Feat pyproxy direct (#692)

* 支持透传，accesstoken

* feat-pyproxy_direct

兼容了--accesstoken和透传的各种情况，修复bug

* 修复无accesstoken也可鉴权的情况

* 修复了js众测中 流式不报错的问题

* Update proxy.py

* 删除冗余注释

* Update proxy.py

* fix[trainer]: fix client trainer info not work with --taks-id / function chat call (#693)

* bug: 升级了 function call 字段的返回结构体以适配 langchain function call (#687)

* 升级了 function call 字段的返回结构体

* 格式化

* update: openai_adapter.ipynb

* feat: 浏览器增加鉴权开关 & 流式数据切分优化 (#689)

* feat: 浏览器增加鉴权开关 & 流式数据切分优化

* fix: 鉴权开关字段修改& 更新版本号

* fix: 鉴权开关字段修改& 更新版本号

---------

Co-authored-by: wangting31 <wangting31@baidu.com>

* fix: remove ipynb useless content

* fix: replace ERNIE-Bot description

* fix: add trainer cli info docs & remove ernie-bot in ipynb

* fix: remove useless or ambiguous content (#690)

* fix: remove ipynb useless content

* fix: replace ERNIE-Bot description

* fix: add trainer cli info docs & remove ernie-bot in ipynb

* fix[trainer]: cli trainer

* fix[resources]: functions passthourgh req warning

* chore: add trainer_ppl_file_tmpl.json

---------

Co-authored-by: Dobiichi-Origami <56953648+Dobiichi-Origami@users.noreply.github.com>
Co-authored-by: stonekim <shikuan@baidu.com>
Co-authored-by: wangting829 <1940087162@qq.com>
Co-authored-by: wangting31 <wangting31@baidu.com>

* feat[stat]: add star_timestamp (#685)

* feat[requestor]: add start_timestamp statis in python requestor

* chore: update version->0.4.1rc0

* chore: bump version-> 0.4.3

---------

Co-authored-by: NuODaniel <zhonghanjun@baidu.com>
Co-authored-by: Dobiichi-Origami <56953648+Dobiichi-Origami@users.noreply.github.com>
Co-authored-by: stonekim <shikuan@baidu.com>
Co-authored-by: wangting829 <1940087162@qq.com>
Co-authored-by: wangting31 <wangting31@baidu.com>

docs/trainer_ppl_file_tmpl.json

@@ -0,0 +1,57 @@
+{
+  "actions": [
+    {
+      "type": "LoadDataSetAction",
+      "datasets": {
+        "sourceType": "Platform",
+        "versions": [
+          {
+            "versionId": "ds-syeynswz2yxav8ix",
+            "samplingRate": 1 
+          }
+        ]
+      },
+      "eval_split_ratio": 10,
+      "corpus_config": {
+        "copy_data": true,
+        "config": [
+          {
+            "proportion": "1:2",
+            "corpus_type": 3,
+            "labels": [
+              "文本创作"
+            ]
+          }
+        ]
+      }
+    },
+    {
+      "type": "TrainAction",
+      "init_params": {
+        "train_mode": "SFT",
+        "train_type": "ERNIE-Speed-8K",
+        "train_config": {
+          "peft_type": "LoRA",
+          "epoch": 1,
+          "learning_rate": 0.0003,
+          "max_seq_len": 4096,
+          "logging_steps": 1,
+          "warmup_ratio": 0.1,
+          "weight_decay": 0.01,
+          "lora_rank": 8,
+          "lora_all_linear": "True"
+        },
+        "is_incr": false,
+        "job_name": "speed_math02",
+        "task_description": "task_desc1",
+        "job_description": "job_desc1"
+      }
+    },
+    {
+      "type": "ModelPublishAction"
+    }
+  ],
+  "case_init_params": {
+    "case_type": "Finetune"
+  }
+}

