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 6, 10:45 UTC
Remove deprecated reasoning_content message field(part-2) (#37480)
Signed-off-by: JartX <sagformas@epdcenter.es> Signed-off-by: Ifta Khairul Alam Adil <ikaadil007@gmail.com> Signed-off-by: Netanel Haber <58652339+netanel-haber@users.noreply.github.com> Signed-off-by: yewentao256 <zhyanwentao@126.com> Signed-off-by: Philip Ottesen <phiott256@gmail.com> Signed-off-by: Woosuk Kwon <woosuk@inferact.ai> Signed-off-by: Michael Goin <mgoin64@gmail.com> Signed-off-by: Giancarlo Delfin <gdelfin@inferact.ai> Signed-off-by: Andy Lo <andy@mistral.ai> Signed-off-by: Thillai Chithambaram <thillaichithambaram.a@gmail.com> Signed-off-by: sihao.li <sihao.li@intel.com> Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: JartX <sagformas@epdcenter.es> Co-authored-by: Netanel Haber <58652339+netanel-haber@users.noreply.github.com> Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com> Co-authored-by: Philip Ottesen <phiott256@gmail.com> Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu> Co-authored-by: Michael Goin <mgoin64@gmail.com> Co-authored-by: Giancarlo Delfin <32987265+TheEpicDolphin@users.noreply.github.com> Co-authored-by: Andy Lo <andy@mistral.ai> Co-authored-by: Thillai Chithambaram <79466435+thillai-c@users.noreply.github.com> Co-authored-by: sihao_li <165983188+1643661061leo@users.noreply.github.com> Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Ifta khairul Alam Adil committed
Mar 19, 2026 at 16:20 UTC
104605cbf2046d09436a41a2367a975f73116138
8 files changed
+36
-38
docs/features/reasoning_outputs.md
+1
-1
@@ -5,7 +5,7 @@ vLLM offers support for reasoning models like [DeepSeek R1](https://huggingface.
5
Reasoning models return an additional `reasoning` field in their outputs, which contains the reasoning steps that led to the final conclusion. This field is not present in the outputs of other models.
6
7
!!! warning
8
- `reasoning` used to be called `reasoning_content`. For now, `reasoning_content` will continue to work. However, we encourage you to migrate to `reasoning` in case `reasoning_content` is removed in future.
8
+ `reasoning` used to be called `reasoning_content`. To migrate, directly replace `reasoning_content` with `reasoning`.
9
10
## Supported Models
11
tests/entrypoints/openai/chat_completion/test_serving_chat.py
+3
-3
@@ -484,7 +484,7 @@ class TestGPTOSSSpeculativeChat:
484
)
485
486
content = ""
487
- reasoning_content = ""
487
+ reasoning = ""
488
async for chunk in stream:
489
delta = chunk.choices[0].delta
490
if delta.content:
@@ -492,9 +492,9 @@ class TestGPTOSSSpeculativeChat:
492
493
chunk_reasoning = getattr(delta, "reasoning", None)
494
if chunk_reasoning:
495
- reasoning_content += delta.reasoning
495
+ reasoning += delta.reasoning
496
497
- assert len(reasoning_content) > 0, "No reasoning was generated."
497
+ assert len(reasoning) > 0, "No reasoning was generated."
498
assert content.strip() == "4"
499
500
tests/reasoning/test_step3p5_reasoning_parser.py
+20
-20
@@ -21,119 +21,119 @@ def step3p5_tokenizer():
21
22
SIMPLE_REASONING = {
23
"output": "This is a reasoning section</think>This is the rest",
24
- "reasoning_content": "This is a reasoning section",
24
+ "reasoning": "This is a reasoning section",
25
"content": "This is the rest",
26
"is_reasoning_end": True,
27
}
28
# need to get into parser again to remove newline after </think>
29
COMPLETE_REASONING = {
30
"output": "This is a reasoning section</think>",
31
- "reasoning_content": "This is a reasoning section",
31
+ "reasoning": "This is a reasoning section",
32
"content": None,
33
"is_reasoning_end": False,
34
}
35
NO_CONTENT = {
36
"output": "This is content",
37
- "reasoning_content": "This is content",
37
+ "reasoning": "This is content",
38
"content": None,
39
"is_reasoning_end": False,
40
}
41
NO_REASONING_STREAMING = {
42
"output": "This is a reasoning section",
43
- "reasoning_content": "This is a reasoning section",
43
+ "reasoning": "This is a reasoning section",
44
"content": None,
45
"is_reasoning_end": False,
46
}
47
MULTIPLE_LINES = {
48
"output": "This\nThat</think>This is the rest\nThat",
49
- "reasoning_content": "This\nThat",
49
+ "reasoning": "This\nThat",
50
"content": "This is the rest\nThat",
51
"is_reasoning_end": True,
52
}
53
SHORTEST_REASONING_NO_STREAMING = {
54
"output": "</think>This is the rest",
55
- "reasoning_content": None,
55
+ "reasoning": None,
56
"content": "This is the rest",
57
"is_reasoning_end": True,
58
}
59
SHORTEST_REASONING = {
60
"output": "</think>This is the rest",
61
- "reasoning_content": None,
61
+ "reasoning": None,
62
"content": "This is the rest",
63
"is_reasoning_end": True,
64
}
65
REASONING_WITH_THINK = {
66
"output": "<think>This is a reasoning section</think>This is the rest",
67
- "reasoning_content": "This is a reasoning section",
67
+ "reasoning": "This is a reasoning section",
68
"content": "This is the rest",
69
"is_reasoning_end": True,
70
}
71
COMPLETE_REASONING_WITH_THINK = {
72
"output": "<think>This is a reasoning section</think>",
73
- "reasoning_content": "This is a reasoning section",
73
+ "reasoning": "This is a reasoning section",
74
"content": None,
75
"is_reasoning_end": False,
76
}
77
MULTIPLE_LINES_WITH_THINK = {
78
"output": "<think>This\nThat</think>This is the rest\nThat",
79
- "reasoning_content": "This\nThat",
79
+ "reasoning": "This\nThat",
80
"content": "This is the rest\nThat",
81
"is_reasoning_end": True,
82
}
83
SHORTEST_REASONING_NO_STREAMING_WITH_THINK = {
84
"output": "</think>This is the rest",
85
- "reasoning_content": None,
85
+ "reasoning": None,
86
"content": "This is the rest",
87
"is_reasoning_end": True,
88
}
89
SHORTEST_REASONING_WITH_THINK = {
90
"output": "</think>This is the rest",
91
- "reasoning_content": None,
91
+ "reasoning": None,
92
"content": "This is the rest",
93
"is_reasoning_end": True,
94
}
95
THINK_NO_END = {
96
"output": "<think>This is a reasoning section",
97
- "reasoning_content": "This is a reasoning section",
97
+ "reasoning": "This is a reasoning section",
98
"content": None,
99
"is_reasoning_end": False,
100
}
101
EMPTY = {
102
"output": "",
103
- "reasoning_content": None,
103
+ "reasoning": None,
104
"content": None,
105
"is_reasoning_end": False,
106
}
107
EMPTY_STREAMING = {
108
"output": "",
109
- "reasoning_content": None,
109
+ "reasoning": None,
110
"content": None,
111
"is_reasoning_end": False,
112
}
113
NEW_LINE = {
114
"output": "\n<think>This is a reasoning section</think>\nThis is the rest",
115
- "reasoning_content": "This is a reasoning section",
115
+ "reasoning": "This is a reasoning section",
116
"content": "This is the rest",
117
"is_reasoning_end": True,
118
}
119
120
NEW_LINE_STREAMING = {
121
"output": "\n<think>This is a reasoning section\n</think>\nThis is the rest",
122
- "reasoning_content": "\nThis is a reasoning section",
122
+ "reasoning": "\nThis is a reasoning section",
123
"content": "This is the rest",
124
"is_reasoning_end": True,
