biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix]: Reasoning output bug according to the chat template change (#13025)

Browse Source 
Signed-off-by: Ce Gao <cegao@tensorchord.ai>
This commit is contained in:
Ce Gao
2025-02-11 15:49:03 +08:00
committed by GitHub
parent 78a141d768
commit fc6485d277
3 changed files with 127 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files

108
tests/entrypoints/openai/reasoning_parsers/test_deepseekr1_reasoning_parser.py
View File

@@ -15,32 +15,62 @@ start_token = "<think>"
end_token = "</think>"
SIMPLE_REASONING = {
"output": "<think>This is a reasoning section</think>This is the rest",
"output": "This is a reasoning section</think>This is the rest",
"reasoning_content": "This is a reasoning section",
"content": "This is the rest",
}
COMPLETE_REASONING = {
"output": "<think>This is a reasoning section</think>",
"output": "This is a reasoning section</think>",
"reasoning_content": "This is a reasoning section",
"content": None,
}
NO_REASONING = {
"output": "This is a reasoning section",
"output": "This is content",
"reasoning_content": None,
"content": "This is a reasoning section",
"content": "This is content",
}
NO_REASONING_STREAMING = {
"output": "This is a reasoning section",
"reasoning_content": "This is a reasoning section",
"content": None,
}
MULTIPLE_LINES = {
"output": "<think>This\nThat</think>This is the rest\nThat",
"output": "This\nThat</think>This is the rest\nThat",
"reasoning_content": "This\nThat",
"content": "This is the rest\nThat",
}
SHORTEST_REASONING_NO_STREAMING = {
"output": "<think></think>This is the rest",
"output": "</think>This is the rest",
"reasoning_content": "",
"content": "This is the rest",
}
SHORTEST_REASONING = {
"output": "<think></think>This is the rest",
"output": "</think>This is the rest",
"reasoning_content": None,
"content": "This is the rest",
}
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",
}
COMPLETE_REASONING_WITH_THINK = {
"output": "<think>This is a reasoning section</think>",
"reasoning_content": "This is a reasoning section",
"content": None,
}
MULTIPLE_LINES_WITH_THINK = {
"output": "<think>This\nThat</think>This is the rest\nThat",
"reasoning_content": "This\nThat",
"content": "This is the rest\nThat",
}
SHORTEST_REASONING_NO_STREAMING_WITH_THINK = {
"output": "</think>This is the rest",
"reasoning_content": "",
"content": "This is the rest",
}
SHORTEST_REASONING_WITH_THINK = {
"output": "</think>This is the rest",
"reasoning_content": None,
"content": "This is the rest",
}
@@ -49,37 +79,37 @@ TEST_CASES = [
pytest.param(
False,
SIMPLE_REASONING,
id="simple_streaming",
id="simple_reasoning",
),
pytest.param(
True,
SIMPLE_REASONING,
id="simple_streaming",
id="simple_reasoning_streaming",
),
pytest.param(
False,
COMPLETE_REASONING,
id="complete_streaming",
id="complete_reasoning",
),
pytest.param(
True,
COMPLETE_REASONING,
id="complete_streaming",
id="complete_reasoning_streaming",
),
pytest.param(
False,
NO_REASONING,
id="no_streaming",
id="no_reasoning_token",
),
pytest.param(
True,
NO_REASONING,
id="no_streaming",
NO_REASONING_STREAMING,
id="no_reasoning_token_streaming",
),
pytest.param(
False,
MULTIPLE_LINES,
id="multiple_lines_streaming",
id="multiple_lines",
),
pytest.param(
True,
@@ -89,23 +119,65 @@ TEST_CASES = [
pytest.param(
True,
SHORTEST_REASONING,
id="shortest_streaming",
id="shortest",
),
pytest.param(
False,
SHORTEST_REASONING_NO_STREAMING,
id="shortest_streaming",
),
pytest.param(
False,
REASONING_WITH_THINK,
id="reasoning_with_think",
),
pytest.param(
True,
REASONING_WITH_THINK,
id="reasoning_with_think_streaming",
),
pytest.param(
False,
COMPLETE_REASONING_WITH_THINK,
id="complete_reasoning_with_think",
),
pytest.param(
True,
COMPLETE_REASONING_WITH_THINK,
id="complete_reasoning_with_think_streaming",
),
pytest.param(
False,
MULTIPLE_LINES_WITH_THINK,
id="multiple_lines_with_think",
),
pytest.param(
True,
MULTIPLE_LINES_WITH_THINK,
id="multiple_lines_with_think_streaming",
),
pytest.param(
False,
SHORTEST_REASONING_NO_STREAMING_WITH_THINK,
id="shortest_with_think",
),
pytest.param(
True,
SHORTEST_REASONING_WITH_THINK,
id="shortest_with_think_streaming",
),
]
# 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,
):
tokenizer = AutoTokenizer.from_pretrained("facebook/opt-125m")
tokenizer.add_tokens([start_token, end_token])
output = tokenizer.tokenize(param_dict["output"])
# decode everything to tokens
output_tokens: List[str] = [