Spaces:
Running
Running
ff_li
commited on
Commit
•
e7de191
1
Parent(s):
f67d239
oai
Browse files- agent/llm/__init__.py +1 -1
- agent/llm/{qwen_oai.py → oai.py} +1 -1
- agent/llm/qwen_dashscope.py +0 -86
- agent/llm/qwen_oai_bak.py +0 -527
- assistant_server.py +2 -2
- workstation_server.py +2 -2
agent/llm/__init__.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
from .base import BaseChatModel
|
2 |
from .qwen_dashscope import QwenChatAtDS
|
3 |
-
from .
|
4 |
|
|
|
1 |
from .base import BaseChatModel
|
2 |
from .qwen_dashscope import QwenChatAtDS
|
3 |
+
from .oai import ChatAsOAI
|
4 |
|
agent/llm/{qwen_oai.py → oai.py}
RENAMED
@@ -5,7 +5,7 @@ from agent.llm.base import BaseChatModel
|
|
5 |
from typing import Dict, List, Literal, Optional, Union
|
6 |
|
7 |
|
8 |
-
class
|
9 |
|
10 |
def __init__(self, model: str):
|
11 |
super().__init__()
|
|
|
5 |
from typing import Dict, List, Literal, Optional, Union
|
6 |
|
7 |
|
8 |
+
class ChatAsOAI(BaseChatModel):
|
9 |
|
10 |
def __init__(self, model: str):
|
11 |
super().__init__()
|
agent/llm/qwen_dashscope.py
DELETED
@@ -1,86 +0,0 @@
|
|
1 |
-
import os
|
2 |
-
from http import HTTPStatus
|
3 |
-
from typing import Dict, Iterator, List, Optional
|
4 |
-
|
5 |
-
import dashscope
|
6 |
-
|
7 |
-
from agent.llm.base import BaseChatModel
|
8 |
-
|
9 |
-
|
10 |
-
class QwenChatAtDS(BaseChatModel):
|
11 |
-
|
12 |
-
def __init__(self, model: str, api_key: str):
|
13 |
-
super().__init__()
|
14 |
-
self.model = model
|
15 |
-
dashscope.api_key = api_key.strip() or os.getenv('DASHSCOPE_API_KEY',
|
16 |
-
default='')
|
17 |
-
assert dashscope.api_key, 'DASHSCOPE_API_KEY is required.'
|
18 |
-
|
19 |
-
def _chat_stream(
|
20 |
-
self,
|
21 |
-
messages: List[Dict],
|
22 |
-
stop: Optional[List[str]] = None,
|
23 |
-
) -> Iterator[str]:
|
24 |
-
stop = stop or []
|
25 |
-
response = dashscope.Generation.call(
|
26 |
-
self.model,
|
27 |
-
messages=messages, # noqa
|
28 |
-
stop_words=[{
|
29 |
-
'stop_str': word,
|
30 |
-
'mode': 'exclude'
|
31 |
-
} for word in stop],
|
32 |
-
top_p=0.8,
|
33 |
-
result_format='message',
|
34 |
-
stream=True,
|
35 |
-
)
|
36 |
-
last_len = 0
|
37 |
-
delay_len = 5
|
38 |
-
in_delay = False
|
39 |
-
text = ''
|
40 |
-
for trunk in response:
|
41 |
-
if trunk.status_code == HTTPStatus.OK:
|
42 |
-
text = trunk.output.choices[0].message.content
|
43 |
-
if (len(text) - last_len) <= delay_len:
|
44 |
-
in_delay = True
|
45 |
-
continue
|
46 |
-
else:
|
47 |
-
in_delay = False
|
48 |
-
real_text = text[:-delay_len]
|
49 |
-
now_rsp = real_text[last_len:]
|
50 |
-
yield now_rsp
|
51 |
-
last_len = len(real_text)
|
52 |
-
else:
|
53 |
-
err = '\nError code: %s. Error message: %s' % (trunk.code,
|
54 |
-
trunk.message)
|
55 |
-
if trunk.code == 'DataInspectionFailed':
|
56 |
-
err += '\n错误码: 数据检查失败。错误信息: 输入数据可能包含不适当的内容。'
|
57 |
-
text = ''
|
58 |
-
yield f'{err}'
|
59 |
-
if text and (in_delay or (last_len != len(text))):
|
60 |
-
yield text[last_len:]
|
61 |
-
|
62 |
-
def _chat_no_stream(
|
63 |
-
self,
|
64 |
-
messages: List[Dict],
|
65 |
-
stop: Optional[List[str]] = None,
