feat(model): Support DeepSeek proxy LLM (#1491)

dbgpt/model/adapter/proxy_adapter.py

@@ -294,6 +294,31 @@ class MoonshotProxyLLMModelAdapter(ProxyLLMModelAdapter):
         return moonshot_generate_stream
 
 
+class DeepseekProxyLLMModelAdapter(ProxyLLMModelAdapter):
+    """Deepseek proxy LLM model adapter.
+
+    See Also: `Deepseek Documentation <https://platform.deepseek.com/api-docs/>`_
+    """
+
+    def support_async(self) -> bool:
+        return True
+
+    def do_match(self, lower_model_name_or_path: Optional[str] = None):
+        return lower_model_name_or_path == "deepseek_proxyllm"
+
+    def get_llm_client_class(
+        self, params: ProxyModelParameters
+    ) -> Type[ProxyLLMClient]:
+        from dbgpt.model.proxy.llms.deepseek import DeepseekLLMClient
+
+        return DeepseekLLMClient
+
+    def get_async_generate_stream_function(self, model, model_path: str):
+        from dbgpt.model.proxy.llms.deepseek import deepseek_generate_stream
+
+        return deepseek_generate_stream
+
+
 register_model_adapter(OpenAIProxyLLMModelAdapter)
 register_model_adapter(TongyiProxyLLMModelAdapter)
 register_model_adapter(OllamaLLMModelAdapter)
@@ -305,3 +330,4 @@ register_model_adapter(BardProxyLLMModelAdapter)
 register_model_adapter(BaichuanProxyLLMModelAdapter)
 register_model_adapter(YiProxyLLMModelAdapter)
 register_model_adapter(MoonshotProxyLLMModelAdapter)
+register_model_adapter(DeepseekProxyLLMModelAdapter)

dbgpt/model/proxy/__init__.py

@@ -12,6 +12,7 @@ def __lazy_import(name):
         "YiLLMClient": "dbgpt.model.proxy.llms.yi",
         "MoonshotLLMClient": "dbgpt.model.proxy.llms.moonshot",
         "OllamaLLMClient": "dbgpt.model.proxy.llms.ollama",
+        "DeepseekLLMClient": "dbgpt.model.proxy.llms.deepseek",
     }
 
     if name in module_path:
@@ -35,4 +36,5 @@ __all__ = [
     "YiLLMClient",
     "MoonshotLLMClient",
     "OllamaLLMClient",
+    "DeepseekLLMClient",
 ]

dbgpt/model/proxy/llms/deepseek.py

@@ -0,0 +1,104 @@
+import os
+from typing import TYPE_CHECKING, Any, Dict, Optional, Union, cast
+
+from dbgpt.core import ModelRequest, ModelRequestContext
+from dbgpt.model.proxy.llms.proxy_model import ProxyModel
+
+from .chatgpt import OpenAILLMClient
+
+if TYPE_CHECKING:
+    from httpx._types import ProxiesTypes
+    from openai import AsyncAzureOpenAI, AsyncOpenAI
+
+    ClientType = Union[AsyncAzureOpenAI, AsyncOpenAI]
+
+# 32K model
+_DEFAULT_MODEL = "deepseek-chat"
+
+
+async def deepseek_generate_stream(
+    model: ProxyModel, tokenizer, params, device, context_len=2048
+):
+    client: DeepseekLLMClient = cast(DeepseekLLMClient, model.proxy_llm_client)
+    context = ModelRequestContext(stream=True, user_name=params.get("user_name"))
+    request = ModelRequest.build_request(
+        client.default_model,
+        messages=params["messages"],
+        temperature=params.get("temperature"),
+        context=context,
+        max_new_tokens=params.get("max_new_tokens"),
+    )
+    async for r in client.generate_stream(request):
+        yield r
+
+
+class DeepseekLLMClient(OpenAILLMClient):
+    """Deepseek LLM Client.
+
+    Deepseek's API is compatible with OpenAI's API, so we inherit from OpenAILLMClient.
+
+    API Reference: https://platform.deepseek.com/api-docs/
+    """
+
+    def __init__(
+        self,
+        api_key: Optional[str] = None,
+        api_base: Optional[str] = None,
+        api_type: Optional[str] = None,
+        api_version: Optional[str] = None,
+        model: Optional[str] = _DEFAULT_MODEL,
+        proxies: Optional["ProxiesTypes"] = None,
+        timeout: Optional[int] = 240,
+        model_alias: Optional[str] = "deepseek_proxyllm",
+        context_length: Optional[int] = None,
+        openai_client: Optional["ClientType"] = None,
+        openai_kwargs: Optional[Dict[str, Any]] = None,
+        **kwargs,
+    ):
+        api_base = (
+            api_base or os.getenv("DEEPSEEK_API_BASE") or "https://api.deepseek.com/v1"
+        )
+        api_key = api_key or os.getenv("DEEPSEEK_API_KEY")
+        model = model or _DEFAULT_MODEL
+        if not context_length:
+            if "deepseek-chat" in model:
+                context_length = 1024 * 32
+            elif "deepseek-coder" in model:
+                context_length = 1024 * 16
+            else:
+                # 8k
+                context_length = 1024 * 8
+
+        if not api_key:
+            raise ValueError(
+                "Deepseek API key is required, please set 'DEEPSEEK_API_KEY' in "
+                "environment variable or pass it to the client."
+            )
+        super().__init__(
+            api_key=api_key,
+            api_base=api_base,
+            api_type=api_type,
+            api_version=api_version,
+            model=model,
+            proxies=proxies,
+            timeout=timeout,
+            model_alias=model_alias,
+            context_length=context_length,
+            openai_client=openai_client,
+            openai_kwargs=openai_kwargs,
+            **kwargs,
+        )
+
+    def check_sdk_version(self, version: str) -> None:
+        if not version >= "1.0":
+            raise ValueError(
+                "Deepseek API requires openai>=1.0, please upgrade it by "
+                "`pip install --upgrade 'openai>=1.0'`"
+            )
+
+    @property
+    def default_model(self) -> str:
+        model = self._model
+        if not model:
+            model = _DEFAULT_MODEL
+        return model