feat[resources]: 支持chat_completion调用ERNIE-Functions-8K

python/pyproject.toml

@@ -22,15 +22,15 @@ aiolimiter = ">=1.1.0"
 importlib-metadata = { version = ">=1.4.0", python = "<=3.7" }
 bce-python-sdk = ">=0.8.79"
 typing-extensions = { version = ">=4.0.0", python = "<=3.10" }
-pydantic = "*"
+pydantic = ">=1.0"
 python-dotenv = [
     { version = "<=0.21.1", python = "<3.8" },
     { version = ">=1.0", python = ">=3.8" }
 ]
 tenacity = "^8.2.3"
-multiprocess = "*"
+multiprocess = ">=0.70.12"
 langchain = { version = ">=0.1.10", python = ">=3.8.1", optional = true }
-langchain-community = { version = ">=0.2.0", python = ">=3.8.1", optional = true}
+langchain-community = { version = ">=0.2.0", python = ">=3.8.1", optional = true }
 numpy = [
     { version = "<1.22.0", python = ">=3.7 <3.8", optional = true },
     { version = ">=1.22.0", python = ">=3.8", optional = true }
@@ -39,26 +39,27 @@ pyarrow = [
     { version = ">=14.0.1", python = ">=3.8", optional = true },
     { version = "<=12.0.1", python = ">=3.7 <3.8", optional = true }
 ]
-locust = { version = "*", optional = true }
-tabulate = { version = "*", optional = true }
+locust = { version = ">=2.9.0", optional = true }
+tabulate = { version = ">=0.9.0", optional = true }
 python-dateutil = { version = "^2.8.2", optional = true }
 rich = ">=13.0.0"
 typer = ">=0.9.0"
 pyyaml = "^6.0.1"
 prompt-toolkit = ">=3.0.38"
 torch = [
-    { version = "<=1.13.1", python = "<3.8", optional = true},
+    { version = "<=1.13.1", python = "<3.8", optional = true },
     { version = ">=1.4.0", python = ">=3.8", optional = true }
 ]
 ltp = { version = ">=4.2.0", optional = true }
 emoji = { version = ">=2.2.0", optional = true }
 sentencepiece = { version = ">=0.1.98", optional = true }
-diskcache = "^5.6.3"
+diskcache = ">=5.6.3"
+cachetools = ">=5.0.0"
 
-ijson = { version = "*", optional = true }
-fastapi = { version = "*", optional = true }
-uvicorn = { version = "*", optional = true }
-filelock = { version = "*", optional = true}
+ijson = { version = ">=3.0", optional = true }
+fastapi = { version = ">=0.85.0", optional = true }
+uvicorn = { version = ">=0.15.0", optional = true }
+filelock = { version = ">=3.7.0", optional = true }
 
 [tool.poetry.scripts]
 qianfan = "qianfan.common.client.main:main"

python/qianfan/common/client/trainer.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 
+import json
 import time
 from typing import Any, Callable, Dict, List, Optional
 
@@ -41,7 +42,7 @@
 from qianfan.model.configs import DeployConfig
 from qianfan.model.consts import ServiceType
 from qianfan.resources.console.consts import DeployPoolType, FinetuneSupportModelType
-from qianfan.trainer import DPO, LLMFinetune, PostPreTrain
+from qianfan.trainer import DPO, Finetune, PostPreTrain, Trainer
 from qianfan.trainer.actions import (
     DeployAction,
     EvaluateAction,
@@ -220,7 +221,7 @@ def list_train_type(
         if cmd == "postpretrain":
             model_list = PostPreTrain.train_type_list()
         elif cmd in ["finetune", "run"]:
-            model_list = LLMFinetune.train_type_list()
+            model_list = Finetune.train_type_list()
         elif cmd in ["dpo"]:
             model_list = DPO.train_type_list()
         else:
@@ -306,7 +307,7 @@ def show_config_limit(
         if cmd == "postpretrain":
             model_list = PostPreTrain.train_type_list()
         elif cmd in ["finetune", "run"]:
-            model_list = LLMFinetune.train_type_list()
+            model_list = Finetune.train_type_list()
         elif cmd in ["dpo"]:
             model_list = DPO.train_type_list()
         else:
@@ -366,6 +367,9 @@ def finetune(
     list_train_type: Optional[bool] = list_train_type_option,
     show_config_limit: Optional[str] = typer.Option(
         None,
+        "--show-config-limit",
+        "--show",
+        "-s",
         callback=show_config_limit,
         is_eager=True,
         help="Show config limit for specified train type.",
@@ -445,7 +449,7 @@ def finetune(
     callback = MyEventHandler(console=console)
 