|
66 |
-
) -> str:
|
67 |
-
stop = stop or []
|
68 |
-
response = dashscope.Generation.call(
|
69 |
-
self.model,
|
70 |
-
messages=messages, # noqa
|
71 |
-
result_format='message',
|
72 |
-
stream=False,
|
73 |
-
stop_words=[{
|
74 |
-
'stop_str': word,
|
75 |
-
'mode': 'exclude'
|
76 |
-
} for word in stop],
|
77 |
-
top_p=0.8,
|
78 |
-
)
|
79 |
-
if response.status_code == HTTPStatus.OK:
|
80 |
-
return response.output.choices[0].message.content
|
81 |
-
else:
|
82 |
-
err = 'Error code: %s, error message: %s' % (
|
83 |
-
response.code,
|
84 |
-
response.message,
|
85 |
-
)
|
86 |
-
return err
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
agent/llm/qwen_oai_bak.py
DELETED
@@ -1,527 +0,0 @@
|
|
1 |
-
import os
|
2 |
-
from typing import Dict, Iterator, List, Optional
|
3 |
-
|
4 |
-
import openai
|
5 |
-
|
6 |
-
from agent.llm.base import BaseChatModel
|
7 |
-
|
8 |
-
import re
|
9 |
-
import copy
|
10 |
-
import json
|
11 |
-
import time
|
12 |
-
from contextlib import asynccontextmanager
|
13 |
-
from typing import Dict, List, Literal, Optional, Union
|
14 |
-
import torch
|
15 |
-
from pydantic import BaseModel, Field
|
16 |
-
from sse_starlette.sse import EventSourceResponse
|
17 |
-
from transformers import AutoTokenizer, AutoModelForCausalLM
|
18 |
-
from transformers.generation import GenerationConfig
|
19 |
-
|
20 |
-
|
21 |
-
def _gc(forced: bool = False, disable_gc: bool = True):
|
22 |
-
if disable_gc and not forced:
|
23 |
-
return
|
24 |
-
|
25 |
-
import gc
|
26 |
-
gc.collect()
|
27 |
-
if torch.cuda.is_available():
|
28 |
-
torch.cuda.empty_cache()
|
29 |
-
|
30 |
-
|
31 |
-
class ChatMessage(BaseModel):
|
32 |
-
role: Literal["user", "assistant", "system", "function"]
|
33 |
-
content: Optional[str]
|
34 |
-
function_call: Optional[Dict] = None
|
35 |
-
|
36 |
-
|
37 |
-
class DeltaMessage(BaseModel):
|
38 |
-
role: Optional[Literal["user", "assistant", "system"]] = None
|
39 |
-
content: Optional[str] = None
|
40 |
-
|
41 |
-
|
42 |
-
class ChatCompletionRequest(BaseModel):
|
43 |
-
model: str
|
44 |
-
messages: List[ChatMessage]
|
45 |
-
functions: Optional[List[Dict]] = None
|
46 |
-
temperature: Optional[float] = None
|
47 |
-
top_p: Optional[float] = None
|
48 |
-
max_length: Optional[int] = None
|
49 |
-
stream: Optional[bool] = False
|
50 |
-
stop: Optional[List[str]] = None
|
51 |
-
|
52 |
-
|
53 |
-
class ChatCompletionResponseChoice(BaseModel):
|
54 |
-
index: int
|
55 |
-
message: ChatMessage
|
56 |
-
finish_reason: Literal["stop", "length", "function_call"]
|
57 |
-
|
58 |
-
|
59 |
-
class ChatCompletionResponseStreamChoice(BaseModel):
|
60 |
-
index: int
|
61 |
-
delta: DeltaMessage
|
62 |
-
finish_reason: Optional[Literal["stop", "length"]]
|
63 |
-
|
64 |
-
|
65 |
-
class ChatCompletionResponse(BaseModel):
|
66 |
-
model: str
|
67 |
-
object: Literal["chat.completion", "chat.completion.chunk"]
|
68 |
-
choices: List[
|
69 |
-
Union[ChatCompletionResponseChoice, ChatCompletionResponseStreamChoice]
|
70 |
-
]
|
71 |
-
created: Optional[int] = Field(default_factory=lambda: int(time.time()))
