Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,189 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
#import subprocess
|
| 3 |
+
from huggingface_hub import InferenceClient
|
| 4 |
+
import gradio as gr
|
| 5 |
+
client = InferenceClient(
|
| 6 |
+
"mistralai/Mixtral-8x7B-Instruct-v0.1"
|
| 7 |
+
)
|
| 8 |
+
|
| 9 |
+
def parse_action(string: str):
|
| 10 |
+
assert string.startswith("action:")
|
| 11 |
+
idx = string.find("action_input=")
|
| 12 |
+
if idx == -1:
|
| 13 |
+
return string[8:], None
|
| 14 |
+
return string[8 : idx - 1], string[idx + 13 :].strip("'").strip('"')
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
from prompts import (
|
| 18 |
+
ACTION_PROMPT,
|
| 19 |
+
COMPRESS_HISTORY_PROMPT,
|
| 20 |
+
LOG_PROMPT,
|
| 21 |
+
LOG_RESPONSE,
|
| 22 |
+
PREFIX,
|
| 23 |
+
TASK_PROMPT,
|
| 24 |
+
)
|
| 25 |
+
|
| 26 |
+
VERBOSE = True
|
| 27 |
+
MAX_HISTORY = 100
|
| 28 |
+
#MODEL = "gpt-3.5-turbo" # "gpt-4"
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
def format_prompt(message, history):
|
| 32 |
+
prompt = "<s>"
|
| 33 |
+
for user_prompt, bot_response in history:
|
| 34 |
+
prompt += f"[INST] {user_prompt} [/INST]"
|
| 35 |
+
prompt += f" {bot_response}</s> "
|
| 36 |
+
prompt += f"[INST] {message} [/INST]"
|
| 37 |
+
return prompt
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
def run_gpt(
|
| 42 |
+
prompt_template,
|
| 43 |
+
stop_tokens,
|
| 44 |
+
max_tokens,
|
| 45 |
+
module_summary,
|
| 46 |
+
purpose,
|
| 47 |
+
**prompt_kwargs,
|
| 48 |
+
):
|
| 49 |
+
|
| 50 |
+
generate_kwargs = dict(
|
| 51 |
+
temperature=0.9,
|
| 52 |
+
max_new_tokens=1048,
|
| 53 |
+
top_p=0.95,
|
| 54 |
+
repetition_penalty=1.0,
|
| 55 |
+
do_sample=True,
|
| 56 |
+
seed=1,
|
| 57 |
+
)
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
content = PREFIX.format(
|
| 61 |
+
purpose=purpose,
|
| 62 |
+
) + prompt_template.format(**prompt_kwargs)
|
| 63 |
+
if VERBOSE:
|
| 64 |
+
print(LOG_PROMPT.format(content))
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
#formatted_prompt = format_prompt(f"{system_prompt}, {prompt}", history)
|
| 68 |
+
#formatted_prompt = format_prompt(f'{content}', history)
|
| 69 |
+
|
| 70 |
+
stream = client.text_generation(content, **generate_kwargs, stream=True, details=True, return_full_text=False)
|
| 71 |
+
resp = ""
|
| 72 |
+
for response in stream:
|
| 73 |
+
resp += response.token.text
|
| 74 |
+
#yield resp
|
| 75 |
+
|
| 76 |
+
if VERBOSE:
|
| 77 |
+
print(LOG_RESPONSE.format(resp))
|
| 78 |
+
return resp
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
def compress_history(purpose, task, history):
|
| 82 |
+
resp = run_gpt(
|
| 83 |
+
COMPRESS_HISTORY_PROMPT,
|
| 84 |
+
stop_tokens=["observation:", "task:", "action:", "thought:"],
|
| 85 |
+
max_tokens=512,
|
| 86 |
+
purpose=purpose,
|
| 87 |
+
task=task,
|
| 88 |
+
history=history,
|
| 89 |
+
)
|
| 90 |
+
history = "observation: {}\n".format(resp)
|
| 91 |
+
return history
|
| 92 |
+
|
| 93 |
+
|
| 94 |
+
def call_main(purpose, task, history, action_input):
|
| 95 |
+
resp = run_gpt(
|
| 96 |
+
ACTION_PROMPT,
|
