[Model][MiniMax-M2] Support MiniMax-M2 Model (#27535)

Signed-off-by: xuebi <xuebi@minimaxi.com>
Co-authored-by: xuebi <xuebi@minimaxi.com>

tests/models/registry.py

@@ -341,6 +341,9 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
     "MiniMaxM1ForCausalLM": _HfExamplesInfo(
         "MiniMaxAI/MiniMax-M1-40k", trust_remote_code=True
     ),
+    "MiniMaxM2ForCausalLM": _HfExamplesInfo(
+        "MiniMaxAI/MiniMax-M2", trust_remote_code=True
+    ),
     "MistralForCausalLM": _HfExamplesInfo("mistralai/Mistral-7B-Instruct-v0.1"),
     "MixtralForCausalLM": _HfExamplesInfo(
         "mistralai/Mixtral-8x7B-Instruct-v0.1",

vllm/entrypoints/openai/tool_parsers/__init__.py

@@ -16,6 +16,7 @@ from .kimi_k2_tool_parser import KimiK2ToolParser
 from .llama4_pythonic_tool_parser import Llama4PythonicToolParser
 from .llama_tool_parser import Llama3JsonToolParser
 from .longcat_tool_parser import LongcatFlashToolParser
+from .minimax_m2_tool_parser import MinimaxM2ToolParser
 from .minimax_tool_parser import MinimaxToolParser
 from .mistral_tool_parser import MistralToolParser
 from .olmo3_tool_parser import Olmo3PythonicToolParser
@@ -56,4 +57,5 @@ __all__ = [
     "SeedOssToolParser",
     "Step3ToolParser",
     "OpenAIToolParser",
+    "MinimaxM2ToolParser",
 ]

vllm/model_executor/models/registry.py

@@ -131,6 +131,7 @@ _TEXT_GENERATION_MODELS = {
     "MiniMaxForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
     "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
     "MiniMaxM1ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
+    "MiniMaxM2ForCausalLM": ("minimax_m2", "MiniMaxM2ForCausalLM"),
     "MistralForCausalLM": ("llama", "LlamaForCausalLM"),
     "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
     # transformers's mpt class has lower case

vllm/reasoning/__init__.py

@@ -11,6 +11,7 @@ from .gptoss_reasoning_parser import GptOssReasoningParser
 from .granite_reasoning_parser import GraniteReasoningParser
 from .hunyuan_a13b_reasoning_parser import HunyuanA13BReasoningParser
 from .identity_reasoning_parser import IdentityReasoningParser
+from .minimax_m2_reasoning_parser import MiniMaxM2ReasoningParser
 from .mistral_reasoning_parser import MistralReasoningParser
 from .olmo3_reasoning_parser import Olmo3ReasoningParser
 from .qwen3_reasoning_parser import Qwen3ReasoningParser
@@ -34,4 +35,5 @@ __all__ = [
     "Step3ReasoningParser",
     "GptOssReasoningParser",
     "SeedOSSReasoningParser",
+    "MiniMaxM2ReasoningParser",
 ]

vllm/reasoning/minimax_m2_reasoning_parser.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from collections.abc import Sequence
+
+from vllm.entrypoints.openai.protocol import (
+    ChatCompletionRequest,
+    DeltaMessage,
+    ResponsesRequest,
+)
+from vllm.logger import init_logger
+from vllm.reasoning.abs_reasoning_parsers import ReasoningParser, ReasoningParserManager
+from vllm.reasoning.basic_parsers import BaseThinkingReasoningParser
+from vllm.transformers_utils.tokenizer import AnyTokenizer
+
+logger = init_logger(__name__)
+
+
+@ReasoningParserManager.register_module("minimax_m2")
+class MiniMaxM2ReasoningParser(BaseThinkingReasoningParser):
+    """
+    Reasoning parser for MiniMax M2 model.
+    """
+
+    @property
+    def start_token(self) -> str:
+        """The token that starts reasoning content."""
+        return "<think>"
+
+    @property
+    def end_token(self) -> str:
+        """The token that ends reasoning content."""
+        return "</think>"
+
+
+@ReasoningParserManager.register_module("minimax_m2_append_think")
+class MiniMaxM2AppendThinkReasoningParser(ReasoningParser):
+    """
+    Reasoning parser for MiniMax M2 model.
+    """
+
+    def __init__(self, tokenizer: AnyTokenizer, *args, **kwargs):
+        super().__init__(tokenizer, *args, **kwargs)
+        self.end_token_id = self.vocab.get("</think>")
+
+    def is_reasoning_end(self, input_ids: list[int]) -> bool:
+        end_token_id = self.end_token_id
+        return any(input_id == end_token_id for input_id in reversed(input_ids))
+
+    def extract_content_ids(self, input_ids: list[int]) -> list[int]:
+        return input_ids
+
+    def extract_reasoning_content_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:
+        if len(previous_token_ids) == 0:
+            delta_text = "<think>" + delta_text
+        return DeltaMessage(content=delta_text)
+
+    def extract_reasoning_content(
+        self, model_output: str, request: ChatCompletionRequest | ResponsesRequest
+    ) -> tuple[str | None, str | None]:
+        return None, "<think>" + model_output