|
72 |
-
|
73 |
-
|
74 |
-
# To work around that unpleasant leading-\n tokenization issue!
|
75 |
-
def add_extra_stop_words(stop_words):
|
76 |
-
if stop_words:
|
77 |
-
_stop_words = []
|
78 |
-
_stop_words.extend(stop_words)
|
79 |
-
for x in stop_words:
|
80 |
-
s = x.lstrip("\n")
|
81 |
-
if s and (s not in _stop_words):
|
82 |
-
_stop_words.append(s)
|
83 |
-
return _stop_words
|
84 |
-
return stop_words
|
85 |
-
|
86 |
-
|
87 |
-
def trim_stop_words(response, stop_words):
|
88 |
-
if stop_words:
|
89 |
-
for stop in stop_words:
|
90 |
-
idx = response.find(stop)
|
91 |
-
if idx != -1:
|
92 |
-
response = response[:idx]
|
93 |
-
return response
|
94 |
-
|
95 |
-
|
96 |
-
TOOL_DESC = """{name_for_model}: Call this tool to interact with the {name_for_human} API. What is the {name_for_human} API useful for? {description_for_model} Parameters: {parameters}"""
|
97 |
-
|
98 |
-
REACT_INSTRUCTION = """Answer the following questions as best you can. You have access to the following APIs:
|
99 |
-
|
100 |
-
{tools_text}
|
101 |
-
|
102 |
-
Use the following format:
|
103 |
-
|
104 |
-
Question: the input question you must answer
|
105 |
-
Thought: you should always think about what to do
|
106 |
-
Action: the action to take, should be one of [{tools_name_text}]
|
107 |
-
Action Input: the input to the action
|
108 |
-
Observation: the result of the action
|
109 |
-
... (this Thought/Action/Action Input/Observation can be repeated zero or more times)
|
110 |
-
Thought: I now know the final answer
|
111 |
-
Final Answer: the final answer to the original input question
|
112 |
-
|
113 |
-
Begin!"""
|
114 |
-
|
115 |
-
_TEXT_COMPLETION_CMD = object()
|
116 |
-
|
117 |
-
|
118 |
-
#
|
119 |
-
# Temporarily, the system role does not work as expected.
|
120 |
-
# We advise that you write the setups for role-play in your query,
|
121 |
-
# i.e., use the user role instead of the system role.
|
122 |
-
#
|
123 |
-
# TODO: Use real system role when the model is ready.
|
124 |
-
#
|
125 |
-
def parse_messages(messages, functions):
|
126 |
-
if all(m.role != "user" for m in messages):
|
127 |
-
raise Exception(f"Invalid request: Expecting at least one user message.", )
|
128 |
-
messages = copy.deepcopy(messages)
|
129 |
-
default_system = "You are a helpful assistant."
|
130 |
-
system = ""
|
131 |
-
if messages[0].role == "system":
|
132 |
-
system = messages.pop(0).content.lstrip("\n").rstrip()
|
133 |
-
if system == default_system:
|
134 |
-
system = ""
|
135 |
-
|
136 |
-
if functions:
|
137 |
-
tools_text = []
|
138 |
-
tools_name_text = []
|
139 |
-
for func_info in functions:
|
140 |
-
name = func_info.get("name", "")
|
141 |
-
name_m = func_info.get("name_for_model", name)
|
142 |
-
name_h = func_info.get("name_for_human", name)
|
143 |
-
desc = func_info.get("description", "")
|
144 |
-
desc_m = func_info.get("description_for_model", desc)
|
145 |
-
tool = TOOL_DESC.format(
|
146 |
-
name_for_model=name_m,
|
147 |
-
name_for_human=name_h,
|
148 |
-
# Hint: You can add the following format requirements in description:
|
149 |
-
# "Format the arguments as a JSON object."
|
150 |
-
# "Enclose the code within triple backticks (`) at the beginning and end of the code."
|
151 |
-
description_for_model=desc_m,
|
152 |
-
parameters=json.dumps(func_info["parameters"], ensure_ascii=False),
|
153 |
-
)
|
154 |
-
tools_text.append(tool)
|
155 |
-
tools_name_text.append(name_m)
|
156 |
-
tools_text = "\n\n".join(tools_text)
|
157 |
-
tools_name_text = ", ".join(tools_name_text)
|
158 |
-
system += "\n\n" + REACT_INSTRUCTION.format(
|
159 |
-
tools_text=tools_text,
|
160 |
-
tools_name_text=tools_name_text,
|
161 |
-
)
|
162 |
-
system = system.lstrip("\n").rstrip()
|
163 |
-
|
164 |
-
dummy_thought = {
|
165 |
-
"en": "\nThought: I now know the final answer.\nFinal answer: ",
|
166 |
-
"zh": "\nThought: 我会作答了。\nFinal answer: ",
|
167 |
-
}
|
168 |
-
|
169 |
-
_messages = messages
|
170 |
-
messages = []
|
171 |
-
for m_idx, m in enumerate(_messages):
|
172 |
-
role, content, func_call = m.role, m.content, m.function_call
|
173 |
-
if content:
|
174 |
-
content = content.lstrip("\n").rstrip()
|
175 |
-
if role == "function":
|
176 |
-
if (len(messages) == 0) or (messages[-1].role != "assistant"):
|
177 |
-
raise Exception("Invalid request: Expecting role assistant before role function.")
