Merge pull request #70 from ghost-in-moss/feat/deepseek-reasoning

version:0.2.0

Author: thirdgerb

RELEASES.md

@@ -1,5 +1,21 @@
 # Releases
 
+# v0.2.0
+
+support deepseek-r1. 
+
+* support deepseek-r1 
+  * consider deepseek api protocol is different from openai, add deepseek api adapter.
+  * implement message stage. 
+  * thread history message to prompt filter by stages `[""]` as default.
+* streamlit chat with ghost support staging message stream. 
+* openai o1 do not support system/developer message now, add new compatible option to the model.
+* now llm model and service both have attribute `compatible` to set universe compatible options.
+* prompt object add first_token attribute for debugging.
+* fix bugs
+  * fix shell does not close conversation correctly
+  * fix sequence pipeline handle multiple complete message wrong.
+
 # v0.1.0
 
 first release version.

ghostos/app/configs/llms_conf.yml

@@ -1,4 +1,31 @@
 # from class: ghostos.framework.llms.providers:LLMsYamlConfig
+services:
+  - base_url: https://api.moonshot.cn/v1
+    driver: openai_driver
+    name: moonshot
+    token: $MOONSHOT_API_KEY
+  - base_url: https://api.openai.com/v1
+    driver: openai_driver
+    name: openai
+    proxy: $OPENAI_PROXY
+    token: $OPENAI_API_KEY
+    compatible:
+      use_developer_role: true
+  - base_url: https://api.anthropic.com/v1
+    driver: lite_llm_driver
+    name: anthropic
+    proxy: $ANTHROPIC_PROXY
+    token: $ANTHROPIC_API_KEY
+  - base_url: https://api.deepseek.com/beta
+    driver: deepseek_driver
+    name: deepseek
+    token: $DEEPSEEK_API_KEY
+  - base_url: $AZURE_ENDPOINT
+    name: azure
+    driver: openai_driver
+    azure:
+      api_key: $AZURE_API_KEY
+      api_version: 2023-07-01-preview
 default: gpt-4o
 models:
   gpt-4o-2024-08-06:
@@ -45,17 +72,11 @@ models:
     temperature: 0.7
     timeout: 30
     use_tools: true
-  deepseek-coder:
-    kwargs: { }
-    max_tokens: 2000
-    message_types: null
-    model: deepseek/deepseek-coder
-    n: 1
-    request_timeout: 40
+  deepseek-reasoner:
+    max_tokens: 3400
+    model: deepseek-reasoner
     service: deepseek
-    temperature: 0.7
-    timeout: 30
-    use_tools: true
+    reasoning: {}
   gpt-3.5-turbo:
     kwargs: { }
     max_tokens: 2000
@@ -93,10 +114,14 @@ models:
     model: o1-mini
     service: openai
     reasoning: { }
+    compatible:
+      allow_system_message: false
   gpt-o1:
     model: o1
     service: openai
     reasoning: { }
+    compatible:
+      allow_system_message: false
   gpt-4o:
     kwargs: { }
     max_tokens: 2000
@@ -141,30 +166,3 @@ models:
     temperature: 0.7
     timeout: 30
     use_tools: true
-services:
-  - base_url: https://api.moonshot.cn/v1
-    driver: openai_driver
-    name: moonshot
-    token: $MOONSHOT_API_KEY
-  - base_url: https://api.openai.com/v1
-    driver: openai_driver
-    name: openai
-    proxy: $OPENAI_PROXY
-    token: $OPENAI_API_KEY
-    compatible:
-      use_developer_role: true
-  - base_url: https://api.anthropic.com/v1
-    driver: lite_llm_driver
-    name: anthropic
-    proxy: $ANTHROPIC_PROXY
-    token: $ANTHROPIC_API_KEY
-  - base_url: https://api.deepseek.com/beta
-    driver: openai_driver
-    name: deepseek
-    token: $DEEPSEEK_API_KEY
-  - base_url: $AZURE_ENDPOINT
-    name: azure
-    driver: openai_driver
-    azure:
-      api_key: $AZURE_API_KEY
-      api_version: 2023-07-01-preview

ghostos/core/llms/__init__.py

@@ -1,6 +1,6 @@
 from ghostos.core.llms.configs import (
     ModelConf, ServiceConf, LLMsConfig,
-    OPENAI_DRIVER_NAME, LITELLM_DRIVER_NAME,
+    OPENAI_DRIVER_NAME, LITELLM_DRIVER_NAME, DEEPSEEK_DRIVER_NAME,
 )
 from ghostos.core.llms.abcd import LLMs, LLMDriver, LLMApi
 from ghostos.core.llms.prompt import (

ghostos/core/llms/abcd.py

@@ -17,7 +17,10 @@ class LLMApi(ABC):
     """
 
     service: ServiceConf
+    """service of the api"""
+
     model: ModelConf
+    """model of the api"""
 
