【功能新增】在线 LLM 模型支持阿里云通义千问 (#1534)

* feat: add qwen-api

* 使Qwen API支持temperature参数；添加测试用例

* 将online-api的sdk列为可选依赖

---------

Co-authored-by: liunux4odoo <liunux@qq.com>

README.md

@@ -133,6 +133,7 @@ docker run -d --gpus all -p 80:8501 registry.cn-beijing.aliyuncs.com/chatchat/ch
 - [MiniMax](https://api.minimax.chat)
 - [讯飞星火](https://xinghuo.xfyun.cn)
 - [百度千帆](https://cloud.baidu.com/product/wenxinworkshop?track=dingbutonglan)
+- [阿里云通义千问](https://dashscope.aliyun.com/)
 
 项目中默认使用的 LLM 类型为 `THUDM/chatglm2-6b`，如需使用其他 LLM 类型，请在 [configs/model_config.py] 中对 `llm_model_dict` 和 `LLM_MODEL` 进行修改。
 

configs/model_config.py.example

@@ -115,7 +115,7 @@ ONLINE_LLM_MODEL = {
         "secret_key": "",
         "provider": "QianFanWorker",
     },
-    # 火山方舟 API
+    # 火山方舟 API，文档参考 https://www.volcengine.com/docs/82379
     "fangzhou-api": {
         "version": "chatglm-6b-model",  # 当前支持 "chatglm-6b-model"， 更多的见文档模型支持列表中方舟部分。
         "version_url": "",  # 可以不填写version，直接填写在方舟申请模型发布的API地址
@@ -123,6 +123,12 @@ ONLINE_LLM_MODEL = {
         "secret_key": "",
         "provider": "FangZhouWorker",
     },
+    # 阿里云通义千问 API，文档参考 https://help.aliyun.com/zh/dashscope/developer-reference/api-details
+    "qwen-api": {
+        "version": "qwen-turbo",  # 可选包括 "qwen-turbo", "qwen-plus"
+        "api_key": "",  # 请在阿里云控制台模型服务灵积API-KEY管理页面创建
+        "provider": "QwenWorker",
+    },
 }
 
 

configs/server_config.py.example

@@ -79,6 +79,9 @@ FSCHAT_MODEL_WORKERS = {
     "fangzhou-api": {
         "port": 21005,
     },
+    "qwen-api": {
+        "port": 21006,
+    },
 }
 
 # fastchat multi model worker server

requirements.txt

@@ -24,6 +24,12 @@ pytest
 scikit-learn
 numexpr
 
+# online api libs
+# zhipuai
+# dashscope>=1.10.0 # qwen
+# qianfan
+# volcengine>=1.0.106 # fangzhou
+
 # uncomment libs if you want to use corresponding vector store
 # pymilvus==2.1.3 # requires milvus==2.1.3
 # psycopg2

requirements_api.txt

@@ -24,6 +24,12 @@ pytest
 scikit-learn
 numexpr
 
+# online api libs
+# zhipuai
+# dashscope>=1.10.0 # qwen
+# qianfan
+# volcengine>=1.0.106 # fangzhou
+
 # uncomment libs if you want to use corresponding vector store
 # pymilvus==2.1.3 # requires milvus==2.1.3
 # psycopg2

server/model_workers/__init__.py

@@ -3,3 +3,4 @@ from .minimax import MiniMaxWorker
 from .xinghuo import XingHuoWorker
 from .qianfan import QianFanWorker
 from .fangzhou import FangZhouWorker
+from .qwen import QwenWorker

server/model_workers/base.py

@@ -92,5 +92,5 @@ class ApiModelWorker(BaseModelWorker):
                 if content := msg[len(ai_start):].strip():
                     result.append({"role": ai_role, "content": content})
             else:
-                raise RuntimeError(f"unknow role in msg: {msg}")
+                raise RuntimeError(f"unknown role in msg: {msg}")
         return result