|
178 |
-
messages[-1].content += f"\nObservation: {content}"
|
179 |
-
if m_idx == len(_messages) - 1:
|
180 |
-
messages[-1].content += "\nThought:"
|
181 |
-
elif role == "assistant":
|
182 |
-
if len(messages) == 0:
|
183 |
-
raise Exception(f"Invalid request: Expecting role user before role assistant.")
|
184 |
-
last_msg = messages[-1].content
|
185 |
-
last_msg_has_zh = len(re.findall(r"[\u4e00-\u9fff]+", last_msg)) > 0
|
186 |
-
if func_call is None:
|
187 |
-
if functions:
|
188 |
-
content = dummy_thought["zh" if last_msg_has_zh else "en"] + content
|
189 |
-
else:
|
190 |
-
f_name, f_args = func_call["name"], func_call["arguments"]
|
191 |
-
if not content:
|
192 |
-
if last_msg_has_zh:
|
193 |
-
content = f"Thought: 我可以使用 {f_name} API。"
|
194 |
-
else:
|
195 |
-
content = f"Thought: I can use {f_name}."
|
196 |
-
content = f"\n{content}\nAction: {f_name}\nAction Input: {f_args}"
|
197 |
-
if messages[-1].role == "user":
|
198 |
-
messages.append(
|
199 |
-
ChatMessage(role="assistant", content=content.lstrip("\n").rstrip())
|
200 |
-
)
|
201 |
-
else:
|
202 |
-
messages[-1].content += content
|
203 |
-
elif role == "user":
|
204 |
-
messages.append(
|
205 |
-
ChatMessage(role="user", content=content.lstrip("\n").rstrip())
|
206 |
-
)
|
207 |
-
else:
|
208 |
-
raise Exception(
|
209 |
-
f"Invalid request: Incorrect role {role}."
|
210 |
-
)
|
211 |
-
|
212 |
-
query = _TEXT_COMPLETION_CMD
|
213 |
-
if messages[-1].role == "user":
|
214 |
-
query = messages[-1].content
|
215 |
-
messages = messages[:-1]
|
216 |
-
|
217 |
-
if len(messages) % 2 != 0:
|
218 |
-
raise Exception("Invalid request")
|
219 |
-
|
220 |
-
history = [] # [(Q1, A1), (Q2, A2), ..., (Q_last_turn, A_last_turn)]
|
221 |
-
for i in range(0, len(messages), 2):
|
222 |
-
if messages[i].role == "user" and messages[i + 1].role == "assistant":
|
223 |
-
usr_msg = messages[i].content.lstrip("\n").rstrip()
|
224 |
-
bot_msg = messages[i + 1].content.lstrip("\n").rstrip()
|
225 |
-
if system and (i == len(messages) - 2):
|
226 |
-
usr_msg = f"{system}\n\nQuestion: {usr_msg}"
|
227 |
-
system = ""
|
228 |
-
for t in dummy_thought.values():
|
229 |
-
t = t.lstrip("\n")
|
230 |
-
if bot_msg.startswith(t) and ("\nAction: " in bot_msg):
|
231 |
-
bot_msg = bot_msg[len(t):]
|
232 |
-
history.append([usr_msg, bot_msg])
|
233 |
-
else:
|
234 |
-
raise Exception("Invalid request: Expecting exactly one user (or function) role before every assistant role.")
