PaddlePaddle · fgeygfe · Dec 5, 2025
diff --git a/docs/features/reasoning_output.md b/docs/features/reasoning_output.md
@@ -11,6 +11,9 @@ Reasoning models return an additional `reasoning_content` field in their output,
 | baidu/ERNIE-4.5-VL-28B-A3B-Paddle | ernie-45-vl |    ✅    |  ❌  |"chat_template_kwargs":{"enable_thinking": true/false}|
 | baidu/ERNIE-4.5-21B-A3B-Thinking  | ernie-x1  |   ✅ Not supported for turning off   | ✅|❌|
 | baidu/ERNIE-4.5-VL-28B-A3B-Thinking  | ernie-45-vl-thinking  |   ✅ Not recommended to turn off   | ✅|"chat_template_kwargs": {"options": {"thinking_mode": "open/close"}}|
+| unsloth/DeepSeek-V3.1-BF16  | deepseek  | ❌ (thinking mode off by default)  | ✅|❌|
+| unsloth/DeepSeek-V3-0324-BF16  | deepseek  | ✅ (thinking mode on by default)  | ✅|❌|
+| unsloth/DeepSeek-R1-BF16  | deepseek  | ✅ (thinking mode on by default)  | ✅|❌|
 
 The reasoning model requires a specified parser to extract reasoning content. Referring to the `thinking switch parameters` of each model can turn off the model's thinking mode.
 
@@ -159,3 +162,31 @@ Model output example
 }
 ```
 More reference documentation related to tool calling usage：  [Tool Calling](./tool_calling.md)
+
+## Error Handling and Invalid Format
+
+The DeepSeek reasoning parser handles various invalid or incomplete format scenarios gracefully:
+
+### Missing Start Tag
+If the model output contains only the end tag without the start tag:
+- **Input**: `abc</think>xyz`
+- **Output**: `reasoning_content="abc"`, `content="xyz"`
+- The parser extracts content before the end tag as reasoning, and content after as reply.
+
+### Missing End Tag
+If the model output contains only the start tag without the end tag:
+- **Input**: `<think>abc`
+- **Output**: `reasoning_content="abc"`, `content=None`
+- The parser treats all content as reasoning when the end tag is missing.
+
+### No Reasoning Tags (Thinking Mode Off)
+When thinking mode is disabled (e.g., DeepSeek-V3.1 by default):
+- **Input**: `direct response`
+- **Output**: `reasoning_content=None`, `content="direct response"`
+- The parser treats the entire output as reply content.
+
+### Protocol Violation with Tool Calls
+If there is non-whitespace content between the reasoning end tag and tool calls:
+- **Input**: `<think>thinking</think>\n\nABC\n<｜tool▁calls▁begin｜>...`
+- **Output**: Tool calls are not parsed, entire output is returned as `content`
+- This ensures tool calls are only parsed when they immediately follow reasoning content.
diff --git a/docs/features/tool_calling.md b/docs/features/tool_calling.md
@@ -8,6 +8,13 @@ This document describes how to configure the server in FastDeploy to use the too
 | baidu/ERNIE-4.5-21B-A3B-Thinking  | ernie-x1  |
 | baidu/ERNIE-4.5-VL-28B-A3B-Thinking  | ernie-45-vl-thinking  |
 
+## Tool Call parser for DeepSeek series models
+| Model Name     | Parser Name       |
+|---------------|-------------|
+| unsloth/DeepSeek-V3.1-BF16  | deepseek  |
+| unsloth/DeepSeek-V3-0324-BF16  | deepseek  |
+| unsloth/DeepSeek-R1-BF16  | deepseek  |
+
 ## Quickstart
 
 ### Starting FastDeploy with Tool Calling Enabled.
@@ -90,6 +97,27 @@ The example output is as follows. It shows that the model's output of the though
 }
 ```
 