     @property
     @abstractmethod
@@ -28,13 +31,15 @@ def name(self) -> str:
     def get_service(self) -> ServiceConf:
         """
         get the service configuration of this API
+        Deprecated.
         """
         pass
 
     @abstractmethod
     def get_model(self) -> ModelConf:
         """
         get new api with the given api_conf and return new LLMAPI
+        Deprecated.
         """
         pass
 
@@ -65,21 +70,34 @@ def chat_completion_chunks(self, prompt: Prompt) -> Iterable[Message]:
         pass
 
     @abstractmethod
-    def reasoning_completion(self, prompt: Prompt, stream: bool) -> Iterable[Message]:
+    def reasoning_completion(self, prompt: Prompt) -> Iterable[Message]:
+        """
+        reasoning completion is not compatible to chat completion.
+        so we need another api.
+        :param prompt:
+        :return:
+        """
+        pass
+
+    @abstractmethod
+    def reasoning_completion_stream(self, prompt: Prompt) -> Iterable[Message]:
         pass
 
     def deliver_chat_completion(self, prompt: Prompt, stream: bool) -> Iterable[Message]:
         """
         逐个发送消息的包.
         """
-        if self.model.reasoning is not None:
-            yield from self.reasoning_completion(prompt, stream)
-
-        elif not stream or not self.model.allow_streaming:
-            message = self.chat_completion(prompt)
-            return [message]
+        if self.model.reasoning:
+            if not stream or not self.model.allow_streaming:
+                yield from self.reasoning_completion(prompt)
+            else:
+                yield from self.reasoning_completion_stream(prompt)
         else:
-            yield from self.chat_completion_chunks(prompt)
+            if not stream or not self.model.allow_streaming:
+                message = self.chat_completion(prompt)
+                return [message]
+            else:
+                yield from self.chat_completion_chunks(prompt)
 
 
 class LLMDriver(ABC):

ghostos/core/llms/configs.py

@@ -11,14 +11,16 @@
 
 __all__ = [
     'ModelConf', 'ServiceConf', 'LLMsConfig',
-    'OPENAI_DRIVER_NAME', 'LITELLM_DRIVER_NAME',
+    'OPENAI_DRIVER_NAME', 'LITELLM_DRIVER_NAME', 'DEEPSEEK_DRIVER_NAME',
 ]
 
 OPENAI_DRIVER_NAME = "openai_driver"
 """default llm driver name for OpenAI llm message protocol """
 
 LITELLM_DRIVER_NAME = "lite_llm_driver"
 
+DEEPSEEK_DRIVER_NAME = "deepseek_driver"
+
 
 class Reasonable(BaseModel):
     """
@@ -28,6 +30,10 @@ class Reasonable(BaseModel):
         "medium",
         description="reasoning effort level",
     )
+    max_completion_tokens: Optional[int] = Field(
+        None,
+        description="max completion tokens",
+    )
 
 
 class ModelConf(Payload):
@@ -46,17 +52,19 @@ class ModelConf(Payload):
     request_timeout: float = Field(default=40, description="request timeout")
     kwargs: Dict[str, Any] = Field(default_factory=dict, description="kwargs")
     use_tools: bool = Field(default=True, description="use tools")
-    max_completion_tokens: Optional[int] = Field(
-        None,
-        description="max completion tokens",
-    )
+
     message_types: Optional[List[str]] = Field(None, description="model allow message types")
     allow_streaming: bool = Field(True, description="if the current model allow streaming")
     reasoning: Optional[Reasonable] = Field(
         default=None,
         description="reasoning configuration",
     )
 
