Merge branch 'main' into consciousness_capsule

Author: thirdgerb

ghostos/framework/llms/__init__.py

@@ -1,5 +1,6 @@
 from ghostos.core.llms import LLMs, Prompt, PromptStorage
 from ghostos.framework.llms.llms import LLMsImpl
-from ghostos.framework.llms.openai_driver import OpenAIDriver, OpenAIAdapter, LitellmAdapter
+from ghostos.framework.llms.openai_driver import OpenAIDriver, OpenAIAdapter
+from ghostos.framework.llms.lite_llm_driver import LitellmAdapter
 from ghostos.framework.llms.providers import ConfigBasedLLMsProvider, PromptStorageInWorkspaceProvider, LLMsYamlConfig
 from ghostos.framework.llms.prompt_storage_impl import PromptStorageImpl

ghostos/framework/llms/deepseek_driver.py

@@ -0,0 +1,48 @@
+from typing import Iterable
+from ghostos.core.llms.configs import ServiceConf, ModelConf, DEEPSEEK_DRIVER_NAME
+from ghostos.core.llms.abcd import LLMApi
+from ghostos.core.llms.prompt import Prompt, PromptPayload
+from ghostos.core.messages import Message, MessageStage
+from ghostos.helpers.timeutils import timestamp_ms
+from ghostos.framework.llms.openai_driver import OpenAIDriver, OpenAIAdapter, ChatCompletion, ChatCompletionChunk
+
+
+class DeepseekAdapter(OpenAIAdapter):
+    def _from_openai_chat_completion_item(self, message: ChatCompletion) -> Iterable[Message]:
+        cc_item = message.choices[0].message
+        if reasoning := cc_item.reasoning_content:
+            reasoning_message = Message.new_tail(content=reasoning)
+            reasoning_message.stage = MessageStage.REASONING.value
+            yield reasoning_message
+
+        yield self._parser.from_chat_completion(cc_item)
+
+    def reasoning_completion_stream(self, prompt: Prompt) -> Iterable[ChatCompletionChunk]:
+        try:
+            chunks: Iterable[ChatCompletionChunk] = self._reasoning_completion_stream(prompt)
+            messages = self._from_openai_chat_completion_chunks(chunks)
+            prompt_payload = PromptPayload.from_prompt(prompt)
+            output = []
+            for chunk in messages:
+                if not prompt.first_token:
+                    prompt.first_token = timestamp_ms()
+                yield chunk
+                if chunk.is_complete():
+                    self.model.set_payload(chunk)
+                    prompt_payload.set_payload(chunk)
+                    output.append(chunk)
+            prompt.added = output
+        except Exception as e:
+            prompt.error = str(e)
+            raise
+        finally:
+            self._storage.save(prompt)
+
+
+class DeepseekDriver(OpenAIDriver):
+
+    def driver_name(self) -> str:
+        return DEEPSEEK_DRIVER_NAME
+
+    def new(self, service: ServiceConf, model: ModelConf, api_name: str = "") -> LLMApi:
+        return DeepseekAdapter(service, model, self._parser, self._storage, self._logger, api_name=api_name)

