new backend
Browse files- OpenAIChatAtomicFlow.py +43 -57
- OpenAIChatAtomicFlow.yaml +21 -19
- run.py +15 -7
OpenAIChatAtomicFlow.py
CHANGED
@@ -1,40 +1,42 @@
|
|
1 |
from copy import deepcopy
|
2 |
-
|
3 |
import hydra
|
4 |
|
5 |
import time
|
6 |
|
7 |
from typing import Dict, Optional, Any
|
8 |
|
9 |
-
from langchain import PromptTemplate
|
10 |
-
from langchain.schema import HumanMessage, AIMessage, SystemMessage
|
11 |
-
|
12 |
from flows.base_flows import AtomicFlow
|
13 |
from flows.datasets import GenericDemonstrationsDataset
|
14 |
|
15 |
from flows.utils import logging
|
16 |
from flows.messages.flow_message import UpdateMessage_ChatMessage
|
17 |
|
|
|
|
|
|
|
|
|
18 |
log = logging.get_logger(__name__)
|
19 |
|
20 |
|
21 |
class OpenAIChatAtomicFlow(AtomicFlow):
|
22 |
-
REQUIRED_KEYS_CONFIG = ["
|
23 |
|
24 |
SUPPORTS_CACHING: bool = True
|
25 |
|
26 |
-
system_message_prompt_template:
|
27 |
-
human_message_prompt_template:
|
28 |
|
29 |
-
|
|
|
30 |
demonstrations: GenericDemonstrationsDataset = None
|
31 |
demonstrations_k: Optional[int] = None
|
32 |
-
demonstrations_response_prompt_template:
|
33 |
|
34 |
def __init__(self,
|
35 |
system_message_prompt_template,
|
36 |
human_message_prompt_template,
|
37 |
init_human_message_prompt_template,
|
|
|
38 |
demonstrations_response_prompt_template=None,
|
39 |
demonstrations=None,
|
40 |
**kwargs):
|
@@ -45,7 +47,7 @@ class OpenAIChatAtomicFlow(AtomicFlow):
|
|
45 |
self.demonstrations_response_prompt_template = demonstrations_response_prompt_template
|
46 |
self.demonstrations = demonstrations
|
47 |
self.demonstrations_k = self.flow_config.get("demonstrations_k", None)
|
48 |
-
|
49 |
assert self.flow_config["name"] not in [
|
50 |
"system",
|
51 |
"user",
|
@@ -59,20 +61,29 @@ class OpenAIChatAtomicFlow(AtomicFlow):
|
|
59 |
@classmethod
|
60 |
def _set_up_prompts(cls, config):
|
61 |
kwargs = {}
|
62 |
-
|
63 |
kwargs["system_message_prompt_template"] = \
|
64 |
hydra.utils.instantiate(config['system_message_prompt_template'], _convert_="partial")
|
65 |
kwargs["init_human_message_prompt_template"] = \
|
66 |
hydra.utils.instantiate(config['init_human_message_prompt_template'], _convert_="partial")
|
67 |
kwargs["human_message_prompt_template"] = \
|
68 |
hydra.utils.instantiate(config['human_message_prompt_template'], _convert_="partial")
|
69 |
-
|
70 |
if "demonstrations_response_prompt_template" in config:
|
71 |
kwargs["demonstrations_response_prompt_template"] = \
|
72 |
hydra.utils.instantiate(config['demonstrations_response_prompt_template'], _convert_="partial")
|
73 |
kwargs["demonstrations"] = GenericDemonstrationsDataset(**config['demonstrations'])
|
74 |
|
75 |
return kwargs
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
76 |
|
77 |
@classmethod
|
78 |
def instantiate_from_config(cls, config):
|
@@ -82,6 +93,7 @@ class OpenAIChatAtomicFlow(AtomicFlow):
|
|
82 |
|
83 |
# ~~~ Set up prompts ~~~
|
84 |
kwargs.update(cls._set_up_prompts(flow_config))
|
|
|
85 |
|
86 |
# ~~~ Instantiate flow ~~~
|
87 |
return cls(**kwargs)
|
@@ -106,7 +118,6 @@ class OpenAIChatAtomicFlow(AtomicFlow):
|
|
106 |
template_kwargs = {}
|
107 |
for input_variable in prompt_template.input_variables:
|
108 |
template_kwargs[input_variable] = input_data[input_variable]
|
109 |
-
|
110 |
msg_content = prompt_template.format(**template_kwargs)
|
111 |
return msg_content
|
112 |
|
@@ -140,19 +151,16 @@ class OpenAIChatAtomicFlow(AtomicFlow):
|
|
140 |
role: str,
|
141 |
content: str) -> None:
|
142 |
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
elif role == self.flow_config["assistant_name"]:
|
149 |
-
self.flow_state["previous_messages"].append(AIMessage(content=content))
|
150 |
else:
|
151 |
