[Bugfix][Parser] Honor a prompt-closed reasoning block in non-streaming parse and reasoning token counts (#61857)
Signed-off-by: Alberto Perdomo <aperdomo@redhat.com>
This commit is contained in:
@@ -9,6 +9,9 @@ import pytest
|
||||
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
|
||||
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
|
||||
from vllm.parser.abstract_parser import DelegatingParser
|
||||
from vllm.parser.engine.parser_engine import ParserEngine
|
||||
from vllm.parser.parser_manager import ParserManager
|
||||
from vllm.parser.qwen3 import qwen3_config
|
||||
from vllm.parser.utils import count_history_tool_calls
|
||||
from vllm.reasoning.basic_parsers import BaseThinkingReasoningParser
|
||||
from vllm.tool_parsers.hermes_tool_parser import Hermes2ProToolParser
|
||||
@@ -417,3 +420,179 @@ def test_parse_auto_tools_no_calls_returns_none(tokenizer):
|
||||
assert reasoning is None
|
||||
assert content == PLAIN_TEXT
|
||||
assert tool_calls is None
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# Prompt that already closes the reasoning block
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
CONTINUE_MESSAGES = [
|
||||
{"role": "user", "content": "hi"},
|
||||
{"role": "assistant", "content": "Sure,"},
|
||||
]
|
||||
REASONED_OUTPUT = "<think>\nlet me think\n</think>\n\n" + PLAIN_TEXT
|
||||
TOOL_TURN_MESSAGES = [
|
||||
{"role": "user", "content": "Weather in Paris?"},
|
||||
{
|
||||
"role": "assistant",
|
||||
"content": "<think>\nneed a tool\n</think>\n\n",
|
||||
"tool_calls": [
|
||||
{
|
||||
"id": "call_1",
|
||||
"type": "function",
|
||||
"function": {"name": "get_weather", "arguments": "{}"},
|
||||
}
|
||||
],
|
||||
},
|
||||
{"role": "tool", "tool_call_id": "call_1", "content": "sunny"},
|
||||
]
|
||||
|
||||
|
||||
def reasoning_parser(tokenizer, name="qwen3"):
|
||||
parser_cls = ParserManager.get_parser(
|
||||
reasoning_parser_name=name, tokenizer=tokenizer
|
||||
)
|
||||
return parser_cls(tokenizer, None)
|
||||
|
||||
|
||||
def encode(tokenizer, text):
|
||||
return tokenizer.encode(text, add_special_tokens=False)
|
||||
|
||||
|
||||
def prompt_case(tokenizer, case):
|
||||
"""(rendered prompt ids, request, model output) for each prompt shape."""
|
||||
if case == "continue_final_message":
|
||||
# The template renders an empty, closed think block before "Sure,".
|
||||
prompt = tokenizer.apply_chat_template(
|
||||
CONTINUE_MESSAGES,
|
||||
tokenize=False,
|
||||
add_generation_prompt=False,
|
||||
continue_final_message=True,
|
||||
)
|
||||
request = make_request(
|
||||
messages=CONTINUE_MESSAGES,
|
||||
add_generation_prompt=False,
|
||||
continue_final_message=True,
|
||||
)
|
||||
return encode(tokenizer, prompt), request, " here is the answer"
|
||||
if case == "template_closed_thinking":
|
||||
# The template switches thinking off, but the parser is not told
|
||||
# through chat_template_kwargs (#53284).
|
||||
prompt = tokenizer.apply_chat_template(
|
||||
[{"role": "user", "content": "hi"}],
|
||||
tokenize=False,
|
||||
add_generation_prompt=True,
|
||||
enable_thinking=False,
|
||||
)
|
||||
return encode(tokenizer, prompt), make_request(), PLAIN_TEXT
|
||||
if case == "earlier_turn_closed_reasoning":