|
235 |
-
if system:
|
236 |
-
assert query is not _TEXT_COMPLETION_CMD
|
237 |
-
query = f"{system}\n\nQuestion: {query}"
|
238 |
-
return query, history
|
239 |
-
|
240 |
-
|
241 |
-
def parse_response(response):
|
242 |
-
func_name, func_args = "", ""
|
243 |
-
i = response.rfind("\nAction:")
|
244 |
-
j = response.rfind("\nAction Input:")
|
245 |
-
k = response.rfind("\nObservation:")
|
246 |
-
if 0 <= i < j: # If the text has `Action` and `Action input`,
|
247 |
-
if k < j: # but does not contain `Observation`,
|
248 |
-
# then it is likely that `Observation` is omitted by the LLM,
|
249 |
-
# because the output text may have discarded the stop word.
|
250 |
-
response = response.rstrip() + "\nObservation:" # Add it back.
|
251 |
-
k = response.rfind("\nObservation:")
|
252 |
-
func_name = response[i + len("\nAction:"): j].strip()
|
253 |
-
func_args = response[j + len("\nAction Input:"): k].strip()
|
254 |
-
if func_name:
|
255 |
-
choice_data = ChatCompletionResponseChoice(
|
256 |
-
index=0,
|
257 |
-
message=ChatMessage(
|
258 |
-
role="assistant",
|
259 |
-
content=response[:i],
|
260 |
-
function_call={"name": func_name, "arguments": func_args},
|
261 |
-
),
|
262 |
-
finish_reason="function_call",
|
263 |
-
)
|
264 |
-
return choice_data
|
265 |
-
z = response.rfind("\nFinal Answer: ")
|
266 |
-
if z >= 0:
|
267 |
-
response = response[z + len("\nFinal Answer: "):]
|
268 |
-
choice_data = ChatCompletionResponseChoice(
|
269 |
-
index=0,
|
270 |
-
message=ChatMessage(role="assistant", content=response),
|
271 |
-
finish_reason="stop",
|
272 |
-
)
|
273 |
-
return choice_data
|
274 |
-
|
275 |
-
|
276 |
-
# completion mode, not chat mode
|
277 |
-
def text_complete_last_message(history, stop_words_ids, gen_kwargs):
|
278 |
-
im_start = "<|im_start|>"
|
279 |
-
im_end = "<|im_end|>"
|
280 |
-
prompt = f"{im_start}system\nYou are a helpful assistant.{im_end}"
|
281 |
-
for i, (query, response) in enumerate(history):
|
282 |
-
query = query.lstrip("\n").rstrip()
|
283 |
-
response = response.lstrip("\n").rstrip()
|
284 |
-
prompt += f"\n{im_start}user\n{query}{im_end}"
|
285 |
-
prompt += f"\n{im_start}assistant\n{response}{im_end}"
|
286 |
-
prompt = prompt[: -len(im_end)]
|
287 |
-
|
288 |
-
_stop_words_ids = [tokenizer.encode(im_end)]
|
289 |
-
if stop_words_ids:
|
290 |
-
for s in stop_words_ids:
|
291 |
-
_stop_words_ids.append(s)
|
292 |
-
stop_words_ids = _stop_words_ids
|
293 |
-
|
294 |
-
input_ids = torch.tensor([tokenizer.encode(prompt)]).to(qmodel.device)
|
295 |
-
output = qmodel.generate(input_ids, stop_words_ids=stop_words_ids, **gen_kwargs).tolist()[0]
|
296 |
-
output = tokenizer.decode(output, errors="ignore")
|
297 |
-
assert output.startswith(prompt)
|
298 |
-
output = output[len(prompt):]
|
299 |
-
output = trim_stop_words(output, ["<|endoftext|>", im_end])
|
300 |
-
print(f"<completion>\n{prompt}\n<!-- *** -->\n{output}\n</completion>")
|
301 |
-
return output
|
302 |
-
|
303 |
-
|
304 |
-
def create_chat_completion(request: ChatCompletionRequest, qmodel, tokenizer):
|
305 |
-
|
306 |
-
gen_kwargs = {}
|
307 |
-
if request.temperature is not None:
|
308 |
-
if request.temperature < 0.01:
|
309 |
-
gen_kwargs['top_k'] = 1 # greedy decoding
|
310 |
-
else:
|
311 |
-
# Not recommended. Please tune top_p instead.
|
312 |
-
gen_kwargs['temperature'] = request.temperature
|
313 |
-
if request.top_p is not None:
|
314 |
-
gen_kwargs['top_p'] = request.top_p
|
315 |
-
|
316 |
-
stop_words = add_extra_stop_words(request.stop)
|
317 |
-
if request.functions:
|
318 |
-
stop_words = stop_words or []
|
319 |
-
if "Observation:" not in stop_words:
|
320 |
-
stop_words.append("Observation:")
|
321 |
-
|
322 |
-
query, history = parse_messages(request.messages, request.functions)
|
323 |
-
|
324 |
-
if request.stream:
|
325 |
-
if request.functions:
|
326 |
-
raise Exception("Invalid request: Function calling is not yet implemented for stream mode.")