125
}
126
127
NEW_LINE_STREAMING_COMPLEX_CONTENT = {
128
"output": "\n This is a \n reasoning section\n\n\n</think>\n\nThis is the rest",
129
- "reasoning_content": "\n This is a \n reasoning section\n\n",
129
+ "reasoning": "\n This is a \n reasoning section\n\n",
130
"content": "\nThis is the rest",
131
"is_reasoning_end": True,
132
}
133
134
MULTI_TURN_PROMPT_CONTENT = {
135
"output": "<think> This is last turn's reasoning section </think> hello <think>",
136
- "reasoning_content": "",
136
+ "reasoning": "",
137
"content": "",
138
"is_reasoning_end": False,
139
}
@@ -296,7 +296,7 @@ def test_reasoning(
296
print(f"content: {content}")
297
test_id = request.node.callspec.id if hasattr(request.node, "callspec") else None
298
if request.node.callspec.id != "multi_turn_prompt_content":
299
- assert reasoning == param_dict["reasoning_content"]
299
+ assert reasoning == param_dict["reasoning"]
300
assert content == param_dict["content"]
301
302
# Test is_reasoning_end
vllm/entrypoints/openai/parser/responses_parser.py
+3
-3
@@ -61,10 +61,10 @@ class ResponsesParser:
61
# Store the finish_reason from the output
62
self.finish_reason = output.finish_reason
63
64
- reasoning_content, content = self.reasoning_parser_instance.extract_reasoning(
64
+ reasoning, content = self.reasoning_parser_instance.extract_reasoning(
65
output.text, request=self.request
66
)
67
- if reasoning_content:
67
+ if reasoning:
68
self.response_messages.append(
69
ResponseReasoningItem(
70
type="reasoning",
@@ -73,7 +73,7 @@ class ResponsesParser:
73
content=[
74
Content(
75
type="reasoning_text",
76
- text=reasoning_content,
76
+ text=reasoning,
77
)
78
],
79
)
vllm/entrypoints/openai/responses/utils.py
+4
-4
@@ -191,13 +191,13 @@ def _construct_single_message_from_response_item(
191
],
192
)
193
elif isinstance(item, ResponseReasoningItem):
194
- reasoning_content = ""
194
+ reasoning = ""
195
if item.encrypted_content:
196
raise ValueError("Encrypted content is not supported.")
197
elif item.content and len(item.content) >= 1:
198
- reasoning_content = item.content[0].text
198
+ reasoning = item.content[0].text
199
elif len(item.summary) >= 1:
200
- reasoning_content = item.summary[0].text
200
+ reasoning = item.summary[0].text
201
logger.warning(
202
"Using summary text as reasoning content for item %s. "
203
"Please use content instead of summary for "
@@ -206,7 +206,7 @@ def _construct_single_message_from_response_item(
206
)
207
return {
208
"role": "assistant",
209
- "reasoning": reasoning_content,
209
+ "reasoning": reasoning,
210
}
211
elif isinstance(item, ResponseOutputMessage):
212
return {
vllm/parser/abstract_parser.py
+1
-1
@@ -199,7 +199,7 @@ class Parser:
199
request: The request object used to generate the output.
200
201
Returns:
202
- A tuple of (reasoning_content, response_content).
202
+ A tuple of (reasoning, response_content).
203
"""
204
205
@abstractmethod
vllm/reasoning/nemotron_v3_reasoning_parser.py
+3
-5
@@ -17,9 +17,7 @@ class NemotronV3ReasoningParser(DeepSeekR1ReasoningParser):
17
def extract_reasoning(
18
self, model_output: str, request: ChatCompletionRequest | ResponsesRequest
19
) -> tuple[str | None, str | None]:
20
- reasoning_content, final_content = super().extract_reasoning(
21
- model_output, request
22
- )
20
+ reasoning, final_content = super().extract_reasoning(model_output, request)
21
chat_template_kwargs = getattr(request, "chat_template_kwargs", None)
22
23
if (
@@ -30,6 +28,6 @@ class NemotronV3ReasoningParser(DeepSeekR1ReasoningParser):
28
)
29
and final_content is None
30
):
33
- reasoning_content, final_content = final_content, reasoning_content
31
+ reasoning, final_content = final_content, reasoning
32
35
- return reasoning_content, final_content
33
+ return reasoning, final_content
vllm/tool_parsers/step3p5_tool_parser.py
+1
-1
@@ -295,7 +295,7 @@ class StreamingXMLToolCallParser:
295
final_delta = DeltaMessage(
296
role=None,
297
content=None,
298
- reasoning_content=None,
298
+ reasoning=None,
299
tool_calls=[
300
DeltaToolCall(
301
index=self.tool_call_index - 1,