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

[MiniMax-M2.1] 新增minimax-m2.1工具调用

See merge request dcutoolkit/deeplearing/vllm!387

vllm/reasoning/minimax_m2_reasoning_parser.py

@@ -6,18 +6,23 @@ from collections.abc import Sequence
 from vllm.entrypoints.openai.protocol import (
     ChatCompletionRequest,
     DeltaMessage,
+    ResponsesRequest,
 )
 from vllm.logger import init_logger
-from vllm.reasoning import ReasoningParser
+from vllm.reasoning 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.
+
+    MiniMax M2 models don't generate <think> start token, only </think> end
+    token. All content before </think> is reasoning, content after is the
+    actual response.
     """
 
     @property
@@ -30,6 +35,45 @@ class MiniMaxM2ReasoningParser(BaseThinkingReasoningParser):
         """The token that ends reasoning content."""
         return "</think>"
 
+    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:
+        """
+        Extract reasoning content from a delta message for streaming.
+
+        MiniMax M2 models don't generate <think> start token, so we assume
+        all content is reasoning until we encounter the </think> end token.
+        """
+        # Skip single end token
+        if len(delta_token_ids) == 1 and delta_token_ids[0] == self.end_token_id:
+            return None
+
+        # Check if end token has already appeared in previous tokens
+        # meaning we're past the reasoning phase
+        if self.end_token_id in previous_token_ids:
+            # We're past the reasoning phase, this is content
+            return DeltaMessage(content=delta_text)
+
+        # Check if end token is in delta tokens
+        if self.end_token_id in delta_token_ids:
+            # End token in delta, split reasoning and content
+            end_index = delta_text.find(self.end_token)
+            reasoning = delta_text[:end_index]
+            content = delta_text[end_index + len(self.end_token) :]
+            return DeltaMessage(
+                reasoning=reasoning if reasoning else None,
+                content=content if content else None,
+            )
+
+        # No end token yet, all content is reasoning
+        return DeltaMessage(reasoning=delta_text)
+
 
 class MiniMaxM2AppendThinkReasoningParser(ReasoningParser):
     """
@@ -40,7 +84,7 @@ class MiniMaxM2AppendThinkReasoningParser(ReasoningParser):
         super().__init__(tokenizer, *args, **kwargs)
         self.end_token_id = self.vocab.get("</think>")
 
-    def is_reasoning_end(self, input_ids: list[int]) -> bool:
+    def is_reasoning_end(self, input_ids: Sequence[int]) -> bool:
         end_token_id = self.end_token_id
         return any(input_id == end_token_id for input_id in reversed(input_ids))
 
@@ -61,6 +105,6 @@ class MiniMaxM2AppendThinkReasoningParser(ReasoningParser):
         return DeltaMessage(content=delta_text)
 
     def extract_reasoning(
-        self, model_output: str, request: ChatCompletionRequest
+        self, model_output: str, request: ChatCompletionRequest | ResponsesRequest
     ) -> tuple[str | None, str | None]:
         return None, "<think>" + model_output
\ No newline at end of file