raise Exception(f"Invalid role: `{role}`.\n"
|
152 |
f"Role should be one of: "
|
153 |
-
f"`{
|
154 |
-
|
155 |
-
f"`{self.flow_config['assistant_name']}`")
|
156 |
|
157 |
# Log the update to the flow messages list
|
158 |
chat_message = UpdateMessage_ChatMessage(
|
@@ -174,49 +182,24 @@ class OpenAIChatAtomicFlow(AtomicFlow):
|
|
174 |
return all_messages[:first_k] + all_messages[-last_k:]
|
175 |
elif first_k:
|
176 |
return all_messages[:first_k]
|
177 |
-
|
178 |
return all_messages[-last_k:]
|
179 |
|
180 |
def _call(self):
|
181 |
-
|
182 |
-
api_key = api_information.api_key
|
183 |
-
|
184 |
-
if api_information.backend_used == 'azure':
|
185 |
-
from backends.azure_openai import SafeAzureChatOpenAI
|
186 |
-
endpoint = api_information.endpoint
|
187 |
-
backend = SafeAzureChatOpenAI(
|
188 |
-
openai_api_type='azure',
|
189 |
-
openai_api_key=api_key,
|
190 |
-
openai_api_base=endpoint,
|
191 |
-
openai_api_version='2023-05-15',
|
192 |
-
deployment_name=self.flow_config["model_name"],
|
193 |
-
**self.flow_config["generation_parameters"],
|
194 |
-
)
|
195 |
-
elif api_information.backend_used == 'openai':
|
196 |
-
from backends.openai import SafeChatOpenAI
|
197 |
-
backend = SafeChatOpenAI(
|
198 |
-
model_name=self.flow_config["model_name"],
|
199 |
-
openai_api_key=api_key,
|
200 |
-
openai_api_type="open_ai",
|
201 |
-
**self.flow_config["generation_parameters"],
|
202 |
-
)
|
203 |
-
else:
|
204 |
-
raise ValueError(f"Unsupported backend: {api_information.backend_used}")
|
205 |
-
|
206 |
messages = self._get_previous_messages()
|
207 |
-
|
208 |
_success = False
|
209 |
attempts = 1
|
210 |
error = None
|
211 |
response = None
|
212 |
while attempts <= self.flow_config['n_api_retries']:
|
213 |
try:
|
214 |
-
response = backend(messages)
|
|
|
215 |
_success = True
|
216 |
break
|
217 |
except Exception as e:
|
218 |
log.error(
|
219 |
-
f"Error {attempts} in calling backend: {e}.
|
220 |
f"Retrying in {self.flow_config['wait_time_between_retries']} seconds..."
|
221 |
)
|
222 |
# log.error(
|
@@ -226,7 +209,7 @@ class OpenAIChatAtomicFlow(AtomicFlow):
|
|
226 |
attempts += 1
|
227 |
time.sleep(self.flow_config['wait_time_between_retries'])
|
228 |
error = e
|
229 |
-
|
230 |
if not _success:
|
231 |
raise error
|
232 |
|
@@ -266,9 +249,12 @@ class OpenAIChatAtomicFlow(AtomicFlow):
|
|
266 |
|
267 |
# ~~~ Call ~~~
|
268 |
response = self._call()
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
|
|
|
|
|
|
|
273 |
|
274 |
return {"api_output": response}
|
|
|
1 |
from copy import deepcopy
|
|
|
2 |
import hydra
|
3 |
|
4 |
import time
|
5 |
|
6 |
from typing import Dict, Optional, Any
|
7 |
|
|
|
|
|
|
|
8 |
from flows.base_flows import AtomicFlow
|
9 |
from flows.datasets import GenericDemonstrationsDataset
|
10 |
|
11 |
from flows.utils import logging
|
12 |
from flows.messages.flow_message import UpdateMessage_ChatMessage
|
13 |
|
14 |
+
from flows.prompt_template import JinjaPrompt
|
15 |
+
|
16 |
+
from backends.llm_lite import LiteLLMBackend
|
17 |
+
|
18 |
log = logging.get_logger(__name__)
|
19 |
|
20 |
|
21 |
class OpenAIChatAtomicFlow(AtomicFlow):
|
22 |
+
REQUIRED_KEYS_CONFIG = ["backend"]
|
23 |
|
24 |
SUPPORTS_CACHING: bool = True
|
25 |
|
26 |
+
system_message_prompt_template: JinjaPrompt
|
27 |
+
human_message_prompt_template: JinjaPrompt
|
28 |
|
29 |
+
backend: LiteLLMBackend
|
30 |
+
init_human_message_prompt_template: Optional[JinjaPrompt] = None
|
31 |
demonstrations: GenericDemonstrationsDataset = None
|
32 |
demonstrations_k: Optional[int] = None
|
33 |
+
demonstrations_response_prompt_template: str = None
|
34 |
|
35 |
def __init__(self,
|
36 |
system_message_prompt_template,
|
37 |
human_message_prompt_template,
|
38 |
