Add nano_v3 vLLM reasoning parser plugin

nemo_rl/utils/nano_v3_reasoning_parser.py

@@ -0,0 +1,33 @@
+import logging
+
+from vllm.reasoning.abs_reasoning_parsers import ReasoningParserManager
+from vllm.reasoning.deepseek_r1_reasoning_parser import DeepSeekR1ReasoningParser
+
+logger = logging.getLogger(__name__)
+
+# This will print when the module is imported/loaded
+print("=" * 60)
+print("NanoV3ReasoningParser plugin loaded and registered!")
+print("=" * 60)
+
+
+@ReasoningParserManager.register_module("nano_v3")
+class NanoV3ReasoningParser(DeepSeekR1ReasoningParser):
+    def __init__(self, tokenizer, *args, **kwargs):
+        super().__init__(tokenizer, *args, **kwargs)
+        print(f"NanoV3ReasoningParser initialized with tokenizer: {type(tokenizer)}")
+
+    def extract_reasoning(self, model_output, request):
+        reasoning_content, final_content = super().extract_reasoning(
+            model_output, request
+        )
+        if (
+            hasattr(request, "chat_template_kwargs")
+            and request.chat_template_kwargs
+            and request.chat_template_kwargs.get("enable_thinking") is False
+            and final_content is None
+        ):
+            print("NanoV3ReasoningParser: Swapping reasoning and final content (enable_thinking=False)")
+            reasoning_content, final_content = final_content, reasoning_content
+
+        return reasoning_content, final_content