tergel commited on
Commit
0c8db5a
·
1 Parent(s): 87af67e

Update app.py

Browse files
Files changed (4) hide show
  1. README.md +4 -6
  2. app.py +127 -42
  3. requirements.txt +3 -1
  4. style.css +12 -0
README.md CHANGED
@@ -1,14 +1,12 @@
1
  ---
2
  title: Concise Reasoning Demo
3
- emoji: 💬
4
- colorFrom: yellow
5
  colorTo: purple
6
  sdk: gradio
7
- sdk_version: 5.0.1
8
  app_file: app.py
9
  pinned: false
10
  license: mit
11
  short_description: demo of LLMs fine-tuned for concise reasoning
12
- ---
13
-
14
- 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: Concise Reasoning Demo
3
+ emoji: 🏍️
4
+ colorFrom: red
5
  colorTo: purple
6
  sdk: gradio
7
+ sdk_version: 5.12.0
8
  app_file: app.py
9
  pinned: false
10
  license: mit
11
  short_description: demo of LLMs fine-tuned for concise reasoning
12
+ ---
 
 
app.py CHANGED
@@ -1,64 +1,149 @@
1
- import gradio as gr
2
- from huggingface_hub import InferenceClient
 
3
 
4
- """
5
- For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
6
- """
7
- client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
8
 
 
 
 
 
9
 
10
- def respond(
11
- message,
12
- history: list[tuple[str, str]],
13
- system_message,
14
- max_tokens,
15
- temperature,
16
- top_p,
17
- ):
18
- messages = [{"role": "system", "content": system_message}]
19
 
20
- for val in history:
21
- if val[0]:
22
- messages.append({"role": "user", "content": val[0]})
23
- if val[1]:
24
- messages.append({"role": "assistant", "content": val[1]})
25
 
26
- messages.append({"role": "user", "content": message})
 
27
 
28
- response = ""
 
29
 
30
- for message in client.chat_completion(
31
- messages,
32
- max_tokens=max_tokens,
33
- stream=True,
34
- temperature=temperature,
35
- top_p=top_p,
36
- ):
37
- token = message.choices[0].delta.content
 
 
 
38
 
39
- response += token
40
- yield response
41
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
42
 
43
- """
44
- For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
45
- """
46
- demo = gr.ChatInterface(
47
- respond,
48
  additional_inputs=[
49
- gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
50
- gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
51
- gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
52
  gr.Slider(
 
 
 
 
 
 
 
 
53
  minimum=0.1,
 
 
 
 
 
 
 
54
  maximum=1.0,
 
55
  value=0.95,
 
 
 
 
 
 
 
 
 
 
 
 
56
  step=0.05,
57
- label="Top-p (nucleus sampling)",
58
  ),
59
  ],
 
 
 
 
 
 
 
 
 
 
 
 
 
 
60
  )
61
 
 
 
 
 
62
 
63
  if __name__ == "__main__":
64
- demo.launch()
 
1
+ import os
2
+ from collections.abc import Iterator
3
+ from threading import Thread
4
 
5
+ import gradio as gr
6
+ import spaces
7
+ import torch
8
+ from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
9
 
10
+ # Token limits
11
+ MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))
12
+ MAX_MAX_NEW_TOKENS = 2048
13
+ DEFAULT_MAX_NEW_TOKENS = 512
14
 
15
+ # Description
16
+ DESCRIPTION = """\
17
+ # Demo for "Self-Training Elicits Concise Reasoning in Large Language Models"
 
 
 
 
 
 
18
 
19
+ This Space showcases the model [tergel/llama-3.2-3b-instruct-gsm8k-fs-gpt4o-bon](https://huggingface.co/tergel/llama-3.2-3b-instruct-gsm8k-fs-gpt4o-bon)
 
 
 
 
20
 
21
+ We provide a simple chat interface allowing you to observe the concise CoT solutions that our model can produce. Feel free to play with it.
22
+ """
23
 
24
+ # Decide on device
25
+ device = "cuda" if torch.cuda.is_available() else "cpu"
26
 
27
+ if not torch.cuda.is_available():
28
+ DESCRIPTION += "\n\n<p>**Warning**: Running on CPU 🥶 – this may be extremely slow. We will upgrade to GPUs soon.</p>"
29
+
30
+ # Load model and tokenizer
31
+ model_id = "tergel/llama-3.2-3b-instruct-gsm8k-fs-gpt4o-bon"
32
+ model = AutoModelForCausalLM.from_pretrained(
33
+ model_id,
34
+ device_map=None if device == "cpu" else "auto",
35
+ torch_dtype=torch.bfloat16 if device == "cuda" else torch.float32,
36
+ )
37
+ model.to(device)
38
 
39
+ tokenizer = AutoTokenizer.from_pretrained(model_id)
40
+ tokenizer.use_default_system_prompt = False
41
 
42
+ @spaces.GPU
43
+ def generate(
44
+ message: str,
45
+ chat_history: list[dict],
46
+ system_prompt: str = "",
47
+ max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,
48
+ temperature: float = 0.7,
49
+ top_p: float = 0.95,
50
+ top_k: int = 40,
51
+ repetition_penalty: float = 1.2,
52
+ ) -> Iterator[str]:
53
+ # Build conversation
54
+ conversation = []
55
+ if system_prompt:
56
+ conversation.append({"role": "system", "content": system_prompt})
57
+ conversation += chat_history
58
+ conversation.append({"role": "user", "content": message})
59
+
60
+ input_ids = tokenizer.apply_chat_template(conversation, return_tensors="pt")
61
+ if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
62
+ input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
63
+ gr.Warning(f"Trimmed input from conversation as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.")
64
+ input_ids = input_ids.to(model.device)
65
+
66
+ streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
67
+ generate_kwargs = dict(
68
+ {"input_ids": input_ids},
69
+ streamer=streamer,
70
+ max_new_tokens=max_new_tokens,
71
+ do_sample=True,
72
+ top_p=top_p,
73
+ top_k=top_k,
74
+ temperature=temperature,
75
+ num_beams=1,
76
+ repetition_penalty=repetition_penalty,
77
+ )
78
+ t = Thread(target=model.generate, kwargs=generate_kwargs)
79
+ t.start()
80
+
81
+ outputs = []
82
+ for text in streamer:
83
+ outputs.append(text)
84
+ yield "".join(outputs)
85
+
86
 
87
+ chat_interface = gr.ChatInterface(
88
+ fn=generate,
 
 
 
89
  additional_inputs=[
90
+ gr.Textbox(label="System prompt", lines=6),
 
 
91
  gr.Slider(
92
+ label="Max new tokens",
93
+ minimum=1,
94
+ maximum=MAX_MAX_NEW_TOKENS,
95
+ step=1,
96
+ value=DEFAULT_MAX_NEW_TOKENS,
97
+ ),
98
+ gr.Slider(
99
+ label="Temperature",
100
  minimum=0.1,
101
+ maximum=4.0,
102
+ step=0.1,
103
+ value=0.7,
104
+ ),
105
+ gr.Slider(
106
+ label="Top-p (nucleus sampling)",
107
+ minimum=0.05,
108
  maximum=1.0,
109
+ step=0.05,
110
  value=0.95,
111
+ ),
112
+ gr.Slider(
113
+ label="Top-k",
114
+ minimum=1,
115
+ maximum=1000,
116
+ step=1,
117
+ value=40,
118
+ ),
119
+ gr.Slider(
120
+ label="Repetition penalty",
121
+ minimum=1.0,
122
+ maximum=2.0,
123
  step=0.05,
124
+ value=1.2,
125
  ),
126
  ],
127
+ stop_btn=None,
128
+ examples=[
129
+ [
130
+ "A robe takes 2 bolts of blue fiber and half that much white fiber. How many bolts in total does it take?"
131
+ ],
132
+ [
133
+ "Claire makes a 3 egg omelet every morning for breakfast. How many dozens of eggs will she eat in 4 weeks?"
134
+ ],
135
+ [
136
+ "James decides to run 3 sprints 3 times a week. He runs 60 meters each sprint. How many total meters does he run a week?"
137
+ ],
138
+ ],
139
+ cache_examples=False,
140
+ type="messages",
141
  )
142
 
143
+ with gr.Blocks(css_paths="style.css", fill_height=True) as demo:
144
+ gr.Markdown(DESCRIPTION)
145
+ chat_interface.render()
146
+
147
 
148
  if __name__ == "__main__":
149
+ demo.queue(max_size=20).launch()
requirements.txt CHANGED
@@ -1 +1,3 @@
1
- huggingface_hub==0.25.2
 
 
 
1
+ huggingface_hub==0.25.2
2
+ torch
3
+ transformers
style.css ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ h1 {
2
+ text-align: center;
3
+ display: block;
4
+ }
5
+
6
+ #duplicate-button {
7
+ margin: auto;
8
+ color: white;
9
+ background: #1565c0;
10
+ border-radius: 100vh;
11
+ }
12
+