[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: