xu song
commited on
Commit
•
2fa4e4c
1
Parent(s):
c619300
update
Browse files- README.md +4 -1
- app.py +30 -22
- app_util.py +12 -13
- config.py +2 -1
- models/cpp_qwen2.py +31 -17
- models/hf_qwen2.py +11 -8
README.md
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
---
|
2 |
title: Self Chat
|
3 |
-
emoji:
|
4 |
colorFrom: yellow
|
5 |
colorTo: purple
|
6 |
sdk: gradio
|
@@ -8,6 +8,9 @@ sdk_version: 4.39.0
|
|
8 |
app_file: app.py
|
9 |
pinned: false
|
10 |
license: apache-2.0
|
|
|
|
|
|
|
11 |
---
|
12 |
|
13 |
An example chatbot using [Gradio](https://gradio.app), [`huggingface_hub`](https://huggingface.co/docs/huggingface_hub/v0.22.2/en/index), and the [Hugging Face Inference API](https://huggingface.co/docs/api-inference/index).
|
|
|
1 |
---
|
2 |
title: Self Chat
|
3 |
+
emoji: 🤖🤖
|
4 |
colorFrom: yellow
|
5 |
colorTo: purple
|
6 |
sdk: gradio
|
|
|
8 |
app_file: app.py
|
9 |
pinned: false
|
10 |
license: apache-2.0
|
11 |
+
tags:
|
12 |
+
- chatbot
|
13 |
+
short_description: Generating synthetic data via self-chat
|
14 |
---
|
15 |
|
16 |
An example chatbot using [Gradio](https://gradio.app), [`huggingface_hub`](https://huggingface.co/docs/huggingface_hub/v0.22.2/en/index), and the [Hugging Face Inference API](https://huggingface.co/docs/api-inference/index).
|
app.py
CHANGED
@@ -1,19 +1,29 @@
|
|
1 |
"""
|
2 |
"""
|
|
|
3 |
import gradio
|
4 |
-
|
5 |
import config
|
6 |
from app_util import *
|
7 |
|
8 |
-
|
9 |
-
|
10 |
user_simulator_doc = """\
|
|
|
|
|
11 |
There are maily two types of user simulator:
|
12 |
- prompt-based user-simulator (role-play)
|
13 |
- model-based user-simulator
|
14 |
|
15 |
-
|
16 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
"""
|
18 |
|
19 |
survey = """\
|
@@ -28,16 +38,16 @@ Essentially, it is a form of model compression.
|
|
28 |
## 有不用概率的知识蒸馏吗?
|
29 |
"""
|
30 |
|
31 |
-
with gr.Blocks() as demo:
|
32 |
# Knowledge Distillation through Self Chatting
|
33 |
# Distilling the Knowledge from LLM through Self Chatting
|
34 |
# Generating Synthetic Data through Self Chat
|
35 |
-
gr.HTML("""<h1 align="center">Generating Synthetic Data
|
36 |
with gr.Row():
|
37 |
with gr.Column(scale=5):
|
38 |
system = gr.Dropdown(
|
39 |
choices=system_list,
|
40 |
-
value=system_list[0],
|
41 |
allow_custom_value=True,
|
42 |
interactive=True,
|
43 |
label="System message",
|
@@ -46,7 +56,8 @@ with gr.Blocks() as demo:
|
|
46 |
|
47 |
chatbot = gr.Chatbot(show_copy_button=True,
|
48 |
show_share_button=True,
|
49 |
-
avatar_images=("assets/man.png", "assets/bot.png")
|
|
|
50 |
|
51 |
# gr.Textbox("For faster inference, you can build locally with ")
|
52 |
# ss
|
@@ -54,30 +65,27 @@ with gr.Blocks() as demo:
|
|
54 |
input_text_1 = gr.Textbox(show_label=False, placeholder="...", lines=10, visible=False)
|
55 |
generate_btn = gr.Button("🤔️ Self-Chat", variant="primary")
|
56 |
with gr.Row():
|
57 |
-
retry_btn = gr.Button("🔄
|
58 |
undo_btn = gr.Button("↩️ Undo", variant="secondary", size="sm", )
|
59 |
clear_btn = gr.Button("🗑️ Clear", variant="secondary", size="sm", ) # 🧹 Clear History (清除历史)
|
60 |
# stop_btn = gr.Button("停止生成", variant="stop", visible=False)
|
61 |
-
|
62 |
-
# "Self-chat is a demo, which makes the model talk to itself. "
|
63 |
-
# "It is based on user simulator and response generator.",
|
64 |
-
# visible=True)
|
65 |
|
66 |
# 也叫 chat-assistant,
|
67 |
-
with gradio.Tab("Response Generator"
|
68 |
with gr.Row():
|
69 |
-
input_text_2 = gr.Textbox(show_label=False, placeholder="Please type
|
70 |
generate_btn_2 = gr.Button("Send", variant="primary")
|
71 |
with gr.Row():
|
72 |
retry_btn_2 = gr.Button("🔄 Regenerate", variant="secondary", size="sm", )
|
73 |
undo_btn_2 = gr.Button("↩️ Undo", variant="secondary", size="sm", )
|
74 |
clear_btn_2 = gr.Button("🗑️ Clear", variant="secondary", size="sm", ) # 🧹 Clear History (清除历史)
|
75 |
-
gr.Markdown(
|
76 |
|
77 |
#
|
78 |
-
with gradio.Tab("User Simulator"
|
79 |
with gr.Row():
|
80 |
-
input_text_3 = gr.Textbox(show_label=False, placeholder="Please type
|
81 |
generate_btn_3 = gr.Button("Send", variant="primary")
|
82 |
with gr.Row():
|
83 |
retry_btn_3 = gr.Button("🔄 Regenerate", variant="secondary", size="sm", )
|
@@ -85,7 +93,7 @@ with gr.Blocks() as demo:
|
|
85 |
clear_btn_3 = gr.Button("🗑️ Clear", variant="secondary", size="sm", ) # 🧹 Clear History (清除历史)
|
86 |
gr.Markdown(user_simulator_doc)
|
87 |
|
88 |
-
with gr.Column(variant="compact"):
|
89 |
# with gr.Column():
|
90 |
model = gr.Dropdown(
|
91 |
["Qwen2-0.5B-Instruct", "llama3.1", "gemini"],
|
@@ -155,8 +163,8 @@ with gr.Blocks() as demo:
|
|
155 |
slider_top_k.change(set_top_k, inputs=[slider_top_k])
|
156 |
|
157 |
|
|
|
158 |
|
159 |
-
|
160 |
-
# demo.queue().launch(share=False, server_name="0.0.0.0")
|
161 |
# demo.queue().launch(concurrency_count=1, max_size=5)
|
162 |
demo.queue().launch()
|
|
|
1 |
"""
|
2 |
"""
|
3 |
+
import random
|
4 |
import gradio
|
|
|
5 |
import config
|
6 |
from app_util import *
|
7 |
|
|
|
|
|
8 |
user_simulator_doc = """\
|
9 |
+
The agent acts as user simulator.
|
10 |
+
|
11 |
There are maily two types of user simulator:
|
12 |
- prompt-based user-simulator (role-play)
|
13 |
- model-based user-simulator
|
14 |
|
15 |
+
This demo is a model-based user simulator.
|
16 |
+
"""
|
17 |
+
# In most cases, large language models (LLMs) are used to serve as assistant generator.
|
18 |
+
# Besides, it can also used as user simulator.
|
19 |
+
|
20 |
+
assistant_simulator_doc = """\
|
21 |
+
The agent acts as assistant simulator.
|
22 |
+
"""
|
23 |
+
|
24 |
+
self_chat_doc = """\
|
25 |
+
Self-chat is a demo which make the model talk to itself.
|
26 |
+
It is a combination of user simulator and response generator.
|
27 |
"""
|
28 |
|
29 |
survey = """\
|
|
|
38 |
## 有不用概率的知识蒸馏吗?
|
39 |
"""
|
40 |
|
41 |
+
with gr.Blocks(head=None) as demo:
|
42 |
# Knowledge Distillation through Self Chatting
|
43 |
# Distilling the Knowledge from LLM through Self Chatting
|
44 |
# Generating Synthetic Data through Self Chat
|
45 |
+
gr.HTML("""<h1 align="center">Generating Synthetic Data via Self-Chat</h1>""")
|
46 |
with gr.Row():
|
47 |
with gr.Column(scale=5):
|
48 |
system = gr.Dropdown(
|
49 |
choices=system_list,
|
50 |
+
# value=system_list[0],
|
51 |
allow_custom_value=True,
|
52 |
interactive=True,
|
53 |
label="System message",
|
|
|
56 |
|
57 |
chatbot = gr.Chatbot(show_copy_button=True,
|
58 |
show_share_button=True,
|
59 |
+
avatar_images=("assets/man.png", "assets/bot.png"),
|
60 |
+
likeable=True)
|
61 |
|
62 |
# gr.Textbox("For faster inference, you can build locally with ")
|
63 |
# ss
|
|
|
65 |
input_text_1 = gr.Textbox(show_label=False, placeholder="...", lines=10, visible=False)
|
66 |
generate_btn = gr.Button("🤔️ Self-Chat", variant="primary")
|
67 |
with gr.Row():
|
68 |
+
retry_btn = gr.Button("🔄 Regenerate", variant="secondary", size="sm", )
|
69 |
undo_btn = gr.Button("↩️ Undo", variant="secondary", size="sm", )
|
70 |
clear_btn = gr.Button("🗑️ Clear", variant="secondary", size="sm", ) # 🧹 Clear History (清除历史)
|
71 |
# stop_btn = gr.Button("停止生成", variant="stop", visible=False)
|
72 |
+
gr.Markdown(self_chat_doc)
|
|
|
|
|
|
|
73 |
|
74 |
# 也叫 chat-assistant,
|
75 |
+
with gradio.Tab("Response Generator"):
|
76 |
with gr.Row():
|
77 |
+
input_text_2 = gr.Textbox(show_label=False, placeholder="Please type user input", scale=7)
|
78 |
generate_btn_2 = gr.Button("Send", variant="primary")
|
79 |
with gr.Row():
|
80 |
retry_btn_2 = gr.Button("🔄 Regenerate", variant="secondary", size="sm", )
|
81 |
undo_btn_2 = gr.Button("↩️ Undo", variant="secondary", size="sm", )
|
82 |
clear_btn_2 = gr.Button("🗑️ Clear", variant="secondary", size="sm", ) # 🧹 Clear History (清除历史)
|
83 |
+
gr.Markdown(assistant_simulator_doc)
|
84 |
|
85 |
#
|
86 |
+
with gradio.Tab("User Simulator"):
|
87 |
with gr.Row():
|
88 |
+
input_text_3 = gr.Textbox(show_label=False, placeholder="Please type assistant response", scale=7)
|
89 |
generate_btn_3 = gr.Button("Send", variant="primary")
|
90 |
with gr.Row():
|
91 |
retry_btn_3 = gr.Button("🔄 Regenerate", variant="secondary", size="sm", )
|
|
|
93 |
clear_btn_3 = gr.Button("🗑️ Clear", variant="secondary", size="sm", ) # 🧹 Clear History (清除历史)
|
94 |
gr.Markdown(user_simulator_doc)
|
95 |
|
96 |
+
with gr.Column(variant="compact", scale=1, min_width=300):
|
97 |
# with gr.Column():
|
98 |
model = gr.Dropdown(
|
99 |
["Qwen2-0.5B-Instruct", "llama3.1", "gemini"],
|
|
|
163 |
slider_top_k.change(set_top_k, inputs=[slider_top_k])
|
164 |
|
165 |
|
166 |
+
demo.load(lambda: gr.update(value=random.choice(system_list)), None, system)
|
167 |
|
168 |
+
# demo.queue().launch(share=False, server_name="0.0.0.0", debug=True)
|
|
|
169 |
# demo.queue().launch(concurrency_count=1, max_size=5)
|
170 |
demo.queue().launch()
|
app_util.py
CHANGED
@@ -1,8 +1,8 @@
|
|
1 |
import json
|
2 |
import gradio as gr
|
3 |
from utils.logging_util import logger
|
4 |
-
from models.cpp_qwen2 import
|
5 |
-
# from models.hf_qwen2 import
|
6 |
|
7 |
|
8 |
#
|
@@ -22,15 +22,16 @@ system_list = [
|
|
22 |
"You are a helpful assistant.",
|
23 |
"你是一个导游。",
|
24 |
"你是一名投资经理。",
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
]
|
32 |
|
33 |
-
|
|
|
34 |
|
35 |
def generate_user_message(chatbot, history):
|
36 |
if history and history[-1]["role"] == "user":
|
@@ -52,7 +53,6 @@ def generate_assistant_message(chatbot, history):
|
|
52 |
auto-mode:query is None
|
53 |
manual-mode:query 是用户输入
|
54 |
"""
|
55 |
-
logger.info(f"generating {json.dumps(history, ensure_ascii=False)}")
|
56 |
user_content = history[-1]["content"]
|
57 |
if history[-1]["role"] != "user":
|
58 |
gr.Warning('You should generate or type user-input first.')
|
@@ -65,13 +65,12 @@ def generate_assistant_message(chatbot, history):
|
|
65 |
|
66 |
assistant_tokens = bot.strip_stoptokens(assistant_tokens)
|
67 |
history.append({"role": "assistant", "content": assistant_content, "tokens": assistant_tokens})
|
68 |
-
print(f"chatbot is {chatbot}")
|
69 |
-
print(f"history is {history}")
|
70 |
yield chatbot, history
|
71 |
|
72 |
|
73 |
def generate(chatbot, history):
|
74 |
-
|
|
|
75 |
streamer = None
|
76 |
if history[-1]["role"] in ["assistant", "system"]:
|
77 |
streamer = generate_user_message(chatbot, history)
|
|
|
1 |
import json
|
2 |
import gradio as gr
|
3 |
from utils.logging_util import logger
|
4 |
+
from models.cpp_qwen2 import Qwen2Simulator as Bot
|
5 |
+
# from models.hf_qwen2 import Qwen2Simulator as Bot
|
6 |
|
7 |
|
8 |
#
|
|
|
22 |
"You are a helpful assistant.",
|
23 |
"你是一个导游。",
|
24 |
"你是一名投资经理。",
|
25 |
+
"你是一名医生。",
|
26 |
+
"你是一个英语老师。",
|
27 |
+
"你是一个程序员。",
|
28 |
+
"你是一个心理咨询师。",
|
29 |
+
"你是一名AI写作助手。"
|
30 |
+
"你是一名作家,擅长写小说。"
|
31 |
]
|
32 |
|
33 |
+
|
34 |
+
bot = Bot(system_list)
|
35 |
|
36 |
def generate_user_message(chatbot, history):
|
37 |
if history and history[-1]["role"] == "user":
|
|
|
53 |
auto-mode:query is None
|
54 |
manual-mode:query 是用户输入
|
55 |
"""
|
|
|
56 |
user_content = history[-1]["content"]
|
57 |
if history[-1]["role"] != "user":
|
58 |
gr.Warning('You should generate or type user-input first.')
|
|
|
65 |
|
66 |
assistant_tokens = bot.strip_stoptokens(assistant_tokens)
|
67 |
history.append({"role": "assistant", "content": assistant_content, "tokens": assistant_tokens})
|
|
|
|
|
68 |
yield chatbot, history
|
69 |
|
70 |
|
71 |
def generate(chatbot, history):
|
72 |
+
request_param = json.dumps({'chatbot': chatbot, 'history': history}, ensure_ascii=False)
|
73 |
+
logger.info(f"request_param: {request_param}")
|
74 |
streamer = None
|
75 |
if history[-1]["role"] in ["assistant", "system"]:
|
76 |
streamer = generate_user_message(chatbot, history)
|
config.py
CHANGED
@@ -1,6 +1,7 @@
|
|
1 |
|
2 |
|
3 |
-
MAX_SEQUENCE_LENGTH = 32768 #
|
|
|
4 |
|
5 |
DEFAULT_MAX_NEW_TOKENS = 128
|
6 |
DEFAULT_TOP_K = 100
|
|
|
1 |
|
2 |
|
3 |
+
# MAX_SEQUENCE_LENGTH = 32768 # 消耗内存太多
|
4 |
+
MAX_SEQUENCE_LENGTH = 8192 #
|
5 |
|
6 |
DEFAULT_MAX_NEW_TOKENS = 128
|
7 |
DEFAULT_TOP_K = 100
|
models/cpp_qwen2.py
CHANGED
@@ -77,7 +77,7 @@ import config
|
|
77 |
|
78 |
class Qwen2Simulator(Simulator):
|
79 |
|
80 |
-
def __init__(self):
|
81 |
local_path = "/workspace/xusong/huggingface/models/Qwen2-0.5B-Instruct-GGUF/qwen2-0_5b-instruct-fp16.gguf"
|
82 |
if os.path.exists(local_path):
|
83 |
self.hf_tokenizer = AutoTokenizer.from_pretrained(
|
@@ -105,30 +105,37 @@ class Qwen2Simulator(Simulator):
|
|
105 |
f"n_threads={self.llm.n_threads}, n_ctx={self.llm.n_ctx}, "
|
106 |
f"env[CACHE]={os.environ.get('CACHE', None)}")
|
107 |
|
108 |
-
|
|
|
|
|
109 |
"<|im_end|>",
|
110 |
"<|im_start|>",
|
111 |
"<|endoftext|>",
|
112 |
]
|
113 |
-
self.
|
|
|
|
|
|
|
|
|
|
|
114 |
self.generation_kwargs = dict(
|
115 |
temperature=config.DEFAULT_TEMPERATURE,
|
116 |
top_p=config.DEFAULT_TOP_P,
|
117 |
top_k=config.DEFAULT_TOP_K,
|
118 |
max_tokens=config.DEFAULT_MAX_NEW_TOKENS,
|
119 |
repeat_penalty=1.1,
|
120 |
-
# qwen2-0.5b-chat 有时内容生成结束没有<|im_end|>,直接跟 <|im_start|>
|
121 |
-
stop=self.stop_words,
|
122 |
)
|
123 |
-
|
124 |
self.user_start_tokens = self.tokenize("<|im_start|>user\n")
|
125 |
self.assistant_start_tokens = self.tokenize("<|im_start|>assistant\n")
|
126 |
# self.llm.generate .set_cache .last_n_tokens_size .reset .ctx ._ctx
|
127 |
|
128 |
# cache = llama_cpp.LlamaDiskCache(capacity_bytes=cache_size)
|
129 |
-
cache = llama_cpp.LlamaRAMCache(capacity_bytes=2 << 30)
|
130 |
self.llm.set_cache(cache)
|
131 |
|
|
|
|
|
|
|
132 |
def tokenize(self, text):
|
133 |
return self.llm.tokenize(text.encode("utf-8"))
|
134 |
|
@@ -136,10 +143,10 @@ class Qwen2Simulator(Simulator):
|
|
136 |
return self.llm.detokenize(tokens).decode("utf-8")
|
137 |
|
138 |
def strip_stoptokens(self, tokens):
|
139 |
-
while tokens and tokens[0] in self.
|
140 |
logger.info(f"head-striping {tokens[0]} {self.detokenize([tokens[0]])}")
|
141 |
tokens.pop(0)
|
142 |
-
while tokens and tokens[-1] in self.
|
143 |
logger.info(f"tail-striping {tokens[-1]} {self.detokenize([tokens[-1]])}")
|
144 |
tokens.pop()
|
145 |
return tokens
|
@@ -154,9 +161,12 @@ class Qwen2Simulator(Simulator):
|
|
154 |
"""
|
155 |
if history[-1]['role'] in ["user"]:
|
156 |
start_tokens = self.assistant_start_tokens
|
|
|
157 |
suffix_tokens = self.user_start_tokens
|
|
|
158 |
elif history[-1]['role'] in ["assistant", "system"]:
|
159 |
start_tokens = self.user_start_tokens
|
|
|
160 |
suffix_tokens = self.assistant_start_tokens
|
161 |
|
162 |
input_ids = []
|
@@ -168,15 +178,16 @@ class Qwen2Simulator(Simulator):
|
|
168 |
+ self.tokenize("<|im_end|>\n")
|
169 |
input_ids += start_tokens
|
170 |
if stream:
|
171 |
-
return self._stream_generate(input_ids, suffix_tokens)
|
172 |
else:
|
173 |
return self._generate(input_ids)
|
174 |
|
175 |
-
def _stream_generate(self, input_ids, suffix_tokens=None):
|
176 |
logger.info(f"generation_kwargs {self.generation_kwargs}")
|
177 |
output = self.llm.create_completion(
|
178 |
input_ids,
|
179 |
stream=True,
|
|
|
180 |
**self.generation_kwargs
|
181 |
)
|
182 |
# TODO: 检测finish reason,如果是length,则shift,并继续生成。
|
@@ -201,37 +212,40 @@ class Qwen2Simulator(Simulator):
|
|
201 |
for system_prompt in system_list:
|
202 |
logger.info(f"pre caching '{system_prompt}'")
|
203 |
input_ids = self.tokenize(f"<|im_start|>system\n{system_prompt}<|im_end|>\n<|im_start|>user\n")
|
204 |
-
|
205 |
input_ids,
|
206 |
stream=False,
|
207 |
max_tokens=1,
|
208 |
top_k=1
|
209 |
)
|
210 |
-
logger.info(
|
211 |
-
|
|
|
212 |
|
213 |
self._disable_cache()
|
214 |
|
215 |
-
|
216 |
def post_cache(self, suffix_tokens):
|
217 |
""" warmup for next turn generation
|
218 |
:param suffix_tokens:
|
219 |
:return:
|
220 |
"""
|
|
|
|
|
221 |
if suffix_tokens:
|
222 |
logger.info(f"before warmup: n_tokens = {self.llm.n_tokens}")
|
223 |
self.llm.eval([151645, 198] + suffix_tokens) # <|im_end|>\n
|
224 |
logger.info(f"after warmup: n_tokens = {self.llm.n_tokens}")
|
225 |
-
|
|
|
226 |
|
227 |
def _disable_cache(self):
|
228 |
llama_cpp.LlamaRAMCache.__setitem__ = lambda *args: None
|
229 |
llama_cpp.Llama.save_state = lambda *args: None
|
230 |
|
231 |
-
bot = Qwen2Simulator()
|
232 |
|
233 |
if __name__ == "__main__":
|
234 |
|
|
|
235 |
messages = [{"role": "system", "content": "你是一个导游。"}]
|
236 |
generated_tokens = None
|
237 |
print("######## requesting", messages)
|
|
|
77 |
|
78 |
class Qwen2Simulator(Simulator):
|
79 |
|
80 |
+
def __init__(self, system_list=None):
|
81 |
local_path = "/workspace/xusong/huggingface/models/Qwen2-0.5B-Instruct-GGUF/qwen2-0_5b-instruct-fp16.gguf"
|
82 |
if os.path.exists(local_path):
|
83 |
self.hf_tokenizer = AutoTokenizer.from_pretrained(
|
|
|
105 |
f"n_threads={self.llm.n_threads}, n_ctx={self.llm.n_ctx}, "
|
106 |
f"env[CACHE]={os.environ.get('CACHE', None)}")
|
107 |
|
108 |
+
|
109 |
+
# qwen2-0.5b-chat 有时内容生成结束没有<|im_end|>,直接跟 <|im_start|>
|
110 |
+
self.assistant_stop_words = [
|
111 |
"<|im_end|>",
|
112 |
"<|im_start|>",
|
113 |
"<|endoftext|>",
|
114 |
]
|
115 |
+
self.assistant_stop_tokens = self.tokenize("".join(self.assistant_stop_words))
|
116 |
+
self.user_stop_words = self.assistant_stop_words + ["?", "?"]
|
117 |
+
self.user_stop_tokens = self.tokenize("".join(self.user_stop_words))
|
118 |
+
logger.info(f"assistant_stop_tokens: {self.assistant_stop_tokens}")
|
119 |
+
logger.info(f"user_stop_tokens: {self.user_stop_tokens}")
|
120 |
+
|
121 |
self.generation_kwargs = dict(
|
122 |
temperature=config.DEFAULT_TEMPERATURE,
|
123 |
top_p=config.DEFAULT_TOP_P,
|
124 |
top_k=config.DEFAULT_TOP_K,
|
125 |
max_tokens=config.DEFAULT_MAX_NEW_TOKENS,
|
126 |
repeat_penalty=1.1,
|
|
|
|
|
127 |
)
|
|
|
128 |
self.user_start_tokens = self.tokenize("<|im_start|>user\n")
|
129 |
self.assistant_start_tokens = self.tokenize("<|im_start|>assistant\n")
|
130 |
# self.llm.generate .set_cache .last_n_tokens_size .reset .ctx ._ctx
|
131 |
|
132 |
# cache = llama_cpp.LlamaDiskCache(capacity_bytes=cache_size)
|
133 |
+
cache = llama_cpp.LlamaRAMCache(capacity_bytes=2 << 30) # 2G
|
134 |
self.llm.set_cache(cache)
|
135 |
|
136 |
+
if system_list is not None:
|
137 |
+
self.pre_cache_system(system_list)
|
138 |
+
|
139 |
def tokenize(self, text):
|
140 |
return self.llm.tokenize(text.encode("utf-8"))
|
141 |
|
|
|
143 |
return self.llm.detokenize(tokens).decode("utf-8")
|
144 |
|
145 |
def strip_stoptokens(self, tokens):
|
146 |
+
while tokens and tokens[0] in self.assistant_stop_tokens:
|
147 |
logger.info(f"head-striping {tokens[0]} {self.detokenize([tokens[0]])}")
|
148 |
tokens.pop(0)
|
149 |
+
while tokens and tokens[-1] in self.assistant_stop_tokens:
|
150 |
logger.info(f"tail-striping {tokens[-1]} {self.detokenize([tokens[-1]])}")
|
151 |
tokens.pop()
|
152 |
return tokens
|
|
|
161 |
"""
|
162 |
if history[-1]['role'] in ["user"]:
|
163 |
start_tokens = self.assistant_start_tokens
|
164 |
+
stop_words = self.assistant_stop_words
|
165 |
suffix_tokens = self.user_start_tokens
|
166 |
+
|
167 |
elif history[-1]['role'] in ["assistant", "system"]:
|
168 |
start_tokens = self.user_start_tokens
|
169 |
+
stop_words = self.user_stop_words
|
170 |
suffix_tokens = self.assistant_start_tokens
|
171 |
|
172 |
input_ids = []
|
|
|
178 |
+ self.tokenize("<|im_end|>\n")
|
179 |
input_ids += start_tokens
|
180 |
if stream:
|
181 |
+
return self._stream_generate(input_ids, stop_words, suffix_tokens)
|
182 |
else:
|
183 |
return self._generate(input_ids)
|
184 |
|
185 |
+
def _stream_generate(self, input_ids, stop_words, suffix_tokens=None):
|
186 |
logger.info(f"generation_kwargs {self.generation_kwargs}")
|
187 |
output = self.llm.create_completion(
|
188 |
input_ids,
|
189 |
stream=True,
|
190 |
+
stop=stop_words,
|
191 |
**self.generation_kwargs
|
192 |
)
|
193 |
# TODO: 检测finish reason,如果是length,则shift,并继续生成。
|
|
|
212 |
for system_prompt in system_list:
|
213 |
logger.info(f"pre caching '{system_prompt}'")
|
214 |
input_ids = self.tokenize(f"<|im_start|>system\n{system_prompt}<|im_end|>\n<|im_start|>user\n")
|
215 |
+
_output = self.llm.create_completion(
|
216 |
input_ids,
|
217 |
stream=False,
|
218 |
max_tokens=1,
|
219 |
top_k=1
|
220 |
)
|
221 |
+
logger.info(
|
222 |
+
f"cache size {self.llm.cache.cache_size}={self.llm.cache.cache_size / 1024 / 1024 / 1024:.2f} GB, "
|
223 |
+
f"process_mem: {psutil.Process(os.getpid()).memory_info().rss / 1024 / 1024 / 1024:.2f} GB")
|
224 |
|
225 |
self._disable_cache()
|
226 |
|
|
|
227 |
def post_cache(self, suffix_tokens):
|
228 |
""" warmup for next turn generation
|
229 |
:param suffix_tokens:
|
230 |
:return:
|
231 |
"""
|
232 |
+
logger.info(f"cache size {self.llm.cache.cache_size}={self.llm.cache.cache_size / 1024 / 1024 / 1024:.2f} GB, "
|
233 |
+
f"process_mem: {psutil.Process(os.getpid()).memory_info().rss / 1024 / 1024 / 1024:.2f} GB")
|
234 |
if suffix_tokens:
|
235 |
logger.info(f"before warmup: n_tokens = {self.llm.n_tokens}")
|
236 |
self.llm.eval([151645, 198] + suffix_tokens) # <|im_end|>\n
|
237 |
logger.info(f"after warmup: n_tokens = {self.llm.n_tokens}")
|
238 |
+
logger.info(f"cache size {self.llm.cache.cache_size}={self.llm.cache.cache_size / 1024 / 1024 / 1024:.2f} GB, "
|
239 |
+
f"process_mem: {psutil.Process(os.getpid()).memory_info().rss / 1024 / 1024 / 1024:.2f} GB")
|
240 |
|
241 |
def _disable_cache(self):
|
242 |
llama_cpp.LlamaRAMCache.__setitem__ = lambda *args: None
|
243 |
llama_cpp.Llama.save_state = lambda *args: None
|
244 |
|
|
|
245 |
|
246 |
if __name__ == "__main__":
|
247 |
|
248 |
+
bot = Qwen2Simulator()
|
249 |
messages = [{"role": "system", "content": "你是一个导游。"}]
|
250 |
generated_tokens = None
|
251 |
print("######## requesting", messages)
|
models/hf_qwen2.py
CHANGED
@@ -14,13 +14,15 @@ class Qwen2Simulator(Simulator):
|
|
14 |
在传递 device_map 时,low_cpu_mem_usage 会自动设置为 True
|
15 |
"""
|
16 |
|
17 |
-
self.tokenizer =
|
18 |
-
self.
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
|
|
|
|
24 |
self.generation_kwargs = dict(
|
25 |
do_sample=True,
|
26 |
temperature=0.7,
|
@@ -93,11 +95,12 @@ class Qwen2Simulator(Simulator):
|
|
93 |
return self.tokenizer.decode(response[0][input_ids_length:], skip_special_tokens=True)
|
94 |
|
95 |
|
96 |
-
|
97 |
# bot = Qwen2Simulator("Qwen/Qwen2-0.5B-Instruct")
|
98 |
|
99 |
|
100 |
if __name__ == "__main__":
|
|
|
101 |
messages = [
|
102 |
{"role": "system", "content": "you are a helpful assistant"},
|
103 |
{"role": "user", "content": "hi, what your name"}
|
|
|
14 |
在传递 device_map 时,low_cpu_mem_usage 会自动设置为 True
|
15 |
"""
|
16 |
|
17 |
+
self.tokenizer = None
|
18 |
+
# self.tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
|
19 |
+
self.model = None
|
20 |
+
# self.model = AutoModelForCausalLM.from_pretrained(
|
21 |
+
# model_name_or_path,
|
22 |
+
# torch_dtype="auto",
|
23 |
+
# device_map="auto"
|
24 |
+
# )
|
25 |
+
# self.model.eval()
|
26 |
self.generation_kwargs = dict(
|
27 |
do_sample=True,
|
28 |
temperature=0.7,
|
|
|
95 |
return self.tokenizer.decode(response[0][input_ids_length:], skip_special_tokens=True)
|
96 |
|
97 |
|
98 |
+
|
99 |
# bot = Qwen2Simulator("Qwen/Qwen2-0.5B-Instruct")
|
100 |
|
101 |
|
102 |
if __name__ == "__main__":
|
103 |
+
bot = Qwen2Simulator(r"E:\data_model\Qwen2-0.5B-Instruct")
|
104 |
messages = [
|
105 |
{"role": "system", "content": "you are a helpful assistant"},
|
106 |
{"role": "user", "content": "hi, what your name"}
|