactor][Frontend] Keep all logic about reasoning into one class (vllm-project#14428)

Signed-off-by: Ce Gao <cegao@tensorchord.ai>

Author: gaocegege

tests/entrypoints/openai/reasoning_parsers/__init__.py renamed to tests/reasoning/__init__.py

@@ -3,74 +3,92 @@
 import pytest
 from transformers import AutoTokenizer
 
-from tests.entrypoints.openai.reasoning_parsers.utils import (
-    run_reasoning_extraction)
-from vllm.entrypoints.openai.reasoning_parsers import (ReasoningParser,
-                                                       ReasoningParserManager)
+from tests.reasoning.utils import run_reasoning_extraction
+from vllm.reasoning import ReasoningParser, ReasoningParserManager
 
 parser_name = "deepseek_r1"
 start_token = "<think>"
 end_token = "</think>"
 
+REASONING_MODEL_NAME = "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"
+
+
+@pytest.fixture(scope="module")
+def deepseek_r1_qwen_tokenizer():
+    return AutoTokenizer.from_pretrained(REASONING_MODEL_NAME)
+
+
 SIMPLE_REASONING = {
     "output": "This is a reasoning section</think>This is the rest",
     "reasoning_content": "This is a reasoning section",
     "content": "This is the rest",
+    "is_reasoning_end": True,
 }
 COMPLETE_REASONING = {
     "output": "This is a reasoning section</think>",
     "reasoning_content": "This is a reasoning section",
     "content": None,
+    "is_reasoning_end": True,
 }
 NO_CONTENT = {
     "output": "This is content",
     "reasoning_content": "This is content",
     "content": None,
+    "is_reasoning_end": False,
 }
 NO_REASONING_STREAMING = {
     "output": "This is a reasoning section",
     "reasoning_content": "This is a reasoning section",
     "content": None,
+    "is_reasoning_end": False,
 }
 MULTIPLE_LINES = {
     "output": "This\nThat</think>This is the rest\nThat",
     "reasoning_content": "This\nThat",
     "content": "This is the rest\nThat",
+    "is_reasoning_end": True,
 }
 SHORTEST_REASONING_NO_STREAMING = {
     "output": "</think>This is the rest",
     "reasoning_content": "",
     "content": "This is the rest",
+    "is_reasoning_end": True,
 }
 SHORTEST_REASONING = {
     "output": "</think>This is the rest",
     "reasoning_content": None,
     "content": "This is the rest",
+    "is_reasoning_end": True,
 }
 REASONING_WITH_THINK = {
     "output": "<think>This is a reasoning section</think>This is the rest",
     "reasoning_content": "This is a reasoning section",
     "content": "This is the rest",
+    "is_reasoning_end": True,
 }
 COMPLETE_REASONING_WITH_THINK = {
     "output": "<think>This is a reasoning section</think>",
     "reasoning_content": "This is a reasoning section",
     "content": None,
+    "is_reasoning_end": True,
 }
 MULTIPLE_LINES_WITH_THINK = {
     "output": "<think>This\nThat</think>This is the rest\nThat",
     "reasoning_content": "This\nThat",
     "content": "This is the rest\nThat",
+    "is_reasoning_end": True,
 }
 SHORTEST_REASONING_NO_STREAMING_WITH_THINK = {
     "output": "</think>This is the rest",
     "reasoning_content": "",
     "content": "This is the rest",
+    "is_reasoning_end": True,
 }
 SHORTEST_REASONING_WITH_THINK = {
     "output": "</think>This is the rest",
     "reasoning_content": None,
     "content": "This is the rest",
+    "is_reasoning_end": True,
 }
 
 TEST_CASES = [
@@ -166,27 +184,39 @@
     ),
 ]
 
-# Global tokenizer initialization to avoid repeated loading
-tokenizer = AutoTokenizer.from_pretrained("facebook/opt-125m")
-tokenizer.add_tokens([start_token, end_token])
-
 
 @pytest.mark.parametrize("streaming, param_dict", TEST_CASES)
 def test_reasoning(
     streaming: bool,
     param_dict: dict,
+    deepseek_r1_qwen_tokenizer,
 ):
-    output = tokenizer.tokenize(param_dict["output"])
+    output = deepseek_r1_qwen_tokenizer.tokenize(param_dict["output"])
     # decode everything to tokens
     output_tokens: list[str] = [
-        tokenizer.convert_tokens_to_string([token]) for token in output
+        deepseek_r1_qwen_tokenizer.convert_tokens_to_string([token])
+        for token in output
     ]
     parser: ReasoningParser = ReasoningParserManager.get_reasoning_parser(
-        parser_name)(tokenizer)
+        parser_name)(deepseek_r1_qwen_tokenizer)
 
