Support agent reference (#546)

docs/source/GetStarted/界面训练推理.md

@@ -8,7 +8,8 @@ swift web-ui
 
 web-ui没有传入参数，所有可控部分都在界面中。但是有几个环境变量可以使用：
 
-> WEBUI_SHARE=1 控制gradio是否是share状态
+> WEBUI_SHARE=1/0 默认为0 控制gradio是否是share状态
 > SWIFT_UI_LANG=en/zh 控制web-ui界面语言
-> WEBUI_SERVER server_name参数， web-ui host ip，0.0.0.0代表所有ip均可访问，127.0.0.1代表只允许本机访问
+> WEBUI_SERVER server_name参数，web-ui host ip，0.0.0.0代表所有ip均可访问，127.0.0.1代表只允许本机访问
 > WEBUI_PORT web-ui的端口号
+> USE_INFERENCE=1/0 默认0. 控制gradio的推理页面是直接加载模型推理或者部署（USE_INFERENCE=0）

docs/source/LLM/Agent微调最佳实践.md

@@ -267,6 +267,150 @@ Begin!
 
 ![image-20240201133359457](resources/image-20240201133359457.png)
 
+### 在命令行中使用Agent
+
+目前命令行的Agent推理支持需要指定`--eval_human true`，因为该参数为false的时候会读取数据集内容，此时无法手动传入`Observation:`后面的API调用结果。
+
+```shell
+swift infer --model_type chatglm3-6b-32k --eval_human true --stop_words Observation: --infer_backend pt
+```
+
+运行命令后，改变system字段：
+
+```shell
+# 单行system
+<<< reset-system
+<<< Answer the following questions as best you can. You have access to the following APIs:\n1. fire_recognition: Call this tool to interact with the fire recognition API. This API is used to recognize whether there is fire in the image. Parameters: [{"name": "image", "description": "The input image to recognize fire", "required": "True"}]\n\n2. fire_alert: Call this tool to interact with the fire alert API. This API will start an alert to warn the building's administraters. Parameters: []\n\n3. call_police: Call this tool to interact with the police calling API. This API will call 110 to catch the thief. Parameters: []\n\n4. call_fireman: Call this tool to interact with the fireman calling API. This API will call 119 to extinguish the fire. Parameters: []\n\nUse the following format:\n\nThought: you should always think about what to do\nAction: the action to take, should be one of the above tools[fire_recognition, fire_alert, call_police, call_fireman]\nAction Input: the input to the action\nObservation: the result of the action\n... (this Thought/Action/Action Input/Observation can be repeated zero or more times)\nThought: I now know the final answer\nFinal Answer: the final answer to the original input question\nBegin!
+```
+
+如果需要以多行方式输入，可以用下面的命令(多行信息以#号结束)：
+
+```shell
+# 多行system
+<<< multi-line#
+<<<[M] reset-system#
+<<<[MS] Answer the following questions as best you can. You have access to the following APIs:
+1. fire_recognition: Call this tool to interact with the fire recognition API. This API is used to recognize whether there is fire in the image. Parameters: [{"name": "image", "description": "The input image to recognize fire", "required": "True"}]
+
+2. fire_alert: Call this tool to interact with the fire alert API. This API will start an alert to warn the building's administraters. Parameters: []
+
+3. call_police: Call this tool to interact with the police calling API. This API will call 110 to catch the thief. Parameters: []
+
+4. call_fireman: Call this tool to interact with the fireman calling API. This API will call 119 to extinguish the fire. Parameters: []
+
+Use the following format:
+
+Thought: you should always think about what to do
+Action: the action to take, should be one of the above tools[fire_recognition, fire_alert, call_police, call_fireman]
+Action Input: the input to the action
+Observation: the result of the action
+... (this Thought/Action/Action Input/Observation can be repeated zero or more times)
+Thought: I now know the final answer
+Final Answer: the final answer to the original input question
+Begin!#
+```
+
+下面就可以进行Agent问答：
+
+```shell
+<<< 输入图片是/tmp/1.jpg，协助判断图片中是否存在着火点
+Thought: I need to use the fire\_recognition API to analyze the input image and determine if there are any signs of fire.
+
+Action: Use the fire\_recognition API to analyze the input image.
+
+Action Input: /tmp/1.jpg
+
+Observation:
+<<< [{'coordinate': [101.1, 200.9], 'on_fire': True}]
+Thought: The fire\_recognition API has returned a result indicating that there is fire in the input image.
+
+Final Answer: There is fire in the input image.
+```
+
+可以看到，模型已经返回了API调用的结果分析。用户可以继续问问题进行多轮Agent场景。也可以指定`--infer_backend vllm`和`--stream true`来使用vllm和流式推理。
+
+### 在部署中使用Agent
+
+由于部署不支持history管理，因此agent的API调用结果拼接需要用户自行进行，下面给出一个OpenAI格式可运行的代码范例。
