[do not merge]Refactor/unittest (#264)

* update unittest

* update requirements

* update requirements

* update requirements

* update requirements

* update requirements

* update requirements

* update requirements

* update requirements

* update requirements

* update requirements

* update requirements

* update requirements

* update ci

* update ci

* update ci

* update ci

* update ci

* update ci

* update ci

* update ci

* update ci

* update ci

* merge gradio4

* update ci

* update ci

* update ci

* update ci

* pass unit test

* merge and pass unit test

---------

Co-authored-by: Zhicheng Zhang <zhangzhicheng.zzc@alibaba-inc.com>

Author: zzhangpurdue

.dev_scripts/dockerci.sh

@@ -0,0 +1,14 @@
+#!/bin/bash
+
+
+# install dependencies for ci
+pip install torch
+export CODE_INTERPRETER_WORK_DIR=${GITHUB_WORKSPACE}
+echo "${CODE_INTERPRETER_WORK_DIR}"
+
+# cp file
+cp tests/samples/luoli15.jpg "${CODE_INTERPRETER_WORK_DIR}/luoli15.jpg"
+ls  "${CODE_INTERPRETER_WORK_DIR}"
+
+# run ci
+pytest

.github/workflows/citest.yaml

@@ -40,7 +40,8 @@ jobs:
   unittest:
     # The type of runner that the job will run on
     runs-on: ubuntu-latest
-    timeout-minutes: 240
+    timeout-minutes: 20
+    environment: testci
     steps:
     - uses: actions/checkout@v3
 
@@ -51,14 +52,21 @@ jobs:
         path: ~/.cache/pip
         key: ${{ runner.os }}-pip-${{ hashFiles('requirements.txt') }}
 
-    - name: Set up Python 3.8
+    - name: Set up Python 3.10.13
       uses: actions/setup-python@v3
       with:
-        python-version: 3.8
+        python-version: "3.10.13"
 
     - name: Install dependencies
       if: steps.cache.outputs.cache-hit != 'true'
       run: pip install -r requirements.txt
 
     - name: Run tests
-      run: pytest
+      env:
+        AMAP_TOKEN: ${{ secrets.AMAP_TOKEN }}
+        BING_SEARCH_V7_SUBSCRIPTION_KEY: ${{ secrets.BING_SEARCH_V7_SUBSCRIPTION_KEY }}
+        DASHSCOPE_API_KEY: ${{ secrets.DASHSCOPE_API_KEY }}
+        MODELSCOPE_API_TOKEN: ${{ secrets.MODELSCOPE_API_TOKEN }}
+
+      shell: bash
+      run: bash .dev_scripts/dockerci.sh

modelscope_agent/agent.py

@@ -86,12 +86,20 @@ def _register_tool(self, tool: Union[str, Dict]):
         """
         Instantiate the global tool for the agent
 
+        Args:
+            tool: the tool should be either in a string format with name as value
+            and in a dict format, example
+            (1) When str: amap_weather
+            (2) When dict: {'amap_weather': {'token': 'xxx'}}
+
+        Returns:
+
         """
         tool_name = tool
         tool_cfg = {}
         if isinstance(tool, Dict):
-            tool_name = tool['name']
-            tool_cfg = tool
+            tool_name = next(iter(tool))
+            tool_cfg = tool[tool_name]
         if tool_name not in TOOL_REGISTRY:
             raise NotImplementedError
         if tool not in self.function_list:

modelscope_agent/llm/__init__.py

@@ -2,10 +2,8 @@
 
 from .base import LLM_REGISTRY, BaseChatModel
 from .custom import CustomLLM
-from .dashscope import DashScopeLLM
-from .dashscope_qwen import QwenChatAtDS
-from .modelscope import ModelScopeLLM
-from .modelscope_chatglm import ModelScopeChatGLM
+from .dashscope import DashScopeLLM, QwenChatAtDS
+from .modelscope import ModelScopeChatGLM, ModelScopeLLM
 from .openai import OpenAi
 
 

modelscope_agent/llm/base.py

@@ -43,7 +43,9 @@ def chat(self,
             assert isinstance(prompt, str)
             messages = [{'role': 'user', 'content': prompt}]
         else:
-            assert prompt is None, 'Do not pass agents and messages at the same time.'
+            assert prompt is None, 'Do not pass prompt and messages at the same time.'
+
+        assert len(messages) > 0, 'messages list must not be empty'
 
         if stream:
             return self._chat_stream(messages, stop=stop, **kwargs)