|
327 |
-
generate = predict(query, history, request.model, stop_words, gen_kwargs, qmodel, tokenizer)
|
328 |
-
return generate
|
329 |
-
# return EventSourceResponse(generate, media_type="text/event-stream")
|
330 |
-
|
331 |
-
stop_words_ids = [tokenizer.encode(s) for s in stop_words] if stop_words else None
|
332 |
-
if query is _TEXT_COMPLETION_CMD:
|
333 |
-
response = text_complete_last_message(history, stop_words_ids=stop_words_ids, gen_kwargs=gen_kwargs)
|
334 |
-
else:
|
335 |
-
response, _ = qmodel.chat(
|
336 |
-
tokenizer,
|
337 |
-
query,
|
338 |
-
history=history,
|
339 |
-
stop_words_ids=stop_words_ids,
|
340 |
-
**gen_kwargs
|
341 |
-
)
|
342 |
-
print(f"<chat>\n{history}\n{query}\n<!-- *** -->\n{response}\n</chat>")
|
343 |
-
_gc()
|
344 |
-
|
345 |
-
response = trim_stop_words(response, stop_words)
|
346 |
-
if request.functions:
|
347 |
-
choice_data = parse_response(response)
|
348 |
-
else:
|
349 |
-
choice_data = ChatCompletionResponseChoice(
|
350 |
-
index=0,
|
351 |
-
message=ChatMessage(role="assistant", content=response),
|
352 |
-
finish_reason="stop",
|
353 |
-
)
|
354 |
-
return ChatCompletionResponse(
|
355 |
-
model=request.model, choices=[choice_data], object="chat.completion"
|
356 |
-
)
|
357 |
-
|
358 |
-
|
359 |
-
def _dump_json(data: BaseModel, *args, **kwargs) -> str:
|
360 |
-
try:
|
361 |
-
return data.model_dump_json(*args, **kwargs)
|
362 |
-
except AttributeError: # pydantic<2.0.0
|
363 |
-
return data.json(*args, **kwargs) # noqa
|
364 |
-
|
365 |
-
|
366 |
-
def predict(
|
367 |
-
query: str, history: List[List[str]], model_id: str, stop_words: List[str], gen_kwargs: Dict, qmodel, tokenizer
|
368 |
-
):
|
369 |
-
choice_data = ChatCompletionResponseStreamChoice(
|
370 |
-
index=0, delta=DeltaMessage(role="assistant"), finish_reason=None
|
371 |
-
)
|
372 |
-
chunk = ChatCompletionResponse(
|
373 |
-
model=model_id, choices=[choice_data], object="chat.completion.chunk"
|
374 |
-
)
|
375 |
-
# yield "{}".format(_dump_json(chunk, exclude_unset=True))
|
376 |
-
yield chunk
|
377 |
-
|
378 |
-
current_length = 0
|
379 |
-
stop_words_ids = [tokenizer.encode(s) for s in stop_words] if stop_words else None
|
380 |
-
if stop_words:
|
381 |
-
# TODO: It's a little bit tricky to trim stop words in the stream mode.
|
382 |
-
raise Exception("Invalid request: custom stop words are not yet supported for stream mode.", )
|
383 |
-
response_generator = qmodel.chat_stream(
|
384 |
-
tokenizer, query, history=history, stop_words_ids=stop_words_ids, **gen_kwargs
|
385 |
-
)
|
386 |
-
for new_response in response_generator:
|
387 |
-
if len(new_response) == current_length:
|
388 |
-
continue
|
389 |
-
|
390 |
-
new_text = new_response[current_length:]
|
391 |
-
current_length = len(new_response)
|
392 |
-
|
393 |
-
choice_data = ChatCompletionResponseStreamChoice(
|
394 |
-
index=0, delta=DeltaMessage(content=new_text), finish_reason=None
|
395 |
-
)
|
396 |
-
chunk = ChatCompletionResponse(
|
397 |
-
model=model_id, choices=[choice_data], object="chat.completion.chunk"
|
398 |
-
)
|
399 |
-
# yield "{}".format(_dump_json(chunk, exclude_unset=True))
|
400 |
-
yield chunk
|
401 |
-
|
402 |
-
choice_data = ChatCompletionResponseStreamChoice(
|
403 |
-
index=0, delta=DeltaMessage(), finish_reason="stop"
|
404 |
-
)
|
405 |
-
chunk = ChatCompletionResponse(
|
406 |
-
model=model_id, choices=[choice_data], object="chat.completion.chunk"
|
407 |
-
)
|
408 |
-
# yield "{}".format(_dump_json(chunk, exclude_unset=True))
|
409 |
-
yield chunk
|
410 |
-
# yield "[DONE]"
|
411 |
-
|
412 |
-
_gc()
|
413 |
-
|
414 |
-
|
415 |
-
class QwenChatAsOAI(BaseChatModel):
|
416 |
-
|
417 |
-
def __init__(self, model: str, api_key: str, model_server: str):
|
418 |
-
self.model = model
|
419 |
-
super().__init__()
|
420 |
-
tokenizer = AutoTokenizer.from_pretrained(
|
421 |
-
self.model,
|
422 |
-
trust_remote_code=True,
|
423 |
-
resume_download=True,
|
424 |
-
)
|
425 |
-
device_map = "cpu"
|
426 |
-
# device_map = "auto"
|
427 |
-
qmodel = AutoModelForCausalLM.from_pretrained(
|
428 |
-
self.model,
|
429 |
-
device_map=device_map,
|
430 |
-
trust_remote_code=True,
|
431 |
-
resume_download=True,
|
432 |
-
).eval()
|
433 |
-
|
434 |
-
qmodel.generation_config = GenerationConfig.from_pretrained(
|
435 |
-
self.model,
|
436 |
-
trust_remote_code=True,
|
437 |
-
resume_download=True,
|
438 |
-
)
|
439 |
-
self.qmodel = qmodel
|
440 |
-
self.tokenizer = tokenizer
|
441 |
-
|
442 |
-
def _chat_stream(
|
443 |
-
self,
|
444 |
-
messages: List[Dict],
|
445 |
-
stop: Optional[List[str]] = None,
|
446 |
-
) -> Iterator[str]:
|
447 |
-
_request = ChatCompletionRequest(model=self.model,
|
448 |
-
messages=messages,
|
449 |
-
stop=stop,
|
450 |
-
stream=True)
|
451 |
-
response = create_chat_completion(_request, self.qmodel, self.tokenizer)
|
452 |
-
# TODO: error handling
|
453 |
-
for chunk in response:
|
454 |
-
if hasattr(chunk.choices[0].delta, 'content'):
|
455 |
-
yield chunk.choices[0].delta.content
|
456 |
-
|
457 |
-
def _chat_no_stream(
|
458 |
-
self,
|
459 |
-
messages: List[Dict],
|
460 |
-
stop: Optional[List[str]] = None,
|
461 |
-
) -> str:
|
462 |
-
_request = ChatCompletionRequest(model=self.model, messages=messages, stop=stop, stream=False)
|
463 |
-
response = create_chat_completion(_request, self.qmodel, self.tokenizer)
|
464 |
-
# TODO: error handling
|
465 |
-
return response.choices[0].message.content
|
466 |
-
|
467 |
-
def chat_with_functions(self,
|
468 |
-
messages: List[Dict],
|
469 |
-
functions: Optional[List[Dict]] = None) -> Dict:
|
470 |
-
if functions:
|
471 |
-
_request = ChatCompletionRequest(model=self.model, messages=messages, functions=functions)
|
472 |
-
response = create_chat_completion(_request, self.qmodel, self.tokenizer)
|
473 |
-
else:
|
474 |
-
_request = ChatCompletionRequest(model=self.model, messages=messages)
|
475 |
-
response = create_chat_completion(_request, self.qmodel, self.tokenizer)
|
476 |
-
# TODO: error handling
|
477 |
-
return response.choices[0].message.model_dump()
|
478 |
-
|
479 |
-
|
480 |
-
class QwenChatAsOAI1(BaseChatModel):
|
481 |
-
|
482 |
-