server/model_workers/qwen.py

@@ -0,0 +1,123 @@
+import json
+import sys
+from configs import TEMPERATURE
+from http import HTTPStatus
+from typing import List, Literal, Dict
+
+from fastchat import conversation as conv
+
+from server.model_workers.base import ApiModelWorker
+from server.utils import get_model_worker_config
+
+
+def request_qwen_api(
+    messages: List[Dict[str, str]],
+    api_key: str = None,
+    version: str = "qwen-turbo",
+    temperature: float = TEMPERATURE,
+    model_name: str = "qwen-api",
+):
+    import dashscope
+
+    config = get_model_worker_config(model_name)
+    api_key = api_key or config.get("api_key")
+    version = version or config.get("version")
+
+    gen = dashscope.Generation()
+    responses = gen.call(
+        model=version,
+        temperature=temperature,
+        api_key=api_key,
+        messages=messages,
+        result_format='message',  # set the result is message format.
+        stream=True,
+    )
+
+    text = ""
+    for resp in responses:
+        if resp.status_code != HTTPStatus.OK:
+            yield {
+                "code": resp.status_code,
+                "text": "api not response correctly",
+            }
+
+        if resp["status_code"] == 200:
+            if choices := resp["output"]["choices"]:
+                yield {
+                    "code": 200,
+                    "text": choices[0]["message"]["content"],
+                }
+        else:
+            yield {
+                "code": resp["status_code"],
+                "text": resp["message"],
+            }
+
+
+class QwenWorker(ApiModelWorker):
+    def __init__(
+            self,
+            *,
+            version: Literal["qwen-turbo", "qwen-plus"] = "qwen-turbo",
+            model_names: List[str] = ["qwen-api"],
+            controller_addr: str,
+            worker_addr: str,
+            **kwargs,
+    ):
+        kwargs.update(model_names=model_names, controller_addr=controller_addr, worker_addr=worker_addr)
+        kwargs.setdefault("context_len", 16384)
+        super().__init__(**kwargs)
+
+        # TODO: 确认模板是否需要修改
+        self.conv = conv.Conversation(
+            name=self.model_names[0],
+            system_message="你是一个聪明、对人类有帮助的人工智能，你可以对人类提出的问题给出有用、详细、礼貌的回答。",
+            messages=[],
+            roles=["user", "assistant", "system"],
+            sep="\n### ",
+            stop_str="###",
+        )
+        config = self.get_config()
+        self.api_key = config.get("api_key")
+        self.version = version
+
+    def generate_stream_gate(self, params):
+        messages = self.prompt_to_messages(params["prompt"])
+
+        for resp in request_qwen_api(messages=messages,
+                                     api_key=self.api_key,
+                                     version=self.version,
+                                     temperature=params.get("temperature")):
+            if resp["code"] == 200:
+                yield json.dumps({
+                    "error_code": 0,
+                    "text": resp["text"]
+                },
+                    ensure_ascii=False
+                ).encode() + b"\0"
+            else:
+                yield json.dumps({
+                    "error_code": resp["code"],
+                    "text": resp["text"]
+                },
+                    ensure_ascii=False
+                ).encode() + b"\0"
+
+    def get_embeddings(self, params):
+        # TODO: 支持embeddings
+        print("embedding")
+        print(params)
+
+
+if __name__ == "__main__":
+    import uvicorn
+    from server.utils import MakeFastAPIOffline
+    from fastchat.serve.model_worker import app
+
+    worker = QwenWorker(
+        controller_addr="http://127.0.0.1:20001",
+        worker_addr="http://127.0.0.1:20007",
+    )
+    sys.modules["fastchat.serve.model_worker"].worker = worker
+    MakeFastAPIOffline(app)
+    uvicorn.run(app, port=20007)

tests/online_api/test_qwen.py

@@ -0,0 +1,19 @@
+import sys
+from pathlib import Path
+root_path = Path(__file__).parent.parent.parent
+sys.path.append(str(root_path))
+
+from server.model_workers.qwen import request_qwen_api
+from pprint import pprint
+import pytest
+
+
+@pytest.mark.parametrize("version", ["qwen-turbo"])
+def test_qwen(version):
+    messages = [{"role": "user", "content": "hello"}]
+    print("\n" + version + "\n")
+
+    for x in request_qwen_api(messages, version=version):
+        print(type(x))
+        pprint(x)
+        assert x["code"] == 200