init_human_message_prompt_template,
|
39 |
+
backend,
|
40 |
demonstrations_response_prompt_template=None,
|
41 |
demonstrations=None,
|
42 |
**kwargs):
|
|
|
47 |
self.demonstrations_response_prompt_template = demonstrations_response_prompt_template
|
48 |
self.demonstrations = demonstrations
|
49 |
self.demonstrations_k = self.flow_config.get("demonstrations_k", None)
|
50 |
+
self.backend = backend
|
51 |
assert self.flow_config["name"] not in [
|
52 |
"system",
|
53 |
"user",
|
|
|
61 |
@classmethod
|
62 |
def _set_up_prompts(cls, config):
|
63 |
kwargs = {}
|
64 |
+
|
65 |
kwargs["system_message_prompt_template"] = \
|
66 |
hydra.utils.instantiate(config['system_message_prompt_template'], _convert_="partial")
|
67 |
kwargs["init_human_message_prompt_template"] = \
|
68 |
hydra.utils.instantiate(config['init_human_message_prompt_template'], _convert_="partial")
|
69 |
kwargs["human_message_prompt_template"] = \
|
70 |
hydra.utils.instantiate(config['human_message_prompt_template'], _convert_="partial")
|
71 |
+
|
72 |
if "demonstrations_response_prompt_template" in config:
|
73 |
kwargs["demonstrations_response_prompt_template"] = \
|
74 |
hydra.utils.instantiate(config['demonstrations_response_prompt_template'], _convert_="partial")
|
75 |
kwargs["demonstrations"] = GenericDemonstrationsDataset(**config['demonstrations'])
|
76 |
|
77 |
return kwargs
|
78 |
+
|
79 |
+
@classmethod
|
80 |
+
def _set_up_backend(cls, config):
|
81 |
+
kwargs = {}
|
82 |
+
|
83 |
+
kwargs["backend"] = \
|
84 |
+
hydra.utils.instantiate(config['backend'], _convert_="partial")
|
85 |
+
|
86 |
+
return kwargs
|
87 |
|
88 |
@classmethod
|
89 |
def instantiate_from_config(cls, config):
|
|
|
93 |
|
94 |
# ~~~ Set up prompts ~~~
|
95 |
kwargs.update(cls._set_up_prompts(flow_config))
|
96 |
+
kwargs.update(cls._set_up_backend(flow_config))
|
97 |
|
98 |
# ~~~ Instantiate flow ~~~
|
99 |
return cls(**kwargs)
|
|
|
118 |
template_kwargs = {}
|
119 |
for input_variable in prompt_template.input_variables:
|
120 |
template_kwargs[input_variable] = input_data[input_variable]
|
|
|
121 |
msg_content = prompt_template.format(**template_kwargs)
|
122 |
return msg_content
|
123 |
|
|
|
151 |
role: str,
|
152 |
content: str) -> None:
|
153 |
|
154 |
+
|
155 |
+
acceptable_roles = [self.flow_config["system_name"],self.flow_config["user_name"],self.flow_config["assistant_name"]]
|
156 |
+
if role in acceptable_roles:
|
157 |
+
self.flow_state["previous_messages"].append({"role": role , "content": content})
|
158 |
+
|
|
|
|
|
159 |
else:
|
160 |
raise Exception(f"Invalid role: `{role}`.\n"
|
161 |
f"Role should be one of: "
|
162 |
+
f"`{acceptable_roles}`, ")
|
163 |
+
|
|
|
164 |
|
165 |
# Log the update to the flow messages list
|
166 |
chat_message = UpdateMessage_ChatMessage(
|
|
|
182 |
return all_messages[:first_k] + all_messages[-last_k:]
|
183 |
elif first_k:
|
184 |
return all_messages[:first_k]
|
|
|
185 |
return all_messages[-last_k:]
|
186 |
|
187 |
def _call(self):
|
188 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
189 |
messages = self._get_previous_messages()
|
|
|
190 |
_success = False
|
191 |
attempts = 1
|
192 |
error = None
|
193 |
response = None
|
194 |
while attempts <= self.flow_config['n_api_retries']:
|
195 |
try:
|
196 |
+
response = self.backend(messages=messages,mock_response=False) #set mock_response to True when debugging (fake API request)
|
197 |
+
response = [ answer["content"] for answer in response] # because n in the generation parameters can be > 1
|
198 |
_success = True
|
199 |
break
|
200 |
except Exception as e:
|
201 |
log.error(
|
202 |
+
f"Error {attempts} in calling backend: {e}. "
|
203 |
f"Retrying in {self.flow_config['wait_time_between_retries']} seconds..."
