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 13, 14:30 UTC
[Bugfix] Fix and add tests for GptOss reasoning parser (#28000)
Signed-off-by: Benjamin Chislett <bchislett@nvidia.com>
Benjamin Chislett committed
Nov 7, 2025 at 14:28 UTC
18903216f5dd4f0378e69667d6f75d4dd14d9c12
2 files changed
+151
-7
tests/reasoning/test_gptoss_reasoning_parser.py
new
+127
@@ -0,0 +1,127 @@
1
+# SPDX-License-Identifier: Apache-2.0
2
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
+
4
+import pytest
5
+from transformers import AutoTokenizer
6
+
7
+from vllm.reasoning import ReasoningParser
8
+from vllm.reasoning.gptoss_reasoning_parser import GptOssReasoningParser
9
+
10
+REASONING_MODEL_NAME = "openai/gpt-oss-120b"
11
+
12
+
13
+@pytest.fixture(scope="module")
14
+def gpt_oss_tokenizer():
15
+ return AutoTokenizer.from_pretrained(REASONING_MODEL_NAME)
16
+
17
+
18
+USER_MESSAGE_START = "<|start|>user<|message|>"
19
+REASONING_SECTION_START = "<|end|><|start|>assistant<|channel|>analysis<|message|>"
20
+ASSISTANT_CONTENT_START_PREFIX = "<|end|><|start|>assistant<|channel|>final"
21
+ASSISTANT_CONTENT_START_SUFFIX = "<|message|>"
22
+ASSISTANT_CONTENT_START = (
23
+ ASSISTANT_CONTENT_START_PREFIX + ASSISTANT_CONTENT_START_SUFFIX
24
+)
25
+
26
+BASIC_CONTENT = {
27
+ "output": REASONING_SECTION_START
28
+ + "This is reasoning"
29
+ + ASSISTANT_CONTENT_START
30
+ + "This is the rest",
31
+ "is_reasoning_end": True,
32
+}
33
+
34
+BASIC_REASONING_ONLY = {
35
+ "output": REASONING_SECTION_START + "This is reasoning" + "<|end|>",
36
+ "is_reasoning_end": False,
37
+}
38
+BASIC_NO_REASONING_NO_ASSISTANT = {
39
+ "output": USER_MESSAGE_START + "This is a user message",
40
+ "is_reasoning_end": False,
41
+}
42
+
43
+# Edge-case where the model omits the assistant tag entirely.
44
+BASIC_NO_REASONING_ASSISTANT = {
45
+ "output": USER_MESSAGE_START + "This is a user message<|end|><|channel|>final",
46
+ "is_reasoning_end": True,
47
+}
48
+
49
+COMPLEX_CONTENT_INCOMPLETE_PREFIX_ONLY = {
50
+ "output": REASONING_SECTION_START
51
+ + "This is reasoning"
52
+ + ASSISTANT_CONTENT_START_PREFIX,
53
+ "is_reasoning_end": False,
54
+}
55
+
56
+COMPLEX_CONTENT_SUFFIX_ONLY = {
57
+ "output": REASONING_SECTION_START
58
+ + "This is reasoning"
59
+ + ASSISTANT_CONTENT_START_SUFFIX,
60
+ "is_reasoning_end": False,
61
+}
62
+
63
+COMPLEX_CONTENT_1_NO_SUFFIX = {
64
+ "output": REASONING_SECTION_START
65
+ + "This is reasoning"
66
+ + ASSISTANT_CONTENT_START_PREFIX
67
+ + "<|constrain|> JSON ",
68
+ "is_reasoning_end": False,
69
+}
70
+
71
+COMPLEX_CONTENT_1 = {
72
+ "output": REASONING_SECTION_START
73
+ + "This is reasoning"
74
+ + ASSISTANT_CONTENT_START_PREFIX
75
+ + "<|constrain|> JSON "
76
+ + ASSISTANT_CONTENT_START_SUFFIX,
77
+ "is_reasoning_end": True,
78
+}
79
+
80
+COMPLEX_CONTENT_1_WITH_CONTENT = {
81
+ "output": REASONING_SECTION_START
82
+ + "This is reasoning"
83
+ + ASSISTANT_CONTENT_START_PREFIX
84
+ + "<|constrain|> JSON "
85
+ + ASSISTANT_CONTENT_START_SUFFIX
86
+ + "This is the rest",
87
+ "is_reasoning_end": True,
88
+}
89
+
90
+COMPLEX_CONTENT_2 = {
91
+ "output": REASONING_SECTION_START
92
+ + "This is reasoning"
93
+ + ASSISTANT_CONTENT_START_PREFIX
94