+## Error Handling and Invalid Format
+
+The DeepSeek tool parser handles various invalid or incomplete format scenarios:
+
+### Protocol Violation
+If there is non-whitespace content between the reasoning end tag (`</think>`) and tool calls:
+- **Input**: `<think>thinking</think>\n\nABC\n<｜tool▁calls▁begin｜>...`
+- **Output**: `tools_called=False`, `tool_calls=None`, `content=<entire_output>`
+- Tool calls are not parsed when protocol is violated. The entire output is returned as content.
+
+### Malformed JSON Arguments
+If the tool call arguments contain invalid JSON:
+- **Input**: `<｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location": "北京", "unit":}<｜tool▁call▁end｜>`
+- **Output**: The parser attempts to use `partial_json_parser` to recover valid JSON. If recovery fails, it returns an empty object `{}` or the raw text.
+- This ensures graceful handling of incomplete JSON during streaming.
+
+### Missing Tool Call End Tag
+If a tool call is incomplete (missing end tag):
+- **Input**: `<｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location": "北京"`
+- **Output**: In streaming mode, the parser waits for more data. In non-streaming mode, it attempts to extract what's available.
+
 ## Parallel Tool Calls
 If the model can generate parallel tool calls, FastDeploy will return a list:
 ```bash

diff --git a/docs/zh/features/reasoning_output.md b/docs/zh/features/reasoning_output.md
@@ -11,6 +11,9 @@
 | baidu/ERNIE-4.5-VL-28B-A3B-Paddle | ernie-45-vl |    ✅    |  ❌  |"chat_template_kwargs":{"enable_thinking": true/false}|
 | baidu/ERNIE-4.5-21B-A3B-Thinking  | ernie-x1  |   ✅不支持关思考  | ✅|❌|
 | baidu/ERNIE-4.5-VL-28B-A3B-Thinking  | ernie-45-vl-thinking  |   ✅不推荐关闭   | ✅|"chat_template_kwargs": {"options": {"thinking_mode": "open/close"}}|
+| unsloth/DeepSeek-V3.1-BF16  | deepseek  | ❌ (默认关闭思考模式)  | ✅|❌|
+| unsloth/DeepSeek-V3-0324-BF16  | deepseek  | ✅ (默认开启思考模式)  | ✅|❌|
+| unsloth/DeepSeek-R1-BF16  | deepseek  | ✅ (默认开启思考模式)  | ✅|❌|
 
 思考模型需要指定解析器,以便于对思考内容进行解析. 参考各个模型的 `思考开关控制参数` 可以关闭模型思考模式.
 
@@ -158,3 +161,31 @@ curl -X POST "http://0.0.0.0:8390/v1/chat/completions" \
 }
 ```
 更多工具调用相关的使用参考文档  [Tool Calling](./tool_calling.md)
+
+## 错误处理和格式不合法情况
+
+DeepSeek 推理解析器能够优雅地处理各种格式不合法或不完整的情况：
+
+### 缺少起始标签
+如果模型输出只包含结束标签而没有起始标签：
+- **输入**: `abc</think>xyz`
+- **输出**: `reasoning_content="abc"`, `content="xyz"`
+- 解析器会将结束标签之前的内容提取为思考内容，之后的内容提取为回复内容。
+
+### 缺少结束标签
+如果模型输出只包含起始标签而没有结束标签：
+- **输入**: `<think>abc`
+- **输出**: `reasoning_content="abc"`, `content=None`
+- 解析器会将所有内容视为思考内容。
+
+### 无思考标签（思考模式关闭）
+当思考模式被关闭时（例如 DeepSeek-V3.1 默认关闭）：
+- **输入**: `direct response`
+- **输出**: `reasoning_content=None`, `content="direct response"`
+- 解析器会将整个输出视为回复内容。
+
+### 协议不规范（工具调用前有非空白字符）
+如果思考结束标签和工具调用之间存在非空白字符：
+- **输入**: `<think>thinking</think>\n\nABC\n<｜tool▁calls▁begin｜>...`
+- **输出**: 工具调用不会被解析，整个输出作为 `content` 返回
+- 这确保了只有在工具调用紧跟在思考内容之后时才会被解析。
diff --git a/docs/zh/features/tool_calling.md b/docs/zh/features/tool_calling.md
@@ -8,6 +8,13 @@
 | baidu/ERNIE-4.5-21B-A3B-Thinking  | ernie-x1  |
 | baidu/ERNIE-4.5-VL-28B-A3B-Thinking  | ernie-45-vl-thinking  |
 
+## DeepSeek系列模型配套工具解释器
+| 模型名称          | 解析器名称       |
+|---------------|-------------|
+| unsloth/DeepSeek-V3.1-BF16  | deepseek  |
+| unsloth/DeepSeek-V3-0324-BF16  | deepseek  |
+| unsloth/DeepSeek-R1-BF16  | deepseek  |
+
 ## 快速开始
 
 ### 启动包含解析器的FastDeploy
@@ -92,6 +99,27 @@ curl -X POST http://0.0.0.0:8000/v1/chat/completions \
     ]
 }
 ```
+## 错误处理和格式不合法情况
+
+DeepSeek 工具解析器能够处理各种格式不合法或不完整的情况：
+
+### 协议不规范
+如果思考结束标签（`</think>`）和工具调用之间存在非空白字符：
+- **输入**: `<think>thinking</think>\n\nABC\n<｜tool▁calls▁begin｜>...`
+- **输出**: `tools_called=False`, `tool_calls=None`, `content=<完整输出>`
+- 当协议不规范时，工具调用不会被解析，整个输出作为 content 返回。
+
+### JSON 参数格式错误
+如果工具调用的参数包含无效的 JSON：
+- **输入**: `<｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location": "北京", "unit":}<｜tool▁call▁end｜>`
+- **输出**: 解析器会尝试使用 `partial_json_parser` 来恢复有效的 JSON。如果恢复失败，会返回空对象 `{}` 或原始文本。
+- 这确保了在流式输出过程中能够优雅地处理不完整的 JSON。
+
+### 缺少工具调用结束标签
+如果工具调用不完整（缺少结束标签）：
+- **输入**: `<｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location": "北京"`
+- **输出**: 在流式模式下，解析器会等待更多数据。在非流式模式下，会尝试提取可用的内容。
+
 ## 并行工具调用
 
 如果模型能够生成多个并行的工具调用，FastDeploy 会返回一个列表：