javascript/src/Base/index.ts

@@ -146,6 +146,7 @@ export class BaseClient {
         let fetchOptions;
         // 如果enableOauth开启， 则放开鉴权
         if (getCurrentEnvironment() === 'node' || this.enableOauth) {
+
             // 检查鉴权信息
             if (!(this.qianfanAccessKey && this.qianfanSecretKey) && !(this.qianfanAk && this.qianfanSk)) {
                 throw new Error('请设置AK/SK或QIANFAN_ACCESS_KEY/QIANFAN_SECRET_KEY');

python/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "qianfan"
-version = "0.4.2"
+version = "0.4.3"
 description = "文心千帆大模型平台 Python SDK"
 authors = []
 license = "Apache-2.0"

python/qianfan/common/client/chat.py

@@ -39,9 +39,9 @@
 )
 from qianfan.consts import DefaultLLMModel
 from qianfan.errors import InternalError
+from qianfan.resources.llm.base import BaseResourceV1
 from qianfan.resources.llm.chat_completion import (
     ChatCompletion,
-    _ChatCompletionV1,
     _ChatCompletionV2,
 )
 from qianfan.resources.typing import Literal, QfMessages, QfResponse
@@ -152,7 +152,7 @@ def _markup(s: str) -> str:
         name: str
         _client = client._real
         if self.version == "1":
-            assert isinstance(_client, _ChatCompletionV1)
+            assert isinstance(_client, BaseResourceV1)
             if _client._model is not None:
                 name = f"Model {_markup(_client._model)}"
             elif _client._endpoint is not None:

python/qianfan/common/client/main.py

@@ -32,7 +32,7 @@
     print_error_msg,
     print_info_msg,
 )
-from qianfan.config import encoding, get_config
+from qianfan.config import encoding
 from qianfan.utils.utils import check_dependency
 
 app = typer.Typer(
@@ -53,29 +53,6 @@
 _enable_traceback = False
 
 
-@app.command(name="cache")
-def clear(
-    clear: Optional[bool] = typer.Option(
-        None,
-        "--clear",
-        help="clear qianfan cache",
-    ),
-) -> None:
-    """
-    clear qianfan cache.
-    """
-    import shutil
-
-    # 要删除的目录路径
-    dir_path = get_config().CACHE_DIR
-    # 删除目录
-    try:
-        shutil.rmtree(dir_path)
-        print_info_msg(f"目录 {dir_path} 已删除")
-    except OSError as e:
-        print_info_msg(f"删除目录 {dir_path} 失败: {e}")
-
-
 @app.command(name="openai")
 @credential_required
 def openai(
@@ -206,6 +183,7 @@ def proxy(
         help="Ciphers to use (see stdlib ssl module's) [default: TLSv1]",
     ),
     access_token: str = typer.Option("", "--access-token", help="Access token"),
+    direct: bool = typer.Option(False, "--direct", help="Direct connection to server"),
 ) -> None:
     """
     Create a proxy server.
@@ -236,6 +214,7 @@ def proxy(
         mock_port=mock_port,
         ssl_config=ssl_config,
         access_token=access_token,
+        direct=direct,
     )
 
 

python/qianfan/common/client/proxy.py

@@ -43,7 +43,7 @@ async def get_stream(
 
 
 @base_app.middleware("http")
-async def base_iam(request: Request, callback: Callable) -> Response:
+async def base_openapi(request: Request, callback: Callable) -> Response:
     """
     用于向base请求中添加访问令牌。
 