modelscope_agent/llm/dashscope.py

@@ -96,3 +96,64 @@ def _chat_no_stream(self,
                 response.message,
             )
             return err
+
+
+@register_llm('dashscope_qwen')
+class QwenChatAtDS(DashScopeLLM):
+    """
+    qwen_model from dashscope
+    """
+
+    def chat_with_raw_prompt(self,
+                             prompt: str,
+                             stop: Optional[List[str]] = None,
+                             **kwargs) -> str:
+        if prompt == '':
+            return ''
+        stop = stop or []
+        top_p = kwargs.get('top_p', 0.8)
+
+        response = dashscope.Generation.call(
+            self.model,
+            prompt=prompt,  # noqa
+            stop_words=[{
+                'stop_str': word,
+                'mode': 'exclude'
+            } for word in stop],
+            top_p=top_p,
+            result_format='message',
+            stream=False,
+            use_raw_prompt=True,
+        )
+        if response.status_code == HTTPStatus.OK:
+            # with open('debug.json', 'w', encoding='utf-8') as writer:
+            #     writer.write(json.dumps(response, ensure_ascii=False))
+            return response.output.choices[0].message.content
+        else:
+            err = 'Error code: %s, error message: %s' % (
+                response.code,
+                response.message,
+            )
+            return err
+
+    def build_raw_prompt(self, messages):
+        im_start = '<|im_start|>'
+        im_end = '<|im_end|>'
+        if messages[0]['role'] == 'system':
+            sys = messages[0]['content']
+            prompt = f'{im_start}system\n{sys}{im_end}'
+        else:
+            prompt = f'{im_start}system\nYou are a helpful assistant.{im_end}'
+
+        for message in messages:
+            if message['role'] == 'user':
+                query = message['content'].lstrip('\n').rstrip()
+                prompt += f'\n{im_start}user\n{query}{im_end}'
+            elif message['role'] == 'assistant':
+                response = message['content'].lstrip('\n').rstrip()
+                prompt += f'\n{im_start}assistant\n{response}{im_end}'
+
+        # add one empty reply for the last round of assistant
+        assert prompt.endswith(f'\n{im_start}assistant\n{im_end}')
+        prompt = prompt[:-len(f'{im_end}')]
+        return prompt

modelscope_agent/llm/dashscope_qwen.py

@@ -2,11 +2,6 @@
 import sys
 from typing import Dict, Iterator, List, Optional
 
-import torch
-from swift import Swift
-from transformers import AutoModelForCausalLM, AutoTokenizer
-
-from modelscope import GenerationConfig, snapshot_download
 from .base import BaseChatModel, register_llm
 
 
@@ -18,6 +13,15 @@ class ModelScopeLLM(BaseChatModel):
 
     def __init__(self, model: str, model_server: str, **kwargs):
         super().__init__(model, model_server)
+        try:
+            import torch
+        except ImportError:
+            raise ImportError(
+                'Please install torch first: `pip install torch` or refer https://pytorch.org/ '
+            )
+
+        from modelscope import (AutoModelForCausalLM, AutoTokenizer,
+                                GenerationConfig, snapshot_download)
 
         # Download model based on model version
         self.model_version = kwargs.get('model_version', None)
@@ -59,6 +63,11 @@ def __init__(self, model: str, model_server: str, **kwargs):
             self.model_dir, trust_remote_code=True)
 
         if self.use_lora:
+            try:
+                from swift import Swift
+            except ImportError:
+                raise ImportError(
+                    'Please install swift first: `pip install ms-swift`')
             self.load_from_lora()
 
         if self.use_raw_generation_config:
@@ -112,3 +121,30 @@ def _inference(self, prompt: str) -> str:
         response = self.tokenizer.decode(result)
 
         return response
+
+
+@register_llm('modelscope_chatglm')
+class ModelScopeChatGLM(ModelScopeLLM):
+
+    def _inference(self, prompt: str) -> str:
+        device = self.model.device
+        input_ids = self.tokenizer(
+            prompt, return_tensors='pt').input_ids.to(device)
+        input_len = input_ids.shape[1]
+
+        eos_token_id = [
+            self.tokenizer.eos_token_id,
+            self.tokenizer.get_command('<|user|>'),
+            self.tokenizer.get_command('<|observation|>')
+        ]
+        result = self.model.generate(
+            input_ids=input_ids,
+            generation_config=self.generation_cfg,
+            eos_token_id=eos_token_id)
+
+        result = result[0].tolist()[input_len:]
+        response = self.tokenizer.decode(result)
+        # 遇到生成'<', '|', 'user', '|', '>'的case
+        response = response.split('<|user|>')[0].split('<|observation|>')[0]
+
+        return response

modelscope_agent/llm/modelscope.py

@@ -23,7 +23,10 @@ class BaseTool(ABC):
     parameters: List[Dict]
 
     def __init__(self, cfg: Optional[Dict] = {}):
-        self.cfg = cfg or {}
+        """
+        :param schema: Format of tools, default to oai format, in case there is a need for other formats
+        """
+        self.cfg = cfg.get(self.name, {})
 
         self.schema = self.cfg.get('schema', 'oai')
         self.function = self._build_function()