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

Signed-off-by: Ce Gao <cegao@tensorchord.ai>
This commit is contained in:
Ce Gao
2025-03-28 00:23:30 -07:00
committed by GitHub
parent 2d9045fce8
commit 32b14baf8a
18 changed files with 171 additions and 200 deletions
@@ -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 = [
@@ -166,23 +184,21 @@ TEST_CASES = [
),
]
# 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,
@@ -190,3 +206,17 @@ def test_reasoning(
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 == []
@@ -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:"
@@ -4,7 +4,7 @@ from typing import Optional, Union
from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
DeltaMessage)
from vllm.entrypoints.openai.reasoning_parsers import ReasoningParser
from vllm.reasoning import ReasoningParser
class StreamingReasoningReconstructor: