[Models] Kimi-K2.5 (#33131)

Signed-off-by: wanglinian <wanglinian@stu.pku.edu.cn>
Signed-off-by: wangln19 <96399074+wangln19@users.noreply.github.com>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Signed-off-by: youkaichao <youkaichao@gmail.com>
Signed-off-by: Roger Wang <hey@rogerw.io>
Co-authored-by: wanglinian <wanglinian@stu.pku.edu.cn>
Co-authored-by: wangln19 <96399074+wangln19@users.noreply.github.com>
Co-authored-by: Zaida Zhou <58739961+zhouzaida@users.noreply.github.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: youkaichao <youkaichao@gmail.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

vllm/reasoning/__init__.py

@@ -53,8 +53,8 @@ _REASONING_PARSERS_TO_REGISTER = {
         "HunyuanA13BReasoningParser",
     ),
     "kimi_k2": (
-        "deepseek_r1_reasoning_parser",
-        "DeepSeekR1ReasoningParser",
+        "kimi_k2_reasoning_parser",
+        "KimiK2ReasoningParser",
     ),
     "minimax_m2": (
         "minimax_m2_reasoning_parser",

vllm/reasoning/kimi_k2_reasoning_parser.py

@@ -0,0 +1,80 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from collections.abc import Sequence
+from typing import TYPE_CHECKING, Any
+
+from transformers import PreTrainedTokenizerBase
+
+from vllm.entrypoints.openai.engine.protocol import DeltaMessage
+from vllm.logger import init_logger
+from vllm.reasoning import ReasoningParser
+from vllm.reasoning.deepseek_r1_reasoning_parser import DeepSeekR1ReasoningParser
+
+from .identity_reasoning_parser import IdentityReasoningParser
+
+if TYPE_CHECKING:
+    from vllm.entrypoints.openai.chat_completion.protocol import (
+        ChatCompletionRequest,
+    )
+else:
+    ChatCompletionRequest = Any
+
+
+logger = init_logger(__name__)
+
+
+class KimiK2ReasoningParser(ReasoningParser):
+    """
+    Kimi K2 parser that delegates to either DeepSeekR1ReasoningParser or
+    IdentityReasoningParser based on `thinking` and `separate_reasoning`.
+
+    Unlike DeepSeekV3ReasoningParser which defaults to NOT thinking,
+    KimiK2ReasoningParser defaults to thinking mode (uses DeepSeekR1ReasoningParser).
+    """
+
+    def __init__(self, tokenizer: PreTrainedTokenizerBase, *args, **kwargs):
+        super().__init__(tokenizer, *args, **kwargs)
+
+        chat_kwargs = kwargs.pop("chat_template_kwargs", {}) or {}
+        # Key difference: default to True instead of False
+        thinking = bool(chat_kwargs.pop("thinking", True))
+
+        if thinking:
+            self._parser = DeepSeekR1ReasoningParser(tokenizer, *args, **kwargs)
+        else:
+            self._parser = IdentityReasoningParser(tokenizer, *args, **kwargs)
+
+    def is_reasoning_end(self, input_ids: Sequence[int]) -> bool:
+        return self._parser.is_reasoning_end(input_ids)
+
+    def is_reasoning_end_streaming(
+        self, input_ids: list[int], delta_ids: list[int]
+    ) -> bool:
+        return self._parser.is_reasoning_end_streaming(input_ids, delta_ids)
+
+    def extract_content_ids(self, input_ids: list[int]) -> list[int]:
+        return self._parser.extract_content_ids(input_ids)
+
+    def extract_reasoning(
+        self, model_output: str, request: "ChatCompletionRequest"
+    ) -> tuple[str | None, str | None]:
+        return self._parser.extract_reasoning(model_output, request)
+
+    def extract_reasoning_streaming(
+        self,
+        previous_text: str,
+        current_text: str,
+        delta_text: str,
+        previous_token_ids: Sequence[int],
+        current_token_ids: Sequence[int],
+        delta_token_ids: Sequence[int],
+    ) -> DeltaMessage | None:
+        return self._parser.extract_reasoning_streaming(
+            previous_text,
+            current_text,
+            delta_text,
+            previous_token_ids,
+            current_token_ids,
+            delta_token_ids,
+        )