[Frontend] Add Streaming Parser Engine and new Kimi k2.5/k2.6/k2.7 Parser (#46610)

Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Chauncey
2026-06-30 07:53:17 +00:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 8cc242335d
commit 2bc20e8aba
6 changed files with 397 additions and 570 deletions
+94 -1
View File
@@ -31,6 +31,7 @@ from vllm.entrypoints.openai.chat_completion.protocol import (
from vllm.parser.engine.registered_adapters import (
Gemma4Parser,
Glm47MoeParser,
KimiK2Parser,
MinimaxM2Parser,
NemotronV3Parser,
Qwen3Parser,
@@ -717,6 +718,96 @@ def _build_glm47_moe(scenario: Scenario, validate: bool = True) -> Sample:
return sample
# ── Kimi K2 (native tool-call section, starts in REASONING) ──────────
_KIMI_K2_VOCAB: dict[str, int] = {
"<think>": 50,
"</think>": 51,
"<|tool_calls_section_begin|>": 60,
"<|tool_calls_section_end|>": 61,
"<|tool_call_begin|>": 62,
"<|tool_call_end|>": 63,
"<|tool_call_argument_begin|>": 64,
}
def _kimi_k2_tool_segments(
tool_calls: list[ToolCallSpec],
) -> list[tuple[str, bool]]:
segs: list[tuple[str, bool]] = [("<|tool_calls_section_begin|>", True)]
for index, tc in enumerate(tool_calls):
args = json.dumps(tc.arguments, ensure_ascii=False, separators=(",", ":"))
segs.extend(
[
("<|tool_call_begin|>", True),
(f"functions.{tc.name}:{index}\n", False),
("<|tool_call_argument_begin|>", True),
(args, False),
("<|tool_call_end|>", True),
]
)
segs.append(("<|tool_calls_section_end|>", True))
return segs
def _kimi_k2_segments(scenario: Scenario) -> list[tuple[str, bool]]:
segs: list[tuple[str, bool]] = []
if scenario.reasoning is not None:
segs.append(("<think>", True))
segs.append((scenario.reasoning, False))
if scenario.content is not None or scenario.tool_calls is not None:
segs.append(("</think>", True))
if scenario.content is not None:
segs.append((scenario.content, False))
if scenario.tool_calls is not None:
segs.extend(_kimi_k2_tool_segments(scenario.tool_calls))
return segs
def _build_kimi_k2(
scenario: Scenario,
validate: bool = True,
thinking: bool = True,
) -> Sample:
expected_reasoning = (
scenario.reasoning.rstrip()
if (thinking and scenario.reasoning is not None)
else None
)
if thinking and scenario.reasoning is None:
expected_reasoning = ""
sample = _make_sample(
sample_id=f"kimi_k2-{scenario.id}",
description=scenario.description,
vocab=_KIMI_K2_VOCAB,
segments=_kimi_k2_segments(scenario),
expected_reasoning=expected_reasoning,
expected_content=_qwen3_expected_content(scenario),
expected_tool_calls=_expected_tc(scenario),
tools=_expected_tools(scenario),
chat_template_kwargs=None if thinking else {"thinking": False},
)
if validate:
_validate_sample(
sample,
KimiK2Parser,
chat_template_kwargs=sample.chat_template_kwargs,
)
return sample
_KIMI_K2_SCENARIOS = [
*SCENARIOS,
Scenario(
id="trailing-reasoning-whitespace",
description="Reasoning trailing whitespace is stripped",
reasoning="Reasoning with trailing whitespace. \n\t",
content="Done.",
),
]
# ── Registry and public API ──────────────────────────────────────────
_BUILDERS: dict[str, Any] = {
@@ -726,6 +817,7 @@ _BUILDERS: dict[str, Any] = {
"nemotron_v3": _build_nemotron_v3,
"seed_oss": _build_seed_oss,
"glm47_moe": _build_glm47_moe,
"kimi_k2": _build_kimi_k2,
}
@@ -733,7 +825,8 @@ _BUILDERS: dict[str, Any] = {
def build_samples(model: str) -> tuple[Sample, ...]:
"""Build all scenario samples for a model, self-validated."""
builder = _BUILDERS[model]
return tuple(builder(s) for s in SCENARIOS)
scenarios = _KIMI_K2_SCENARIOS if model == "kimi_k2" else SCENARIOS
return tuple(builder(s) for s in scenarios)
def build_sample(model: str, scenario: Scenario) -> Sample:
@@ -7,7 +7,6 @@ import pytest
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
from vllm.entrypoints.openai.engine.protocol import DeltaMessage
from vllm.reasoning.identity_reasoning_parser import IdentityReasoningParser
from vllm.reasoning.kimi_k2_reasoning_parser import KimiK2ReasoningParser
from vllm.tokenizers import get_tokenizer
@@ -33,20 +32,6 @@ def kimi_k2_tokenizer():
return get_tokenizer(tokenizer_name=REASONING_MODEL_NAME, trust_remote_code=True)
def test_parser_selection_thinking_enabled(kimi_k2_tokenizer):
parser = KimiK2ReasoningParser(
kimi_k2_tokenizer, chat_template_kwargs={"thinking": True}
)
assert parser._identity_parser is None
def test_parser_selection_thinking_disabled(kimi_k2_tokenizer):
parser = KimiK2ReasoningParser(
kimi_k2_tokenizer, chat_template_kwargs={"thinking": False}
)
assert isinstance(parser._identity_parser, IdentityReasoningParser)
def test_extract_reasoning_with_think_tags(kimi_k2_tokenizer):
parser = KimiK2ReasoningParser(kimi_k2_tokenizer)
request = ChatCompletionRequest(model="test-model", messages=[], temperature=1.0)
@@ -65,7 +50,7 @@ def test_extract_reasoning_empty_thinking(kimi_k2_tokenizer):
reasoning, content = parser.extract_reasoning(
"<think></think>final answer", request
)
assert reasoning == ""
assert reasoning is None
assert content == "final answer"
@@ -96,8 +81,8 @@ def test_streaming_reasoning_then_content(kimi_k2_tokenizer):
"""Token-by-token streaming: reasoning tokens then content after </think>."""
parser = KimiK2ReasoningParser(kimi_k2_tokenizer)
think_id = parser._start_token_id
end_think_id = parser._end_token_id
think_id = parser._parser_engine._start_token_id
end_think_id = parser._parser_engine._end_token_id
# Use a real token ID from the tokenizer for regular content
regular_id = kimi_k2_tokenizer.encode("hello", add_special_tokens=False)[0]
@@ -154,8 +139,8 @@ def test_streaming_tool_section_ends_reasoning(kimi_k2_tokenizer):
"""<|tool_calls_section_begin|> in delta ends reasoning during streaming."""
parser = KimiK2ReasoningParser(kimi_k2_tokenizer)
think_id = parser._start_token_id
tool_begin_id = parser._tool_section_start_token_id
think_id = parser._parser_engine._start_token_id
tool_begin_id = parser._parser_engine._tool_section_start_token_id
regular_id = kimi_k2_tokenizer.encode("hello", add_special_tokens=False)[0]
# Tool section token arrives — should transition from reasoning to content
@@ -169,50 +154,3 @@ def test_streaming_tool_section_ends_reasoning(kimi_k2_tokenizer):
)
assert isinstance(result, DeltaMessage)
assert result.content == "<|tool_calls_section_begin|>"
def test_streaming_end_token_id_buffered(mock_kimi_k2_tokenizer):
"""When stop sequences buffer text, </think> ID arrives before its text.
The token ID is present in delta_token_ids but the actual string is not
yet in delta_text (still buffered). The parser must return None to wait
for the next delta, instead of calling find() which returns -1 and
silently corrupting the text split.
"""
parser = KimiK2ReasoningParser(mock_kimi_k2_tokenizer)
think_id = parser._start_token_id
end_think_id = parser._end_token_id
# Simulate: </think> ID arrived but text not yet flushed.
# Two token IDs in delta to bypass the single-special-token guard.
result = parser.extract_reasoning_streaming(
previous_text="some reasoning",
current_text="some reasoning extra",
delta_text="extra", # </think> text not yet flushed
previous_token_ids=[think_id],
current_token_ids=[think_id, end_think_id, 999],
delta_token_ids=[end_think_id, 999],
)
assert result is None
def test_streaming_tool_section_id_buffered(mock_kimi_k2_tokenizer):
"""When stop sequences buffer text, tool section start ID arrives before its text.
Same buffering scenario as above but for <|tool_calls_section_begin|>.
Without the guard, find() returns -1 and delta_text[:tool_index] silently
drops the last character of reasoning.
"""
parser = KimiK2ReasoningParser(mock_kimi_k2_tokenizer)
think_id = parser._start_token_id
tool_begin_id = parser._tool_section_start_token_id
result = parser.extract_reasoning_streaming(
previous_text="some reasoning",
current_text="some reasoning extra",
delta_text="extra", # tool section text not yet flushed
previous_token_ids=[think_id],
current_token_ids=[think_id, tool_begin_id, 999],
delta_token_ids=[tool_begin_id, 999],
)
assert result is None