|
204 |
)
|
205 |
# log.error(
|
|
|
209 |
attempts += 1
|
210 |
time.sleep(self.flow_config['wait_time_between_retries'])
|
211 |
error = e
|
212 |
+
|
213 |
if not _success:
|
214 |
raise error
|
215 |
|
|
|
249 |
|
250 |
# ~~~ Call ~~~
|
251 |
response = self._call()
|
252 |
+
|
253 |
+
#loop is in case there was more than one answer (n>1 in generation parameters)
|
254 |
+
for answer in response:
|
255 |
+
self._state_update_add_chat_message(
|
256 |
+
role=self.flow_config["assistant_name"],
|
257 |
+
content=answer
|
258 |
+
)
|
259 |
|
260 |
return {"api_output": response}
|
OpenAIChatAtomicFlow.yaml
CHANGED
@@ -1,17 +1,6 @@
|
|
1 |
# This is an abstract flow, therefore some required fields are not defined (and must be defined by the concrete flow)
|
2 |
enable_cache: True
|
3 |
|
4 |
-
model_name: "gpt-4"
|
5 |
-
generation_parameters:
|
6 |
-
n: 1
|
7 |
-
max_tokens: 2000
|
8 |
-
temperature: 0.3
|
9 |
-
|
10 |
-
model_kwargs:
|
11 |
-
top_p: 0.2
|
12 |
-
frequency_penalty: 0
|
13 |
-
presence_penalty: 0
|
14 |
-
|
15 |
n_api_retries: 6
|
16 |
wait_time_between_retries: 20
|
17 |
|
@@ -19,26 +8,39 @@ system_name: system
|
|
19 |
user_name: user
|
20 |
assistant_name: assistant
|
21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
system_message_prompt_template:
|
23 |
-
_target_:
|
24 |
-
|
25 |
|
26 |
init_human_message_prompt_template:
|
27 |
-
_target_:
|
28 |
-
template_format: jinja2
|
29 |
|
30 |
human_message_prompt_template:
|
31 |
-
_target_:
|
32 |
template: "{{query}}"
|
33 |
input_variables:
|
34 |
- "query"
|
35 |
-
template_format: jinja2
|
36 |
input_interface_initialized:
|
37 |
- "query"
|
38 |
|
39 |
query_message_prompt_template:
|
40 |
-
_target_:
|
41 |
-
|
42 |
|
43 |
previous_messages:
|
44 |
first_k: null # Note that the first message is the system prompt
|
|
|
1 |
# This is an abstract flow, therefore some required fields are not defined (and must be defined by the concrete flow)
|
2 |
enable_cache: True
|
3 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
n_api_retries: 6
|
5 |
wait_time_between_retries: 20
|
6 |
|
|
|
8 |
user_name: user
|
9 |
assistant_name: assistant
|
10 |
|
11 |
+
backend:
|
12 |
+
_target_: backends.llm_lite.LiteLLMBackend
|
13 |
+
api_infos: ???