     if trainer_pipeline_file is not None:
-        trainer = LLMFinetune.load(file=trainer_pipeline_file)
+        trainer = Finetune.load(file=trainer_pipeline_file)
         trainer.register_event_handler(callback)
     else:
         ds = None
@@ -465,7 +469,7 @@ def finetune(
                 pool_type=DeployPoolType[deploy_pool_type],
                 service_type=ServiceType[deploy_service_type],
             )
-        trainer = LLMFinetune(
+        trainer = Finetune(
             dataset=ds,
             train_type=train_type,
             event_handler=callback,
@@ -515,6 +519,36 @@ def finetune(
     time.sleep(0.1)
 
 
+@trainer_app.command()
+@credential_required
+def info(
+    trainer_id: Optional[str] = typer.Option(None, help="trainer id"),
+    task_id: Optional[str] = typer.Option(None, help="task id"),
+) -> None:
+    """
+    get a trainer info.
+    """
+    console = replace_logger_handler()
+    trainer: Optional[Trainer] = None
+    if trainer_id:
+        trainer = Finetune.load(id=trainer_id)
+    elif task_id:
+        trainers = Finetune.list()
+        for t in trainers:
+            for action in t.actions:
+                if isinstance(action, TrainAction) and action.task_id == task_id:
+                    trainer = t
+                    break
+    else:
+        console.log("Must provide either trainer id or task id.")
+    if trainer:
+        json_str = json.dumps(trainer.info(), ensure_ascii=False, indent=2)
+        print(json_str)
+
+    # wait a second for the log to be flushed
+    time.sleep(0.1)
+
+
 @trainer_app.command()
 @credential_required
 def postpretrain(
@@ -665,11 +699,6 @@ def postpretrain(
     time.sleep(0.1)
 
 
-@trainer_app.command(
-    "run",
-    deprecated=True,
-    help="Run a dpo trainer task.",
-)
 @trainer_app.command()
 @credential_required
 def dpo(
@@ -768,7 +797,7 @@ def dpo(
     callback = MyEventHandler(console=console)
 
     if trainer_pipeline_file is not None:
-        trainer = LLMFinetune.load(file=trainer_pipeline_file)
+        trainer = Finetune.load(file=trainer_pipeline_file)
         trainer.register_event_handler(callback)
     else:
         ds = None
@@ -788,7 +817,7 @@ def dpo(
                 pool_type=DeployPoolType[deploy_pool_type],
                 service_type=ServiceType[deploy_service_type],
             )
-        trainer = LLMFinetune(
+        trainer = Finetune(
             dataset=ds,
             train_type=train_type,
             event_handler=callback,

python/qianfan/model/model.py

@@ -386,13 +386,13 @@ def compress(
                 comp_task_detail_resp["result"]["status"]
                 == console_const.ModelCompTaskStatus.Succeeded.value
             ):
-                new_model_version_id = comp_task_detail_resp["result"].get("modelId")
+                new_model_id = comp_task_detail_resp["result"].get("modelId")
                 log_info(
                     f"compress task {model_comp_task_id} run with status"
                     f" {comp_task_detail_resp['result']['status']}"
-                    f" new model_version_id: {new_model_version_id}"
+                    f" new model_id: {new_model_id}"
                 )
-                new_model = Model(id=new_model_version_id)
+                new_model = Model(id=new_model_id)
                 new_model.auto_complete_info()
                 return new_model
             else:

python/qianfan/resources/llm/base.py

@@ -163,11 +163,11 @@ def __init__(
         self, version: Optional[Literal["1", "2", 1, 2]] = None, **kwargs: Any
     ) -> None:
         self._version = str(version) if version else "1"
-        self._real = self._real_base(self._version)(**kwargs)
-        self._backup = self._real_base("1")(**kwargs)
+        self._real = self._real_base(self._version, **kwargs)(**kwargs)
+        self._backup = self._real_base("1", **kwargs)(**kwargs)
 