@@ -54,8 +54,21 @@ async def base_iam(request: Request, callback: Callable) -> Response:
     Returns:
         Response: 处理后的响应对象。
     """
-    if "access_token" in request.url._url:
-        key = request.url._url.split("?access_token=")[1]
+    if not proxy.direct and proxy.access_token is not None:
+        try:
+            key = request.url._url.split("?access_token=")[1]
+        except Exception:
+            return JSONResponse(
+                {
+                    "error": {
+                        "message": "No ACCESS_TOKEN provided, please check",
+                        "type": "invalid_request_error",
+                        "param": None,
+                        "code": "NO_ACCESS_TOKEN",
+                    }
+                },
+                status_code=401,
+            )
         if key != proxy.access_token:
             return JSONResponse(
                 {
@@ -72,14 +85,15 @@ async def base_iam(request: Request, callback: Callable) -> Response:
             )
         else:
             new_scope = dict(request.scope)
-            new_scope["query_string"] = b""
+            if proxy._config.ACCESS_KEY and proxy._config.SECRET_KEY:
+                new_scope["query_string"] = b""
+            else:
+                proxy._direct = True
             request = StarletteRequest(scope=new_scope, receive=request.receive)
 
     else:
         pass
-
     resp = await proxy.get_response(request, DefaultValue.BaseURL)
-
     if isinstance(resp, AsyncIterator):
         return StreamingResponse(get_stream(resp), media_type="text/event-stream")
 
@@ -115,6 +129,7 @@ def entry(
     mock_port: int,
     ssl_config: Dict[str, Any],
     access_token: Optional[str],
+    direct: bool,
 ) -> None:
     import os
 
@@ -131,6 +146,8 @@ def entry(
 
     if access_token is not None:
         proxy.access_token = access_token
+    if direct:
+        proxy._direct = True
 
     proxy.mock_port = mock_port
 

python/qianfan/common/client/trainer.py

@@ -503,6 +503,7 @@ def finetune(
         if train_lora_all_linear is not None:
             trainer.train_action.train_config.lora_all_linear = train_lora_all_linear
 
+    console.log(f"trainer id: {trainer.id} is created")
     if daemon:
         trainer.start()
         console.print(
@@ -535,7 +536,7 @@ def info(
     elif task_id:
         trainers = Finetune.list()
         for t in trainers:
-            for action in t.actions:
+            for action in t.actions.values():
                 if isinstance(action, TrainAction) and action.task_id == task_id:
                     trainer = t
                     break

python/qianfan/common/client/utils.py

@@ -224,7 +224,9 @@ def credential_required(func: Callable) -> Callable:
 
     @wraps(func)
     def wrapper(*args: Any, **kwargs: Any) -> Any:
-        check_credential()
+        direct = kwargs.get("direct", False)
+        if not direct:
+            check_credential()
         return func(*args, **kwargs)
 
     return wrapper

python/qianfan/extensions/proxy/proxy.py

@@ -25,10 +25,15 @@ class ClientProxy(object):
     _retry_base_config: Optional[RetryConfig] = None
     _retry_console_config: Optional[RetryConfig] = None
     _access_token: Optional[str] = None
+    _direct: bool = False
 
     def __init__(self) -> None:
         pass
 
+    @property
+    def direct(self) -> Optional[bool]:
+        return self._direct
+
     @property
     def access_token(self) -> Optional[str]:
         return self._access_token
@@ -87,9 +92,14 @@ def _sign(self, request: QfRequest) -> None:
         url, path = request.url, urlparse(request.url).path
 
         request.url = path
-        iam_sign(str(self._config.ACCESS_KEY), str(self._config.SECRET_KEY), request)
+        if not (request.query.get("client_id") or request.query.get("client_secret")):
+            iam_sign(
+                str(self._config.ACCESS_KEY), str(self._config.SECRET_KEY), request
+            )
         request.url = url
         if not request.headers.get("Authorization", None):
+            self._auth._ak = request.query.get("client_id")
+            self._auth._sk = request.query.get("client_secret")
             request.query["access_token"] = self._auth.access_token()
 