def __init__(self, model: str, api_key: str, model_server: str):
|
483 |
-
super().__init__()
|
484 |
-
if model_server.strip().lower() != 'openai':
|
485 |
-
openai.api_base = model_server
|
486 |
-
openai.api_key = api_key.strip() or os.getenv('OPENAI_API_KEY',
|
487 |
-
default='EMPTY')
|
488 |
-
self.model = model
|
489 |
-
|
490 |
-
def _chat_stream(
|
491 |
-
self,
|
492 |
-
messages: List[Dict],
|
493 |
-
stop: Optional[List[str]] = None,
|
494 |
-
) -> Iterator[str]:
|
495 |
-
response = openai.ChatCompletion.create(model=self.model,
|
496 |
-
messages=messages,
|
497 |
-
stop=stop,
|
498 |
-
stream=True)
|
499 |
-
# TODO: error handling
|
500 |
-
for chunk in response:
|
501 |
-
if hasattr(chunk.choices[0].delta, 'content'):
|
502 |
-
yield chunk.choices[0].delta.content
|
503 |
-
|
504 |
-
def _chat_no_stream(
|
505 |
-
self,
|
506 |
-
messages: List[Dict],
|
507 |
-
stop: Optional[List[str]] = None,
|
508 |
-
) -> str:
|
509 |
-
response = openai.ChatCompletion.create(model=self.model,
|
510 |
-
messages=messages,
|
511 |
-
stop=stop,
|
512 |
-
stream=False)
|
513 |
-
# TODO: error handling
|
514 |
-
return response.choices[0].message.content
|
515 |
-
|
516 |
-
def chat_with_functions(self,
|
517 |
-
messages: List[Dict],
|
518 |
-
functions: Optional[List[Dict]] = None) -> Dict:
|
519 |
-
if functions:
|
520 |
-
response = openai.ChatCompletion.create(model=self.model,
|
521 |
-
messages=messages,
|
522 |
-
functions=functions)
|
523 |
-
else:
|
524 |
-
response = openai.ChatCompletion.create(model=self.model,
|
525 |
-
messages=messages)
|
526 |
-
# TODO: error handling
|
527 |
-
return response.choices[0].message
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
assistant_server.py
CHANGED
@@ -4,11 +4,11 @@ from pathlib import Path
|
|
4 |
import gradio as gr
|
5 |
import jsonlines
|
6 |
from agent.actions import RetrievalQA
|
7 |
-
from agent.llm import
|
8 |
from agent.memory import Memory
|
9 |
from utils import service, cache_file, max_ref_token
|
10 |
|
11 |
-
llm =
|
12 |
mem = Memory(llm=llm, stream=False)
|
13 |
|
14 |
with open('css/main.css', 'r') as f:
|
|
|
4 |
import gradio as gr
|
5 |
import jsonlines
|
6 |
from agent.actions import RetrievalQA
|
7 |
+
from agent.llm import ChatAsOAI
|
8 |
from agent.memory import Memory
|
9 |
from utils import service, cache_file, max_ref_token
|
10 |
|
11 |
+
llm = ChatAsOAI(model="gpt-3.5-turbo")
|
12 |
mem = Memory(llm=llm, stream=False)
|
13 |
|
14 |
with open('css/main.css', 'r') as f:
|
workstation_server.py
CHANGED
@@ -7,7 +7,7 @@ import gradio as gr
|
|
7 |
import jsonlines
|
8 |
from agent.actions import (ContinueWriting, ReAct, RetrievalQA, WriteFromScratch)
|
9 |
from agent.actions.function_calling import FunctionCalling
|
10 |
-
from agent.llm import
|
11 |
from agent.log import logger
|
12 |
from agent.memory import Memory
|
13 |
from agent.tools import call_plugin, list_of_all_functions
|
@@ -15,7 +15,7 @@ from agent.utils.utils import (format_answer, get_last_one_line_context,
|
|
15 |
has_chinese_chars, save_text_to_file)
|
16 |
from utils import service, extract_and_cache_document, code_interpreter_ws, cache_root, max_ref_token, max_days, download_root
|
17 |
|
18 |
-
llm =
|
19 |
|
20 |
mem = Memory(llm=llm, stream=False)
|
21 |
|
|
|
7 |
import jsonlines
|
8 |
from agent.actions import (ContinueWriting, ReAct, RetrievalQA, WriteFromScratch)
|
9 |
from agent.actions.function_calling import FunctionCalling
|
10 |
+
from agent.llm import ChatAsOAI
|
11 |
from agent.log import logger
|
12 |
from agent.memory import Memory
|
13 |
from agent.tools import call_plugin, list_of_all_functions
|
|
|
15 |
has_chinese_chars, save_text_to_file)
|
16 |
from utils import service, extract_and_cache_document, code_interpreter_ws, cache_root, max_ref_token, max_days, download_root
|
17 |
|
18 |
+
llm = ChatAsOAI(model="gpt-3.5-turbo")
|
19 |
|
20 |
mem = Memory(llm=llm, stream=False)
|
21 |
|