ghostos/framework/llms/lite_llm_driver.py

@@ -0,0 +1,50 @@
+from typing import List
+from ghostos.core.llms.configs import ServiceConf, ModelConf, LITELLM_DRIVER_NAME
+from ghostos.core.llms.abcd import LLMApi
+from ghostos.core.messages import Role, Message
+from ghostos.core.llms.prompt import Prompt
+from ghostos.framework.llms.openai_driver import OpenAIDriver, OpenAIAdapter, ChatCompletionMessageParam, ChatCompletion
+
+
+class LitellmAdapter(OpenAIAdapter):
+    """
+    adapter class wrap openai api to ghostos.blueprint.kernel.llms.LLMApi
+    """
+
+    def _chat_completion(self, chat: Prompt, stream: bool) -> ChatCompletion:
+        import litellm
+        messages = chat.get_messages()
+        messages = self.parse_message_params(messages)
+        response = litellm.completion(
+            model=self.model.model,
+            messages=list(messages),
+            timeout=self.model.timeout,
+            temperature=self.model.temperature,
+            n=self.model.n,
+            # not support stream yet
+            stream=False,
+            api_key=self.service.token,
+        )
+        return response.choices[0].message
+
+    def parse_message_params(self, messages: List[Message]) -> List[ChatCompletionMessageParam]:
+        parsed = super().parse_message_params(messages)
+        outputs = []
+        count = 0
+        for message in parsed:
+            # filter all the system message to __system__ user message.
+            if count > 0 and "role" in message and message["role"] == Role.SYSTEM.value:
+                message["role"] = Role.USER.value
+                message["name"] = "__system__"
+            outputs.append(message)
+            count += 1
+        return outputs
+
+
+class LiteLLMDriver(OpenAIDriver):
+
+    def driver_name(self) -> str:
+        return LITELLM_DRIVER_NAME
+
+    def new(self, service: ServiceConf, model: ModelConf, api_name: str = "") -> LLMApi:
+        return LitellmAdapter(service, model, self._parser, self._storage, self._logger, api_name=api_name)

ghostos/framework/llms/openai_driver.py

@@ -3,7 +3,7 @@
 from httpx import Client
 from httpx_socks import SyncProxyTransport
 from openai import NOT_GIVEN, NotGiven
-from openai.types.chat import ChatCompletion, ChatCompletionReasoningEffort
+from openai.types.chat import ChatCompletion
 from openai.types.chat.chat_completion_stream_options_param import ChatCompletionStreamOptionsParam
 from openai.types.chat.chat_completion_message_param import ChatCompletionMessageParam
 from openai.types.chat.chat_completion_tool_param import ChatCompletionToolParam
@@ -12,22 +12,18 @@
 from ghostos.contracts.logger import LoggerItf, get_ghostos_logger
 from ghostos.helpers import timestamp_ms
 from ghostos.core.messages import (
-    MessageStage,
     Message, OpenAIMessageParser, DefaultOpenAIMessageParser,
     CompletionUsagePayload, Role,
 )
 from ghostos.core.llms import (
-    LLMDriver, LLMApi, ModelConf, ServiceConf, OPENAI_DRIVER_NAME, LITELLM_DRIVER_NAME, DEEPSEEK_DRIVER_NAME,
-    Prompt, PromptPayload, PromptStorage,
-    Compatible,
+    LLMApi, LLMDriver,
+    ModelConf, ServiceConf, Compatible,
+    OPENAI_DRIVER_NAME, LITELLM_DRIVER_NAME, DEEPSEEK_DRIVER_NAME,
     FunctionalToken,
+    Prompt, PromptPayload, PromptStorage
 )
 
-__all__ = [
-    'OpenAIDriver', 'OpenAIAdapter',
-    'LitellmAdapter', 'LiteLLMDriver',
-    'DeepseekDriver', 'DeepseekAdapter',
-]
+__all__ = ['OpenAIDriver', 'OpenAIAdapter']
 