     async def get_request(self, request: Request, url_route: str) -> QfRequest:
@@ -112,9 +122,9 @@ async def get_request(self, request: Request, url_route: str) -> QfRequest:
         # 获取请求头
         if self.mock_port != -1:
             url_route = f"http://127.0.0.1:{self.mock_port}"
-
         url = url_route + request.url.path
         host = urlparse(url_route).netloc
+
         headers = {
             "Content-Type": "application/json",
             "Host": host,
@@ -124,7 +134,7 @@ async def get_request(self, request: Request, url_route: str) -> QfRequest:
         json_body = await request.json()
         return QfRequest(
             url=url,
-            headers=headers,
+            headers=headers if not self._direct else dict(request.headers),
             method=request.method,
             query=dict(request.query_params),
             json_body=json_body,
@@ -149,11 +159,23 @@ async def get_response(
         try:
             async with self._rate_limiter:
                 qf_req = await self.get_request(request, url_route)
-                self._sign(qf_req)
+                if self._direct:
+                    pass
+                else:
+                    self._sign(qf_req)
                 logging.debug(f"request: {qf_req}")
-
                 if qf_req.json_body.get("stream", False):
-                    return self._client.arequest_stream(qf_req)
+                    resp, session = await self._client.arequest(qf_req)
+                    if (
+                        "Content-Type" in resp.headers
+                        and "application/json" in resp.headers["Content-Type"]
+                    ):  # 判断返回中是否有流式数据
+                        resp, session = await self._client.arequest(qf_req)
+                        async with session:
+                            json_body = await resp.json()
+                        return json_body
+                    else:
+                        return self._client.arequest_stream(qf_req)
                 else:
                     resp, session = await self._client.arequest(qf_req)
                     async with session:

python/qianfan/model/model.py

@@ -87,12 +87,6 @@ def __init__(
         self.name = name
         if id is None or set_id is None:
             self.auto_complete_info()
-        if (
-            (id is None and set_id is None)
-            or self.task_id is None
-            or self.job_id is None
-        ):
-            log_warn("set_id/id or job_id/task_id should be provided")
 
     def exec(
         self, input: Optional[Dict] = None, **kwargs: Dict

python/qianfan/resources/console/consts.py

@@ -140,10 +140,12 @@ class TrainStatus(str, Enum):
     """训练完成"""
     Running = "Running"
     """训练进行中"""
-    Fail = "Fail"
+    Fail = "Failed"
     """训练失败"""
     Stop = "Stopped"
     """训练停止"""
+    Waiting = "Waiting"
+    """排队中"""
 
 
 class ModelState(str, Enum):

python/qianfan/resources/llm/function.py

@@ -114,7 +114,7 @@ def _default_model(cls) -> str:
     def do(
         self,
         messages: Union[List[Dict], QfMessages],
-        functions: List[Dict],
+        functions: List[Dict] = [],
         **kwargs: Any,
     ) -> Union[QfResponse, Iterator[QfResponse]]:
         """
@@ -152,18 +152,29 @@ def do(
         ```
 
         """
-        if len(functions) <= 0:
-            raise errors.InvalidArgumentError(
-                "functions should be a list of functions, "
-                "each function is a dictionary with name and description."
-            )
-        if kwargs.get("stream") is True:
-            raise errors.InvalidArgumentError("Function does not support stream mode.")
-
         if isinstance(messages, QfMessages):
             temp_messages = messages._to_list()
         else:
             temp_messages = messages
+        for k in [
+            "auto_concat_truncate",
+            "truncated_continue_prompt",
+            "truncate_overlong_msgs",
+        ]:
+            if k in kwargs:
+                del kwargs[k]
+
+        for k in ["request_id"]:
+            if k in kwargs and kwargs.get(k) is None:
+                del kwargs[k]
+
+        if not functions:
+            # 没有传入functions，不特殊处理，直接走普通的base_resource请求模式
+            kwargs["messages"] = temp_messages
+            return super()._do(**kwargs)
+
+        if kwargs.get("stream") is True:
+            raise errors.InvalidArgumentError("Function does not support stream mode.")
 
         functions_schemas = self._render_functions_prompt(functions)
         temp_messages[0] = self._render_user_query_msg(