     reasoning, content = run_reasoning_extraction(parser,
                                                   output_tokens,
                                                   streaming=streaming)
 
     assert reasoning == param_dict["reasoning_content"]
     assert content == param_dict["content"]
+
+    # Test is_reasoning_end
+    output_ids = deepseek_r1_qwen_tokenizer.convert_tokens_to_ids(output)
+    is_reasoning_end = parser.is_reasoning_end(output_ids)
+    assert is_reasoning_end == param_dict["is_reasoning_end"]
+
+    # Test extract_content
+    if param_dict["content"] is not None:
+        content = parser.extract_content_ids(output_ids)
+        assert content == deepseek_r1_qwen_tokenizer.convert_tokens_to_ids(
+            deepseek_r1_qwen_tokenizer.tokenize(param_dict["content"]))
+    else:
+        content = parser.extract_content_ids(output)
+        assert content == []

tests/entrypoints/openai/reasoning_parsers/test_deepseekr1_reasoning_parser.py renamed to tests/reasoning/test_deepseekr1_reasoning_parser.py

@@ -2,10 +2,8 @@
 import pytest
 from transformers import AutoTokenizer
 
-from tests.entrypoints.openai.reasoning_parsers.utils import (
-    DeltaMessage, run_reasoning_extraction)
-from vllm.entrypoints.openai.reasoning_parsers import (ReasoningParser,
-                                                       ReasoningParserManager)
+from tests.reasoning.utils import DeltaMessage, run_reasoning_extraction
+from vllm.reasoning import ReasoningParser, ReasoningParserManager
 
 parser_name = "granite"
 START_REASONING = "Here is my thought process:"

tests/entrypoints/openai/reasoning_parsers/test_granite_reasoning_parser.py renamed to tests/reasoning/test_granite_reasoning_parser.py

@@ -4,7 +4,7 @@
 
 from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
                                               DeltaMessage)
-from vllm.entrypoints.openai.reasoning_parsers import ReasoningParser
+from vllm.reasoning import ReasoningParser
 
 
 class StreamingReasoningReconstructor:

tests/entrypoints/openai/reasoning_parsers/utils.py renamed to tests/reasoning/utils.py

@@ -23,6 +23,7 @@
 from vllm.logger import init_logger
 from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
 from vllm.plugins import load_general_plugins
+from vllm.reasoning import ReasoningParserManager
 from vllm.test_utils import MODEL_WEIGHTS_S3_BUCKET, MODELS_ON_S3
 from vllm.transformers_utils.utils import check_gguf_file
 from vllm.usage.usage_lib import UsageContext
@@ -1120,7 +1121,7 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
         parser.add_argument(
             "--reasoning-parser",
             type=str,
-            choices=["deepseek_r1", "granite"],
+            choices=list(ReasoningParserManager.reasoning_parsers),
             default=None,
             help=
             "Select the reasoning parser depending on the model that you're "

vllm/engine/arg_utils.py

@@ -2080,8 +2080,9 @@ def _build_logits_processors(
             guided_decoding.backend = guided_decoding.backend or \
                 self.decoding_config.guided_decoding_backend
 
-            logger.debug("Reasoning backend: %s",
-                         self.decoding_config.reasoning_backend)
+            if self.decoding_config.reasoning_backend is not None:
+                logger.debug("Building with reasoning backend %s",
+                             self.decoding_config.reasoning_backend)
 
             processor = get_local_guided_decoding_logits_processor(
                 guided_params=guided_decoding,

vllm/engine/llm_engine.py

@@ -68,7 +68,6 @@
                                               TranscriptionRequest,
                                               TranscriptionResponse,
                                               UnloadLoRAAdapterRequest)
-from vllm.entrypoints.openai.reasoning_parsers import ReasoningParserManager
 # yapf: enable
 from vllm.entrypoints.openai.serving_chat import OpenAIServingChat
 from vllm.entrypoints.openai.serving_completion import OpenAIServingCompletion
@@ -85,6 +84,7 @@
 from vllm.entrypoints.openai.tool_parsers import ToolParserManager
 from vllm.entrypoints.utils import load_aware_call, with_cancellation
 from vllm.logger import init_logger
+from vllm.reasoning import ReasoningParserManager
 from vllm.transformers_utils.config import (
     maybe_register_config_serialize_by_value)
 from vllm.transformers_utils.tokenizer import MistralTokenizer