+    compatible: Optional[Compatible] = Field(
+        default=None,
+        description="the model api compatible configuration",
+    )
+
     payloads: Dict[str, Dict] = Field(
         default_factory=dict,
         description="custom payload objects. save strong typed but optional dict."
@@ -67,6 +75,7 @@ class ModelConf(Payload):
 class Compatible(BaseModel):
     use_developer_role: bool = Field(default=False, description="use developer role instead of system")
     allow_system_in_messages: bool = Field(default=True, description="allow system messages in history")
+    allow_system_message: bool = Field(default=True, description="support system message or not")
 
 
 class Azure(BaseModel):
@@ -93,8 +102,8 @@ class ServiceConf(BaseModel):
         description="the adapter driver name of this service. change it only if you know what you are doing",
     )
 
-    compatible: Compatible = Field(
-        default_factory=Compatible,
+    compatible: Optional[Compatible] = Field(
+        default=None,
         description="the model api compatible configuration",
     )
 

ghostos/core/llms/prompt.py

@@ -49,8 +49,9 @@ class Prompt(BaseModel):
     error: Optional[str] = Field(default=None, description="error message")
     created: int = Field(default_factory=timestamp)
     model: Optional[ModelConf] = Field(default=None, description="model conf")
-    run_start: int = Field(default=0, description="start time")
-    run_end: int = Field(default=0, description="end time")
+    run_start: float = Field(default=0.0, description="start time")
+    first_token: float = Field(default=0.0, description="first token")
+    run_end: float = Field(default=0.0, description="end time")
 
     def system_prompt(self) -> str:
         contents = []

ghostos/core/messages/__init__.py

@@ -3,6 +3,7 @@
     FunctionCaller, FunctionOutput,
     MessageClass, MessageKind,
     MessageClassesParser,
+    MessageStage,
 )
 from ghostos.core.messages.message_classes import (
     MessageKindParser,

ghostos/core/messages/message.py

@@ -13,6 +13,7 @@
 
 __all__ = [
     "Message", "Role", "MessageType",
+    "MessageStage",
     "MessageClass", "MessageClassesParser",
     "MessageKind",
     "FunctionCaller", "FunctionOutput",
@@ -187,6 +188,12 @@ def is_protocol_type(cls, value: str) -> bool:
 # todo: 1. 传输协议和存储协议分开.
 # todo: 2. 传输用弱类型.
 # todo: 3. delta 用于流式传输, content part 用来解决富文本, item 解决消息体.
+
+class MessageStage(str, enum.Enum):
+    DEFAULT = ""
+    REASONING = "reasoning"
+
+
 class Message(BaseModel):
     """ message protocol """
 
@@ -242,6 +249,7 @@ def new_head(
             name: Optional[str] = None,
             msg_id: Optional[str] = None,
             call_id: Optional[str] = None,
+            stage: str = "",
     ):
         """
         create a head chunk message
@@ -252,6 +260,7 @@ def new_head(
         :param name:
         :param msg_id:
         :param call_id:
+        :param stage:
         # :param created:
         :return:
         """
@@ -271,6 +280,7 @@ def new_head(
             type=typ_,
             call_id=call_id,
             msg_id=msg_id,
+            stage=stage,
             created=created,
         )
 
@@ -286,6 +296,7 @@ def new_tail(
             # todo: change to call id
             call_id: Optional[str] = None,
             attrs: Optional[Dict[str, Any]] = None,
+            stage: str = "",
     ):
         """
         create a tail message, is the complete message of chunks.
@@ -297,6 +308,7 @@ def new_tail(
         :param msg_id:
         :param call_id:
         :param attrs:
+        :param stage:
         :return:
         """
         msg = cls.new_head(
@@ -307,6 +319,7 @@ def new_tail(
             typ_=type_,
             msg_id=msg_id,
             call_id=call_id,
+            stage=stage,
         )
         msg.seq = "complete"
         msg.attrs = attrs
@@ -322,6 +335,7 @@ def new_chunk(
             name: Optional[str] = None,
             call_id: Optional[str] = None,
             msg_id: Optional[str] = None,
+            stage: str = "",
     ):
         """
         create a chunk message.
@@ -333,6 +347,7 @@ def new_chunk(
             call_id=call_id,
             msg_id=msg_id or "",
             seq="chunk",
+            stage=stage,
         )
 
     def get_content(self) -> str:
@@ -406,6 +421,8 @@ def update(self, pack: "Message") -> None:
         if not self.type:
             # only update when self type is empty (default)
             self.type = pack.type
+        if pack.stage:
+            self.stage = pack.stage
 
         if not self.role:
             self.role = pack.role