|
||||
# Only a previous assistant turn closes reasoning; this one is open.
|
||||
prompt = tokenizer.apply_chat_template(
|
||||
TOOL_TURN_MESSAGES,
|
||||
tools=[{"type": "function", "function": {"name": "get_weather"}}],
|
||||
tokenize=False,
|
||||
add_generation_prompt=True,
|
||||
)
|
||||
return encode(tokenizer, prompt), make_request(), REASONED_OUTPUT
|
||||
assert case == "reasoning_open"
|
||||
prompt = tokenizer.apply_chat_template(
|
||||
[{"role": "user", "content": "hi"}],
|
||||
tokenize=False,
|
||||
add_generation_prompt=True,
|
||||
)
|
||||
return encode(tokenizer, prompt), make_request(), REASONED_OUTPUT
|
||||
|
||||
|
||||
@pytest.mark.parametrize("case", ["continue_final_message", "template_closed_thinking"])
|
||||
def test_parse_prompt_closed_reasoning_is_content(tokenizer, case):
|
||||
prompt_ids, request, output = prompt_case(tokenizer, case)
|
||||
parser = reasoning_parser(tokenizer)
|
||||
parser.set_prompt_token_ids(prompt_ids)
|
||||
|
||||
reasoning, content, _ = parser.parse(
|
||||
output, request, model_output_token_ids=encode(tokenizer, output)
|
||||
)
|
||||
|
||||
assert reasoning is None
|
||||
assert content == output
|
||||
assert parser.count_reasoning_tokens(encode(tokenizer, output)) == 0
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
("parser_name", "case"),
|
||||
[
|
||||
("qwen3", "reasoning_open"),
|
||||
("qwen3", "earlier_turn_closed_reasoning"),
|
||||
# BaseThinkingReasoningParser scans the prompt without turn
|
||||
# boundaries, so it must not decide non-streaming classification.
|
||||
("deepseek_r1", "earlier_turn_closed_reasoning"),
|
||||
],
|
||||
)
|
||||
def test_parse_prompt_open_reasoning_still_extracted(tokenizer, parser_name, case):
|
||||
prompt_ids, request, output = prompt_case(tokenizer, case)
|
||||
parser = reasoning_parser(tokenizer, parser_name)
|
||||
parser.set_prompt_token_ids(prompt_ids)
|
||||
|
||||
reasoning, content, _ = parser.parse(
|
||||
output, request, model_output_token_ids=encode(tokenizer, output)
|
||||
)
|
||||
|
||||
assert reasoning.strip() == "let me think"
|
||||
assert content.strip() == PLAIN_TEXT
|
||||
assert parser.count_reasoning_tokens(encode(tokenizer, output)) > 0
|
||||
|
||||
|
||||
def test_parse_prompt_ignored_without_turn_boundaries(tokenizer):
|
||||
"""Without turn boundaries, an earlier turn's </think> looks like this
|
||||
turn's, so the engine must not start the output in content."""
|
||||
prompt_ids, request, output = prompt_case(
|
||||
tokenizer, "earlier_turn_closed_reasoning"
|
||||
)
|
||||
parser = ParserEngine(tokenizer, parser_engine_config=qwen3_config(thinking=True))
|
||||
parser.set_prompt_token_ids(prompt_ids)
|
||||
|
||||
reasoning, content, _ = parser.parse(
|
||||
output, request, model_output_token_ids=encode(tokenizer, output)
|
||||
)
|
||||
|
||||
assert reasoning.strip() == "let me think"
|
||||
assert content.strip() == PLAIN_TEXT
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"case",
|
||||
[
|
||||
"continue_final_message",
|
||||
"template_closed_thinking",
|
||||
"reasoning_open",
|
||||
"earlier_turn_closed_reasoning",
|
||||
],
|
||||
)
|
||||
def test_prompt_reasoning_streaming_matches_non_streaming(tokenizer, case):
|
||||
prompt_ids, request, output = prompt_case(tokenizer, case)
|
||||
output_ids = encode(tokenizer, output)
|
||||
|
||||
parser = reasoning_parser(tokenizer)
|
||||
parser.set_prompt_token_ids(prompt_ids)
|
||||
reasoning, content, _ = parser.parse(
|
||||
output, request, model_output_token_ids=output_ids
|
||||
)
|
||||
non_streaming = (reasoning or "", content or "")
|
||||
non_streaming_count = parser.count_reasoning_tokens(output_ids)
|
||||
|
||||
parser = reasoning_parser(tokenizer)
|
||||
streamed_reasoning = streamed_content = ""
|
||||
for i, token_id in enumerate(output_ids):
|
||||
delta = parser.parse_delta(
|
||||
tokenizer.decode([token_id]),
|
||||
[token_id],
|
||||
request,
|
||||
prompt_token_ids=prompt_ids,
|
||||
finished=i == len(output_ids) - 1,
|
||||
)
|
||||
if delta is not None:
|
||||
streamed_reasoning += delta.reasoning or ""
|
||||
streamed_content += delta.content or ""
|
||||
|
||||
assert (streamed_reasoning, streamed_content) == non_streaming
|
||||
assert parser.count_reasoning_tokens(output_ids) == non_streaming_count
|
||||
|
||||
Reference in New Issue
Block a user