Merge branch 'v0.11.0-dev' into 'v0.11.0-dev'

Update MiniMax-M2/M2.1/M2.5

See merge request dcutoolkit/deeplearing/vllm!437

tests/models/registry.py

@@ -54,7 +54,7 @@ class _HfExamplesInfo:
 
     skip_tokenizer_init: bool = False
     """
-    If true, skip initialization of tokenizer and detokenizer. 
+    If true, skip initialization of tokenizer and detokenizer.
     """
 
     dtype: ModelDType = "auto"
@@ -103,7 +103,7 @@ class _HfExamplesInfo:
 
     use_original_num_layers: bool = False
     """
-    If True, use the original number of layers from the model config 
+    If True, use the original number of layers from the model config
     instead of minimal layers for testing.
     """
 
@@ -300,6 +300,8 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
                                                 revision="a59aa9cbc53b9fb8742ca4e9e1531b9802b6fdc3"),  # noqa: E501
     "MiniMaxM1ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "MiniMaxAI/MiniMax-M1-40k"),
                                             trust_remote_code=True),
+    "MiniMaxM2ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "MiniMaxAI/MiniMax-M2"),
+                                            trust_remote_code=True),
     "MistralForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "mistralai/Mistral-7B-Instruct-v0.1")),
     "MixtralForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "mistralai/Mixtral-8x7B-Instruct-v0.1"),  # noqa: E501
                                           {"tiny": os.path.join(models_path_prefix, "TitanML/tiny-mixtral")}),  # noqa: E501

vllm/reasoning/__init__.py

@@ -12,7 +12,7 @@ from .mistral_reasoning_parser import MistralReasoningParser
 from .qwen3_reasoning_parser import Qwen3ReasoningParser
 from .seedoss_reasoning_parser import SeedOSSReasoningParser
 from .step3_reasoning_parser import Step3ReasoningParser
-from .minimax_m2_reasoning_parser import MiniMaxM2ReasoningParser
+from .minimax_m2_reasoning_parser import MiniMaxM2ReasoningParser,MiniMaxM2AppendThinkReasoningParser
 
 __all__ = [
     "ReasoningParser",
@@ -29,4 +29,5 @@ __all__ = [
     "GptOssReasoningParser",
     "SeedOSSReasoningParser",
     "MiniMaxM2ReasoningParser",
+    "MiniMaxM2AppendThinkReasoningParser"
 ]

vllm/reasoning/minimax_m2_reasoning_parser.py

@@ -74,7 +74,7 @@ class MiniMaxM2ReasoningParser(BaseThinkingReasoningParser):
         # No end token yet, all content is reasoning
         return DeltaMessage(reasoning=delta_text)
 
-
+@ReasoningParserManager.register_module("minimax_m2_append_think")
 class MiniMaxM2AppendThinkReasoningParser(ReasoningParser):
     """
     Reasoning parser for MiniMax M2 model.