|
14 |
+
model_name: "gpt-3.5-turbo"
|
15 |
+
n: 1
|
16 |
+
max_tokens: 2000
|
17 |
+
temperature: 0.3
|
18 |
+
|
19 |
+
|
20 |
+
top_p: 0.2
|
21 |
+
frequency_penalty: 0
|
22 |
+
presence_penalty: 0
|
23 |
+
stream: True
|
24 |
+
|
25 |
+
|
26 |
system_message_prompt_template:
|
27 |
+
_target_: flows.prompt_template.JinjaPrompt
|
28 |
+
|
29 |
|
30 |
init_human_message_prompt_template:
|
31 |
+
_target_: flows.prompt_template.JinjaPrompt
|
|
|
32 |
|
33 |
human_message_prompt_template:
|
34 |
+
_target_: flows.prompt_template.JinjaPrompt
|
35 |
template: "{{query}}"
|
36 |
input_variables:
|
37 |
- "query"
|
|
|
38 |
input_interface_initialized:
|
39 |
- "query"
|
40 |
|
41 |
query_message_prompt_template:
|
42 |
+
_target_: flows.prompt_template.JinjaPrompt
|
43 |
+
|
44 |
|
45 |
previous_messages:
|
46 |
first_k: null # Note that the first message is the system prompt
|
run.py
CHANGED
@@ -3,7 +3,8 @@ import os
|
|
3 |
import hydra
|
4 |
|
5 |
import flows
|
6 |
-
from flows.flow_launchers import FlowLauncher
|
|
|
7 |
from flows.utils.general_helpers import read_yaml_file
|
8 |
|
9 |
from flows import logging
|
@@ -23,25 +24,33 @@ flow_verse.sync_dependencies(dependencies)
|
|
23 |
if __name__ == "__main__":
|
24 |
# ~~~ Set the API information ~~~
|
25 |
# OpenAI backend
|
26 |
-
|
27 |
-
|
28 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
|
30 |
root_dir = "."
|
31 |
cfg_path = os.path.join(root_dir, "SimpleQA.yaml")
|
32 |
cfg = read_yaml_file(cfg_path)
|
33 |
|
|
|
|
|
34 |
# ~~~ Instantiate the Flow ~~~
|
35 |
flow_with_interfaces = {
|
36 |
"flow": hydra.utils.instantiate(cfg['flow'], _recursive_=False, _convert_="partial"),
|
37 |
"input_interface": (
|
38 |
None
|
39 |
-
if
|
40 |
else hydra.utils.instantiate(cfg['input_interface'], _recursive_=False)
|
41 |
),
|
42 |
"output_interface": (
|
43 |
None
|
44 |
-
if
|
45 |
else hydra.utils.instantiate(cfg['output_interface'], _recursive_=False)
|
46 |
),
|
47 |
}
|
@@ -58,7 +67,6 @@ if __name__ == "__main__":
|
|
58 |
flow_with_interfaces=flow_with_interfaces,
|
59 |
data=data,
|
60 |
path_to_output_file=path_to_output_file,
|
61 |
-
api_information=api_information,
|
62 |
)
|
63 |
|
64 |
# ~~~ Print the output ~~~
|
|
|
3 |
import hydra
|
4 |
|
5 |
import flows
|
6 |
+
from flows.flow_launchers import FlowLauncher
|
7 |
+
from backends.api_info import ApiInfo
|
8 |
from flows.utils.general_helpers import read_yaml_file
|
9 |
|
10 |
from flows import logging
|
|
|
24 |
if __name__ == "__main__":
|
25 |
# ~~~ Set the API information ~~~
|
26 |
# OpenAI backend
|
27 |
+
|
28 |
+
api_information = [ApiInfo(backend_used="openai",
|
29 |
+
api_key = os.getenv("OPENAI_API_KEY"))]
|
30 |
+
|
31 |
+
# # Azure backend
|
32 |
+
# api_information = ApiInfo(backend_used = "azure",
|
33 |
+
# api_base = os.getenv("AZURE_API_BASE"),
|
34 |
+
# api_key = os.getenv("AZURE_OPENAI_KEY"),
|
35 |
+
# api_version = os.getenv("AZURE_API_VERSION") )
|
36 |
|
37 |
root_dir = "."
|
38 |
cfg_path = os.path.join(root_dir, "SimpleQA.yaml")
|
39 |
cfg = read_yaml_file(cfg_path)
|
40 |
|
41 |
+
cfg["flow"]["backend"]["api_infos"] = api_information
|
42 |
+
# ~~~ Instantiate the Flow ~~~
|
43 |
# ~~~ Instantiate the Flow ~~~
|
44 |
flow_with_interfaces = {
|
45 |
"flow": hydra.utils.instantiate(cfg['flow'], _recursive_=False, _convert_="partial"),
|
46 |
"input_interface": (
|
47 |
None
|
48 |
+
if cfg.get( "input_interface", None) is None
|
49 |
else hydra.utils.instantiate(cfg['input_interface'], _recursive_=False)
|
50 |
),
|
51 |
"output_interface": (
|
52 |
None
|
53 |
+
if cfg.get( "output_interface", None) is None
|
54 |
else hydra.utils.instantiate(cfg['output_interface'], _recursive_=False)
|
55 |
),
|
56 |
}
|
|
|
67 |
flow_with_interfaces=flow_with_interfaces,
|
68 |
data=data,
|
69 |
path_to_output_file=path_to_output_file,
|
|
|
70 |
)
|
71 |
|
72 |
# ~~~ Print the output ~~~
|