vllm/entrypoints/openai/api_server.py

@@ -23,8 +23,6 @@
     ChatCompletionStreamResponse, ChatMessage, DeltaFunctionCall, DeltaMessage,
     DeltaToolCall, ErrorResponse, FunctionCall, PromptTokenUsageInfo,
     RequestResponseMetadata, ToolCall, UsageInfo)
-from vllm.entrypoints.openai.reasoning_parsers import (ReasoningParser,
-                                                       ReasoningParserManager)
 from vllm.entrypoints.openai.serving_engine import (OpenAIServing,
                                                     clamp_prompt_logprobs)
 from vllm.entrypoints.openai.serving_models import OpenAIServingModels
@@ -33,6 +31,7 @@
     MistralToolCall)
 from vllm.logger import init_logger
 from vllm.outputs import CompletionOutput, RequestOutput
+from vllm.reasoning import ReasoningParser, ReasoningParserManager
 from vllm.sampling_params import BeamSearchParams, SamplingParams
 from vllm.sequence import Logprob
 from vllm.transformers_utils.tokenizer import AnyTokenizer, MistralTokenizer

vllm/entrypoints/openai/serving_chat.py

@@ -5,10 +5,10 @@
 from typing import TYPE_CHECKING
 
 from vllm.logger import init_logger
-from vllm.model_executor.guided_decoding.reasoner import get_reasoner
 from vllm.model_executor.guided_decoding.utils import (
     convert_lark_to_gbnf, grammar_is_likely_lark,
     has_lmf_unsupported_json_features, has_xgrammar_unsupported_json_features)
+from vllm.reasoning import ReasoningParserManager
 
 if TYPE_CHECKING:
     from transformers import PreTrainedTokenizer
@@ -107,7 +107,11 @@ async def get_guided_decoding_logits_processor(
         model_config: ModelConfig,
         reasoning_backend: str | None = None) -> LogitsProcessor | None:
 
-    reasoner = get_reasoner(tokenizer, reasoning_backend)
+    reasoner = None
+    if reasoning_backend is not None:
+        reasoner_class = ReasoningParserManager.get_reasoning_parser(
+            reasoning_backend)
+        reasoner = reasoner_class(tokenizer)
 
     guided_params = maybe_backend_fallback(guided_params)
 
@@ -146,8 +150,11 @@ def get_local_guided_decoding_logits_processor(
         reasoning_backend: str | None = None) -> LogitsProcessor | None:
     guided_params = maybe_backend_fallback(guided_params)
 
-    # Get the reasoner if needed, it will be None if reasoning_
-    reasoner = get_reasoner(tokenizer, reasoning_backend)
+    reasoner = None
+    if reasoning_backend is not None:
+        reasoner_class = ReasoningParserManager.get_reasoning_parser(
+            reasoning_backend)
+        reasoner = reasoner_class(tokenizer)
 
     # CFG grammar not supported by LMFE, so we use outlines instead
     if guided_params.backend_name == 'outlines':

vllm/model_executor/guided_decoding/__init__.py

@@ -12,7 +12,7 @@
 
 from vllm.model_executor.guided_decoding.outlines_logits_processors import (
     CFGLogitsProcessor, JSONLogitsProcessor, RegexLogitsProcessor)
-from vllm.model_executor.guided_decoding.reasoner import Reasoner
+from vllm.reasoning import ReasoningParser
 from vllm.sampling_params import GuidedDecodingParams
 
 
@@ -61,7 +61,7 @@ class GuidedDecodingMode(Enum):
 async def get_outlines_guided_decoding_logits_processor(
     guided_params: GuidedDecodingParams,
     tokenizer: PreTrainedTokenizerBase,
-    reasoner: Optional[Reasoner],
+    reasoner: Optional[ReasoningParser],
 ) -> Union[JSONLogitsProcessor, RegexLogitsProcessor, CFGLogitsProcessor,
            None]:
     """
@@ -92,7 +92,7 @@ async def get_outlines_guided_decoding_logits_processor(
 def get_local_outlines_guided_decoding_logits_processor(
     guided_params: GuidedDecodingParams,
     tokenizer: PreTrainedTokenizerBase,
-    reasoner: Optional[Reasoner],
+    reasoner: Optional[ReasoningParser],
 ) -> Union[JSONLogitsProcessor, RegexLogitsProcessor, CFGLogitsProcessor,
            None]:
     """
@@ -141,7 +141,7 @@ def _get_logits_processor(
     tokenizer: PreTrainedTokenizerBase,
     mode: GuidedDecodingMode,
     whitespace_pattern: Union[str, None],
-    reasoner: Optional[Reasoner],
+    reasoner: Optional[ReasoningParser],
 ) -> Union[JSONLogitsProcessor, RegexLogitsProcessor, CFGLogitsProcessor]:
     if mode == GuidedDecodingMode.JSON:
         return JSONLogitsProcessor(guide, tokenizer, whitespace_pattern,