+ + "<|constrain|>ReplyAction "
95
+ + ASSISTANT_CONTENT_START_SUFFIX
96
+ + "This is the rest",
97
+ "is_reasoning_end": True,
98
+}
99
+
100
+TEST_CASES = [
101
+ BASIC_CONTENT,
102
+ BASIC_REASONING_ONLY,
103
+ COMPLEX_CONTENT_INCOMPLETE_PREFIX_ONLY,
104
+ COMPLEX_CONTENT_SUFFIX_ONLY,
105
+ COMPLEX_CONTENT_1_NO_SUFFIX,
106
+ COMPLEX_CONTENT_1,
107
+ COMPLEX_CONTENT_1_WITH_CONTENT,
108
+ COMPLEX_CONTENT_2,
109
+]
110
+
111
+
112
+@pytest.mark.parametrize(
113
+ "output, is_reasoning_end",
114
+ [(t["output"], t["is_reasoning_end"]) for t in TEST_CASES],
115
+)
116
+def test_gptoss_is_reasoning_end(
117
+ output,
118
+ is_reasoning_end,
119
+ gpt_oss_tokenizer,
120
+):
121
+ output = gpt_oss_tokenizer.tokenize(output)
122
+ parser: ReasoningParser = GptOssReasoningParser(gpt_oss_tokenizer)
123
+
124
+ # Test is_reasoning_end
125
+ output_ids = gpt_oss_tokenizer.convert_tokens_to_ids(output)
126
+ actual_is_reasoning_end = parser.is_reasoning_end(output_ids)
127
+ assert is_reasoning_end == actual_is_reasoning_end
vllm/reasoning/gptoss_reasoning_parser.py
+24
-7
@@ -67,18 +67,35 @@ class GptOssReasoningParser(ReasoningParser):
67
68
def __init__(self, tokenizer: PreTrainedTokenizerBase, *args, **kwargs):
69
super().__init__(tokenizer, *args, **kwargs)
70
- self.reasoning_end_token_ids = self.model_tokenizer.encode(
71
- "<|start|>assistant<|channel|>final<|message|>"
70
+ # The model can output some special tokens between "final" and "<|message|>"
71
+ # So we need to look for both sequences to determine the end of reasoning.
72
+ self.reasoning_end_token_ids_prefix = self.model_tokenizer.encode(
73
+ "<|channel|>final"
74
)
75
+ self.reasoning_end_token_ids_suffix = self.model_tokenizer.encode("<|message|>")
76
+ self.reasoning_max_num_between_tokens = 20
77
78
def is_reasoning_end(self, input_ids: list[int]) -> bool:
75
- end_token_ids = self.reasoning_end_token_ids
76
- assert len(end_token_ids) > 0, "reasoning_end_token_ids is empty"
79
+ end_token_ids_prefix = self.reasoning_end_token_ids_prefix
80
+ end_token_ids_suffix = self.reasoning_end_token_ids_suffix
81
+ assert len(end_token_ids_prefix) > 0, "reasoning_end_token_ids_prefix is empty"
82
+ assert len(end_token_ids_suffix) > 0, "reasoning_end_token_ids_suffix is empty"
83
# Check if the end sequence is present in the input_ids.
84
# We search from the end of input_ids to find the last match.
79
- for i in range(len(input_ids) - len(end_token_ids), -1, -1):
80
- if input_ids[i : i + len(end_token_ids)] == end_token_ids:
81
- return True
85
+ for i in range(len(input_ids) - len(end_token_ids_prefix), -1, -1):
86
+ if input_ids[i : i + len(end_token_ids_prefix)] == end_token_ids_prefix:
87
+ # We have found the prefix, now we look for the suffix after the prefix.
88
+ suffix_start = i + len(end_token_ids_prefix)
89
+ for j in range(
90
+ suffix_start, len(input_ids) - len(end_token_ids_suffix) + 1
91
+ ):
92
+ if j - suffix_start >= self.reasoning_max_num_between_tokens:
93
+ break
94
+ if (
95
+ input_ids[j : j + len(end_token_ids_suffix)]
96
+ == end_token_ids_suffix
97
+ ):
98
+ return True
99
return False
100
101
def extract_content_ids(self, input_ids: list[int]) -> list[int]: