feat: support ds v3.2 encoding

Author: soaringk

rtp_llm/models/deepseek_v2.py

@@ -714,3 +714,4 @@ def get_weight_cls():
 register_model("deepseek-v3-mtp", DeepSeekV3Mtp, ["DeepseekV3ForCausalLMNextN"])
 register_model("kimi_k2", DeepSeekV2, [])
 register_model("deepseek_v31", DeepSeekV2, [])
+register_model("deepseek_v32", DeepSeekV2, [])  # DeepSeek-V3.2 uses same architecture as V3.1, with encoding script update

rtp_llm/openai/renderers/__init__.py

@@ -3,6 +3,7 @@
 from .chatglm4_renderer import ChatGlm4Renderer
 from .chatglm45_renderer import ChatGlm45Renderer
 from .deepseekv31_renderer import DeepseekV31Renderer
+from .deepseekv32_renderer import DeepseekV32Renderer
 from .internvl_renderer import InternVLRenderer
 from .kimik2_renderer import KimiK2Renderer
 from .llava_renderer import LlavaRenderer

rtp_llm/openai/renderers/deepseekv32_renderer.py

@@ -0,0 +1,249 @@
+import importlib.util
+import logging
+import os
+import sys
+from typing import Optional
+
+from typing_extensions import override
+
+from rtp_llm.frontend.tokenizer_factory.tokenizers import BaseTokenizer
+from rtp_llm.openai.api_datatype import ChatCompletionRequest
+from rtp_llm.openai.renderer_factory_register import register_renderer
+from rtp_llm.openai.renderers.custom_renderer import RenderedInputs, RendererParams
+from rtp_llm.openai.renderers.reasoning_tool_base_renderer import (
+    ReasoningToolBaseRenderer,
+)
+from rtp_llm.openai.renderers.sglang_helpers.function_call.base_format_detector import (
+    BaseFormatDetector,
+)
+from rtp_llm.openai.renderers.sglang_helpers.reasoning_parser import ReasoningParser
+
+
+class DeepseekV32Renderer(ReasoningToolBaseRenderer):
+    """DeepSeek V3.2 Renderer
+
+    This renderer uses a dedicated Python encoding script instead of Jinja templates.
+    The encoding script is loaded from the checkpoint's "encode" folder.
+
+    Key features:
+    1. Loads encoding_dsv32.py from checkpoint/encode folder
+    2. Uses encode_messages function for rendering
+    3. Supports thinking mode and tool calls
+    """
+
+    def __init__(
+        self,
+        tokenizer: BaseTokenizer,
+        renderer_params: RendererParams,
+    ):
+        # Load the encoding module before calling super().__init__()
+        self.encoding_module = self._load_encoding_module(renderer_params.ckpt_path)
+        super().__init__(tokenizer, renderer_params)
+
+    def _load_encoding_module(self, ckpt_path: str):
+        """
+        Load the encoding_dsv32.py module from the checkpoint's encode folder.
+
+        Args:
+            ckpt_path: Path to the checkpoint directory
+
+        Returns:
+            The loaded encoding module
+
+        Raises:
+            FileNotFoundError: If the encoding script is not found
+            ImportError: If the encoding script cannot be loaded
+        """
+        encode_folder = os.path.join(ckpt_path, "encode")
+        encoding_script_path = os.path.join(encode_folder, "encoding_dsv32.py")
+
+        if not os.path.exists(encoding_script_path):
+            raise FileNotFoundError(
+                f"DeepSeek V3.2 encoding script not found at {encoding_script_path}. "
+                f"Please ensure the checkpoint includes the 'encode' folder with encoding_dsv32.py"
+            )
+
+        try:
+            spec = importlib.util.spec_from_file_location(
+                "encoding_dsv32", encoding_script_path
+            )
+            if spec is None or spec.loader is None:
+                raise ImportError(f"Failed to load spec from {encoding_script_path}")
+
+            module = importlib.util.module_from_spec(spec)
+            sys.modules["encoding_dsv32"] = module
+            spec.loader.exec_module(module)
+
+            logging.info(f"Successfully loaded DeepSeek V3.2 encoding module from {encoding_script_path}")
+            return module
+        except Exception as e:
+            raise ImportError(
+                f"Failed to load DeepSeek V3.2 encoding module from {encoding_script_path}: {str(e)}"
+            )
+
+    @override
+    def _setup_chat_template(self):
+        """
+        DeepSeek V3.2 doesn't use Jinja templates.
+        The chat_template attribute is set to None to indicate custom rendering.
+        """
+        self.chat_template = None
+
+    def _build_prompt(self, request: ChatCompletionRequest) -> str:
+        """
+        Build prompt string using the DeepSeek V3.2 encoding script.
+
+        Args:
+            request: Chat completion request
+
+        Returns:
+            str: Rendered prompt string
+        """
+        # Convert request messages to the format expected by encoding_dsv32
+        messages = []
+        for msg in request.messages:
+            message_dict = {"role": msg.role.value, "content": msg.content}
+
+            # Add tool_calls if present (on assistant messages)
+            if hasattr(msg, "tool_calls") and msg.tool_calls:
+                message_dict["tool_calls"] = [
+                    {
+                        "type": "function",
+                        "id": tc.id,
+                        "function": {
+                            "name": tc.function.name,
+                            "arguments": tc.function.arguments,
+                        }
+                    }
+                    for tc in msg.tool_calls
+                ]
+
+            # Add reasoning_content if present
+            if hasattr(msg, "reasoning_content") and msg.reasoning_content:
+                message_dict["reasoning_content"] = msg.reasoning_content
+
+            messages.append(message_dict)
+
+        # Add tools from request level to the first system message
+        # According to encoding_dsv32 format, tools must be attached to a system message
+        if request.tools:
+            tools_data = [
+                {
+                    "type": "function",
+                    "function": {
+                        "name": tool.function.name,
+                        "description": tool.function.description,
+                        "parameters": tool.function.parameters,
+                    }
+                }
+                for tool in request.tools
+            ]
+
+            # Find the first system message and add tools to it
+            has_system = False
+            for msg in messages:
+                if msg["role"] == "system":
+                    msg["tools"] = tools_data
+                    has_system = True
+                    break
+
+            # If no system message exists, create one with tools
+            if not has_system:
+                messages.insert(0, {
+                    "role": "system",
+                    "content": "",
+                    "tools": tools_data
+                })
+
+        # Determine thinking mode
+        thinking_mode = "thinking" if self.in_think_mode(request) else "chat"
+
+        # Configure encoding
+        # drop_thinking=True: Remove reasoning_content from historical assistant messages
+        # add_default_bos_token=True: Always add BOS token since we encode full messages
+        encode_config = {
+            "thinking_mode": thinking_mode,
+            "drop_thinking": True,
+            "add_default_bos_token": True,
+        }
+
+        # Override with custom configs if provided
+        # Note: context parameter is not used since RTP-LLM always provides full message history
+        if request.chat_template_kwargs:
+            encode_config.update(request.chat_template_kwargs)
+
+        if (
+            request.extra_configs
+            and request.extra_configs.chat_template_kwargs
+            and isinstance(request.extra_configs.chat_template_kwargs, dict)
+        ):
+            encode_config.update(request.extra_configs.chat_template_kwargs)
+
+        try:
+            # Use the encoding module to encode messages
+            rendered_prompt = self.encoding_module.encode_messages(
+                messages, **encode_config
+            )
+
+            logging.debug(
+                f"DeepSeek V3.2 rendered prompt (thinking_mode={thinking_mode}): {rendered_prompt[:200]}..."
+            )
+
+            return rendered_prompt
+        except Exception as e:
+            logging.error(f"Failed to render DeepSeek V3.2 prompt: {str(e)}")
+            raise ValueError(f"Error rendering DeepSeek V3.2 prompt: {str(e)}")
+
+    @override
+    def render_chat(self, request: ChatCompletionRequest) -> RenderedInputs:
+        """
+        Render chat messages using the DeepSeek V3.2 encoding script.
+
+        Args:
+            request: Chat completion request
+
+        Returns:
+            RenderedInputs with encoded token IDs and rendered prompt
+        """
+        prompt = self._build_prompt(request)
+        input_ids = self.tokenizer.encode(prompt)
+        return RenderedInputs(input_ids=input_ids, rendered_prompt=prompt)
+
+    @override
+    def _create_detector(
+        self, _request: ChatCompletionRequest
+    ) -> Optional[BaseFormatDetector]:
+        """
+        DeepSeek V3.2 uses custom DSML format for tool calls.
+        For now, we don't use a detector. Tool parsing will be handled by the encoding module.
+        """
+        return None
+
+    @override
+    def _create_reasoning_parser(
+        self, request: ChatCompletionRequest
+    ) -> Optional[ReasoningParser]:
+        """
+        Create reasoning parser if in thinking mode.
+
+        Args:
+            request: Chat completion request
+
+        Returns:
+            ReasoningParser if thinking mode is enabled, None otherwise
+        """
+        if not self.in_think_mode(request):
+            return None
+
+        try:
+            # Check if the rendered prompt should use thinking mode
+            rendered_result = self.render_chat(request)
+            if "<think>" in rendered_result.rendered_prompt:
+                return ReasoningParser(model_type="deepseek-v3", force_reasoning=True)
+        except Exception:
+            return None
+
+        return None
+
+
+register_renderer("deepseek_v32", DeepseekV32Renderer)