| 97 |
+
stop_tokens=["observation:", "task:"],
|
| 98 |
+
max_tokens=256,
|
| 99 |
+
purpose=purpose,
|
| 100 |
+
task=task,
|
| 101 |
+
history=history,
|
| 102 |
+
)
|
| 103 |
+
lines = resp.strip().strip("\n").split("\n")
|
| 104 |
+
for line in lines:
|
| 105 |
+
if line == "":
|
| 106 |
+
continue
|
| 107 |
+
if line.startswith("thought: "):
|
| 108 |
+
history += "{}\n".format(line)
|
| 109 |
+
elif line.startswith("action: "):
|
| 110 |
+
action_name, action_input = parse_action(line)
|
| 111 |
+
history += "{}\n".format(line)
|
| 112 |
+
return action_name, action_input, history, task
|
| 113 |
+
else:
|
| 114 |
+
assert False, "unknown action: {}".format(line)
|
| 115 |
+
return "MAIN", None, history, task
|
| 116 |
+
|
| 117 |
+
|
| 118 |
+
def call_set_task(purpose, task, history, action_input):
|
| 119 |
+
task = run_gpt(
|
| 120 |
+
TASK_PROMPT,
|
| 121 |
+
stop_tokens=[],
|
| 122 |
+
max_tokens=64,
|
| 123 |
+
purpose=purpose,
|
| 124 |
+
task=task,
|
| 125 |
+
history=history,
|
| 126 |
+
).strip("\n")
|
| 127 |
+
history += "observation: task has been updated to: {}\n".format(task)
|
| 128 |
+
return "MAIN", None, history, task
|
| 129 |
+
|
| 130 |
+
|
| 131 |
+
NAME_TO_FUNC = {
|
| 132 |
+
"MAIN": call_main,
|
| 133 |
+
"UPDATE-TASK": call_set_task,
|
| 134 |
+
"SEARCH": call_search,
|
| 135 |
+
|
| 136 |
+
}
|
| 137 |
+
|
| 138 |
+
|
| 139 |
+
def run_action(purpose, task, history, action_name, action_input):
|
| 140 |
+
if action_name == "COMPLETE":
|
| 141 |
+
exit(0)
|
| 142 |
+
|
| 143 |
+
# compress the history when it is long
|
| 144 |
+
if len(history.split("\n")) > MAX_HISTORY:
|
| 145 |
+
if VERBOSE:
|
| 146 |
+
print("COMPRESSING HISTORY")
|
| 147 |
+
history = compress_history(purpose, task, history)
|
| 148 |
+
|
| 149 |
+
assert action_name in NAME_TO_FUNC
|
| 150 |
+
|
| 151 |
+
print("RUN: ", action_name, action_input)
|
| 152 |
+
return NAME_TO_FUNC[action_name](purpose, task, history, action_input)
|
| 153 |
+
|
| 154 |
+
|
| 155 |
+
def run(purpose,history):
|
| 156 |
+
task=None
|
| 157 |
+
history = ""
|
| 158 |
+
#if not history:
|
| 159 |
+
# history = []
|
| 160 |
+
action_name = "UPDATE-TASK" if task is None else "MAIN"
|
| 161 |
+
action_input = None
|
| 162 |
+
while True:
|
| 163 |
+
print("")
|
| 164 |
+
print("")
|
| 165 |
+
print("---")
|
| 166 |
+
print("purpose:", purpose)
|
| 167 |
+
print("task:", task)
|
| 168 |
+
print("---")
|
| 169 |
+
print(history)
|
| 170 |
+
print("---")
|
| 171 |
+
|
| 172 |
+
action_name, action_input, history, task = run_action(
|
| 173 |
+
purpose,
|
| 174 |
+
task,
|
| 175 |
+
history,
|
| 176 |
+
action_name,
|
| 177 |
+
action_input,
|
| 178 |
+
)
|
| 179 |
+
|
| 180 |
+
|
| 181 |
+
gr.ChatInterface(
|
| 182 |
+
fn=run,
|
| 183 |
+
chatbot=gr.Chatbot(show_label=False, show_share_button=False, show_copy_button=True, likeable=True, layout="panel"),
|
| 184 |
+
title="Mixtral 46.7B\nMicro-Agent\nInternet Search <br> development test",
|
| 185 |
+
examples=examples,
|
| 186 |
+
concurrency_limit=20,
|
| 187 |
+
).launch(show_api=False)
|
| 188 |
+
|
| 189 |
+
|