 
 class FunctionalTokenPrompt(str):
@@ -438,97 +434,3 @@ def driver_name(self) -> str:
     def new(self, service: ServiceConf, model: ModelConf, api_name: str = "") -> LLMApi:
         get_ghostos_logger().debug(f"new llm api %s at service %s", model.model, service.name)
         return OpenAIAdapter(service, model, self._parser, self._storage, self._logger, api_name=api_name)
-
-
-class LitellmAdapter(OpenAIAdapter):
-    """
-    adapter class wrap openai api to ghostos.blueprint.kernel.llms.LLMApi
-    """
-
-    def _chat_completion(self, chat: Prompt, stream: bool) -> ChatCompletion:
-        import litellm
-        messages = chat.get_messages()
-        messages = self.parse_message_params(messages)
-        response = litellm.completion(
-            model=self.model.model,
-            messages=list(messages),
-            timeout=self.model.timeout,
-            temperature=self.model.temperature,
-            n=self.model.n,
-            # not support stream yet
-            stream=False,
-            api_key=self.service.token,
-        )
-        return response.choices[0].message
-
-    def parse_message_params(self, messages: List[Message]) -> List[ChatCompletionMessageParam]:
-        parsed = super().parse_message_params(messages)
-        outputs = []
-        count = 0
-        for message in parsed:
-            # filter all the system message to __system__ user message.
-            if count > 0 and "role" in message and message["role"] == Role.SYSTEM.value:
-                message["role"] = Role.USER.value
-                message["name"] = "__system__"
-            outputs.append(message)
-            count += 1
-        return outputs
-
-
-# todo: move to lite_llm_driver. shall not locate here at very first.
-class LiteLLMDriver(OpenAIDriver):
-
-    def driver_name(self) -> str:
-        return LITELLM_DRIVER_NAME
-
-    def new(self, service: ServiceConf, model: ModelConf, api_name: str = "") -> LLMApi:
-        return LitellmAdapter(service, model, self._parser, self._storage, self._logger, api_name=api_name)
-
-
-class DeepseekAdapter(OpenAIAdapter):
-    # def reasoning_completion(self, prompt: Prompt, stream: bool) -> Iterable[Message]:
-    #     if not stream:
-    #         yield from self._reasoning_completion_none_stream(prompt)
-    #     else:
-    #         yield from self._reasoning_completion_stream(prompt)
-    #
-
-    def _from_openai_chat_completion_item(self, message: ChatCompletion) -> Iterable[Message]:
-        cc_item = message.choices[0].message
-        if reasoning := cc_item.reasoning_content:
-            reasoning_message = Message.new_tail(content=reasoning)
-            reasoning_message.stage = MessageStage.REASONING.value
-            yield reasoning_message
-
-        yield self._parser.from_chat_completion(cc_item)
-
-    def reasoning_completion_stream(self, prompt: Prompt) -> Iterable[ChatCompletionChunk]:
-        try:
-            chunks: Iterable[ChatCompletionChunk] = self._reasoning_completion_stream(prompt)
-            messages = self._from_openai_chat_completion_chunks(chunks)
-            prompt_payload = PromptPayload.from_prompt(prompt)
-            output = []
-            for chunk in messages:
-                if not prompt.first_token:
-                    prompt.first_token = timestamp_ms()
-                yield chunk
-                if chunk.is_complete():
-                    self.model.set_payload(chunk)
-                    prompt_payload.set_payload(chunk)
-                    output.append(chunk)
-            prompt.added = output
-        except Exception as e:
-            prompt.error = str(e)
-            raise
-        finally:
-            self._storage.save(prompt)
-
-
-# todo: move to deepseek driver
-class DeepseekDriver(OpenAIDriver):
-
-    def driver_name(self) -> str:
-        return DEEPSEEK_DRIVER_NAME
-
-    def new(self, service: ServiceConf, model: ModelConf, api_name: str = "") -> LLMApi:
-        return DeepseekAdapter(service, model, self._parser, self._storage, self._logger, api_name=api_name)

ghostos/framework/llms/providers.py

@@ -4,7 +4,9 @@
 from ghostos.core.llms import LLMs, LLMsConfig, PromptStorage
 from ghostos.core.messages.openai import OpenAIMessageParser
 from ghostos.framework.llms.llms import LLMsImpl
-from ghostos.framework.llms.openai_driver import OpenAIDriver, LiteLLMDriver, DeepseekDriver
+from ghostos.framework.llms.openai_driver import OpenAIDriver
+from ghostos.framework.llms.lite_llm_driver import LiteLLMDriver
+from ghostos.framework.llms.deepseek_driver import DeepseekDriver
 from ghostos.framework.llms.prompt_storage_impl import PromptStorageImpl
 from ghostos.contracts.workspace import Workspace
 from ghostos.contracts.logger import LoggerItf