+
+服务端：
+
+```shell
+swift deploy --model_type chatglm3-6b-32k --stop_words Observation:
+```
+
+客户端：
+
+```python
+from openai import OpenAI
+client = OpenAI(
+    api_key='EMPTY',
+    base_url='http://localhost:8000/v1',
+)
+model_type = client.models.list().data[0].id
+print(f'model_type: {model_type}')
+
+system = """Answer the following questions as best you can. You have access to the following APIs:
+1. fire_recognition: Call this tool to interact with the fire recognition API. This API is used to recognize whether there is fire in the image. Parameters: [{\"name\": \"image\", \"description\": \"The input image to recognize fire\", \"required\": \"True\"}]
+
+2. fire_alert: Call this tool to interact with the fire alert API. This API will start an alert to warn the building's administraters. Parameters: []
+
+3. call_police: Call this tool to interact with the police calling API. This API will call 110 to catch the thief. Parameters: []
+
+4. call_fireman: Call this tool to interact with the fireman calling API. This API will call 119 to extinguish the fire. Parameters: []
+
+Use the following format:
+
+Thought: you should always think about what to do
+Action: the action to take, should be one of the above tools[fire_recognition, fire_alert, call_police, call_fireman]
+Action Input: the input to the action
+Observation: the result of the action
+... (this Thought/Action/Action Input/Observation can be repeated zero or more times)
+Thought: I now know the final answer
+Final Answer: the final answer to the original input question
+Begin!"""
+messages = [{
+    'role': 'system',
+    'content': system
+}, {
+    'role': 'user',
+    'content': '输入图片是/tmp/1.jpg，协助判断图片中是否存在着火点'
+}]
+resp = client.chat.completions.create(
+    model=model_type,
+    messages=messages,
+    stop=['Observation:'],
+    seed=42)
+response = resp.choices[0].message.content
+print(f'response: {response}')
+
+# # 流式
+messages.append({'role': 'assistant', 'content': response + "\n[{'coordinate': [101.1, 200.9], 'on_fire': True}]"})
+print(messages)
+stream_resp = client.chat.completions.create(
+    model=model_type,
+    messages=messages,
+    stop=['Observation:'],
+    stream=True,
+    seed=42)
+
+print('response: ', end='')
+for chunk in stream_resp:
+    print(chunk.choices[0].delta.content, end='', flush=True)
+print()
+## Output:
+# model_type: chatglm3-6b-32k
+# response: Thought: I need to check if there is fire in the image
+# Action: Use fire\_recognition API
+# Action Input: /tmp/2.jpg
+# Observation:
+# [{'role': 'system', 'content': 'Answer the following questions as best you can. You have access to the following APIs:\n1. fire_recognition: Call this tool to interact with the fire recognition API. This API is used to recognize whether there is fire in the image. Parameters: [{"name": "image", "description": "The input image to recognize fire", "required": "True"}]\n\n2. fire_alert: Call this tool to interact with the fire alert API. This API will start an alert to warn the building\'s administraters. Parameters: []\n\n3. call_police: Call this tool to interact with the police calling API. This API will call 110 to catch the thief. Parameters: []\n\n4. call_fireman: Call this tool to interact with the fireman calling API. This API will call 119 to extinguish the fire. Parameters: []\n\nUse the following format:\n\nThought: you should always think about what to do\nAction: the action to take, should be one of the above tools[fire_recognition, fire_alert, call_police, call_fireman]\nAction Input: the input to the action\nObservation: the result of the action\n... (this Thought/Action/Action Input/Observation can be repeated zero or more times)\nThought: I now know the final answer\nFinal Answer: the final answer to the original input question\nBegin!'}, {'role': 'user', 'content': '输入图片是/tmp/2.jpg，协助判断图片中是否存在着火点'}, {'role': 'assistant', 'content': "Thought: I need to check if there is fire in the image\nAction: Use fire\\_recognition API\nAction Input: /tmp/2.jpg\nObservation:\n[{'coordinate': [101.1, 200.9], 'on_fire': True}]"}]
+# response:
+# Final Answer: There is fire in the image at coordinates [101.1, 200.9]
+```
+
+
+
 ## 总结
 
 通过SWIFT支持的Agent训练能力，我们使用ms-agent和ms-bench对qwen-7b-chat模型进行了微调。可以看到微调后模型保留了通用知识问答能力，并在system字段增加了API的情况下可以正确调用并完成任务。需要注意的是：

swift/llm/deploy.py

@@ -178,7 +178,10 @@ async def _generate_full():
             for output in result.outputs:
                 choice = ChatCompletionResponseChoice(
                     index=output.index,
-                    message=ChatMessage(role='assistant', content=output.text),
+                    message=ChatMessage(
+                        role='assistant',
+                        content=template.tokenizer.decode(
+                            output.token_ids, True)),
                     finish_reason=output.finish_reason,
                 )
                 choices.append(choice)
@@ -193,7 +196,7 @@ async def _generate_full():
             for output in result.outputs:
                 choice = CompletionResponseChoice(
                     index=output.index,
-                    text=output.text,
+                    text=template.tokenizer.decode(output.token_ids, True),
                     finish_reason=output.finish_reason,
                 )
                 choices.append(choice)
@@ -219,7 +222,8 @@ async def _generate_stream():
             if isinstance(request, ChatCompletionRequest):
                 choices = []
                 for output in result.outputs:
-                    delta_text = output.text[print_idx_list[output.index]:]
+                    text = template.tokenizer.decode(output.token_ids, True)
+                    delta_text = text[print_idx_list[output.index]:]
                     print_idx_list[output.index] += len(delta_text)
                     choice = ChatCompletionResponseStreamChoice(
                         index=output.index,
@@ -236,7 +240,8 @@ async def _generate_stream():
             else:
                 choices = []
                 for output in result.outputs:
-                    delta_text = output.text[print_idx_list[output.index]:]
+                    text = template.tokenizer.decode(output.token_ids, True)
+                    delta_text = text[print_idx_list[output.index]:]
                     print_idx_list[output.index] += len(delta_text)
                     choice = CompletionResponseStreamChoice(
                         index=output.index,

swift/llm/infer.py

@@ -315,6 +315,7 @@ def llm_infer(args: InferArguments) -> None:
             if not template.support_multi_round:
                 history = []
                 infer_kwargs = {}
+
             read_media_file(infer_kwargs, args.infer_media_type)
             if args.infer_backend == 'vllm':
                 request_list = [{
@@ -340,6 +341,8 @@ def llm_infer(args: InferArguments) -> None:
                     new_history = resp_list[0]['history']
                     print(response)
             else:
+                if args.stop_words:
+                    infer_kwargs['stop_words'] = args.stop_words
                 template_info = TEMPLATE_MAPPING[args.template_type]
                 support_stream = template_info.get('support_stream', True)
                 if args.stream and support_stream:

swift/llm/utils/argument.py

@@ -607,6 +607,7 @@ class InferArguments:
     top_p: float = 0.7
     repetition_penalty: float = 1.
     num_beams: int = 1
+    stop_words: List[str] = None
 
     # other
     use_flash_attn: Optional[bool] = None

swift/llm/utils/utils.py

@@ -468,7 +468,7 @@ def inference_stream(model: PreTrainedModel,
 
     # agent support
     is_observation = history[-1][-1].endswith(
-        'Observation:') if history else False
+        'Observation:') if history and history[-1][-1] else False
     if is_observation:
         history[-1][-1] = history[-1][-1] + query
         act_length = len(history[-1][-1])
@@ -576,7 +576,7 @@ def inference_stream(model: PreTrainedModel,
     ) and response[-len(template.suffix[-1]):] == template.suffix[-1]:
         response = response[:-len(template.suffix[-1])]
     if not is_observation:
-        history[-1] = (query, response)
+        history[-1] = [query, response]
     else:
         history[-1][-1] = history[-1][-1][:act_length] + response
     yield response, history
@@ -607,7 +607,7 @@ def inference(model: PreTrainedModel,
         history = deepcopy(history)
 
     is_observation = history[-1][-1].endswith(
-        'Observation:') if history else False
+        'Observation:') if history and history[-1][-1] else False
     if is_observation:
         history[-1][-1] = history[-1][-1] + query
         query = None
@@ -701,7 +701,7 @@ def inference(model: PreTrainedModel,
     ) and response[-len(template.suffix[-1]):] == template.suffix[-1]:
         response = response[:-len(template.suffix[-1])]
     if not is_observation:
-        history.append((query, response))
+        history.append([query, response])
     else:
         history[-1][-1] = history[-1][-1] + response
     return response, history

swift/llm/utils/vllm_utils.py

@@ -6,7 +6,7 @@
 
 import torch
 import vllm
-from modelscope import GenerationConfig, snapshot_download
+from modelscope import GenerationConfig
 from packaging import version
 from torch import dtype as Dtype
 from tqdm import tqdm
@@ -16,7 +16,7 @@
 
 from swift.utils import get_logger, seed_everything
 from .argument import InferArguments
-from .model import MODEL_MAPPING, get_model_tokenizer
+from .model import get_model_tokenizer
 from .template import Template, get_template
 from .utils import _is_chinese_char
 
@@ -210,10 +210,22 @@ def inference_stream_vllm(
                   str) and template.suffix[-1] not in generation_config.stop:
         generation_config.stop.append(template.suffix[-1])
 
+    request_temp = []
     for i, request in enumerate(request_list):
         history = request.get('history', None)
         if history is None:
             history = []
+
+        # agent support
+        is_observation = history[-1][-1].endswith(
+            'Observation:') if history and history[-1][-1] else False
+        act_length = None
+        if is_observation:
+            history[-1][-1] = history[-1][-1] + request['query']
+            act_length = len(history[-1][-1])
+            request['query'] = None
+        request_temp.append((is_observation, act_length))
+
         request['history'] = history
         inputs = template.encode(request)[0]
         input_ids = inputs['input_ids']
@@ -240,9 +252,13 @@ def inference_stream_vllm(
             safe_response = response[:print_idx_list[i]]
             query = request['query']
             history = request['history']
-            if resp_list[i] is None:
+            if resp_list[i] is None and not request_temp[i][0]:
                 history.append(None)
-            history[-1] = (query, safe_response)
+            if not request_temp[i][0]:
+                history[-1] = [query, safe_response]
+            else:
+                history[-1][
+                    -1] = history[-1][-1][:request_temp[i][1]] + safe_response
             resp_list[i] = {'response': safe_response, 'history': history}
             if output.finished:
                 prog_bar.update()
@@ -284,6 +300,12 @@ def inference_vllm(llm_engine: LLMEngine,
         history = request.get('history', None)
         if history is None:
             history = []
+
+        is_observation = history[-1][-1].endswith(
+            'Observation:') if history and history[-1][-1] else False
+        if is_observation:
+            history[-1][-1] = history[-1][-1] + request['query']
+            request['query'] = None
         request['history'] = history
         inputs = template.encode(request)[0]
         input_ids = inputs['input_ids']
@@ -308,7 +330,10 @@ def inference_vllm(llm_engine: LLMEngine,
         response = tokenizer.decode(output.outputs[0].token_ids, True)
         query = request['query']
         history = request['history']
-        history.append((query, response))
+        if not is_observation:
+            history.append([query, response])
+        else:
+            history[-1][-1] = history[-1][-1] + response
         resp_list[i] = {'response': response, 'history': history}
         if verbose:
             print(
@@ -355,6 +380,7 @@ def prepare_vllm_engine_template(
         temperature=args.temperature,
         top_k=args.top_k,
         top_p=args.top_p,
+        stop=args.stop_words,
         repetition_penalty=args.repetition_penalty,
         num_beams=args.num_beams)
     logger.info(f'generation_config: {generation_config}')