     @classmethod
-    def _real_base(cls, version: str) -> Type[BaseResource]:
+    def _real_base(cls, version: str, **kwargs: Any) -> Type[BaseResource]:
         """
         return the real base class
         """

python/qianfan/resources/llm/chat_completion.py

@@ -34,6 +34,7 @@
     BatchRequestFuture,
     VersionBase,
 )
+from qianfan.resources.llm.function import Function
 from qianfan.resources.tools.tokenizer import Tokenizer
 from qianfan.resources.typing import JsonBody, QfLLMInfo, QfMessages, QfResponse, QfRole
 from qianfan.utils.logging import log_error, log_info
@@ -1537,7 +1538,24 @@ class ChatCompletion(VersionBase):
     _real: Union[_ChatCompletionV1, _ChatCompletionV2]
 
     @classmethod
-    def _real_base(cls, version: str) -> Type:
+    def _real_base(cls, version: str, **kwargs: Any) -> Type:
+        # convert to qianfan.Function
+        if kwargs.get("use_function"):
+            return Function
+        else:
+            model = kwargs.get("model", "")
+            func_model_info_list = {
+                k.lower(): v for k, v in Function._supported_models().items()
+            }
+            func_model_info = func_model_info_list.get(model.lower())
+            if model and func_model_info:
+                if func_model_info and func_model_info.endpoint:
+                    return Function
+            endpoint = kwargs.get("endpoint", "")
+            for m in func_model_info_list.values():
+                if endpoint and m.endpoint == endpoint:
+                    return Function
+
         if version == "1":
             return _ChatCompletionV1
         elif version == "2":

python/qianfan/resources/llm/function.py

@@ -80,6 +80,8 @@ def _supported_models(cls) -> Dict[str, QfLLMInfo]:
                     "user_id",
                     "stop",
                     "max_output_tokens",
+                    "enable_user_memory",
+                    "user_memory_extract_level",
                 },
                 max_input_chars=11200,
                 max_input_tokens=7168,

python/qianfan/tests/trainer_test.py

@@ -139,7 +139,7 @@ def test_trainer_sft_run():
 
     eh = MyEventHandler()
     sft_task = LLMFinetune(
-        train_type="ERNIE-Bot-turbo-0725",
+        train_type="ERNIE-Speed-8K",
         dataset=ds,
         train_config=train_config,
         event_handler=eh,
@@ -157,11 +157,11 @@ def test_trainer_sft_run():
 def test_trainer_sft_run_from_bos():
     with pytest.raises(InvalidArgumentError):
         sft_task = LLMFinetune(
-            train_type="ERNIE-Bot-turbo-0725",
+            train_type="ERNIE-Speed-8K",
         )
         sft_task.run()
     sft_task = LLMFinetune(
-        train_type="ERNIE-Bot-turbo-0725", dataset_bos_path="bos:/sdk-test/"
+        train_type="ERNIE-Speed-8K", dataset_bos_path="bos:/sdk-test/"
     )
     sft_task.run()
     res = sft_task.result
@@ -235,7 +235,7 @@ def test_trainer_resume():
     ds = Dataset.load(source=qianfan_data_source, organize_data_as_group=True)
 
     sft_task = LLMFinetune(
-        train_type="ERNIE-Bot-turbo-0725",
+        train_type="ERNIE-Speed-8K",
         dataset=ds,
     )
     ppl = sft_task.ppls[0]