File size: 4,568 Bytes
56ab742
5b9f390
56ab742
 
 
 
 
5b9f390
 
 
 
 
 
 
 
 
 
 
 
 
 
56ab742
5b9f390
4115285
 
 
5b9f390
 
 
 
1dfc873
5b9f390
1dfc873
 
5b9f390
 
 
 
 
 
 
56ab742
5b9f390
2698ee0
5b9f390
 
2698ee0
5b9f390
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2698ee0
5b9f390
77bd93c
5b9f390
 
 
 
1dfc873
5b9f390
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
# login as a privileged user.
import os
HF_TOKEN = os.environ.get("HF_TOKEN")

from huggingface_hub import login
login(token=HF_TOKEN)

from threading import Thread
from typing import Iterator

import gradio as gr
import spaces
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer

from pyreft import ReftModel

MAX_MAX_NEW_TOKENS = 2048
DEFAULT_MAX_NEW_TOKENS = 1024
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))


DESCRIPTION = """\
# Ethos-Chat with ReFT

### Ethos-Chat is a [GOODY-2](https://www.goody2.ai/chat) imitator built with ReFT. It is trained with 5 training examples under 30 seconds. You can train your own ReFT agent and share it on HuggingFace by following this [tutorial](https://github.com/stanfordnlp/pyreft/tree/main/examples/gradio/train_and_share.ipynb)!
"""

LICENSE = """
<p/>

---
As a derivate work of [Llama-2-7b-chat](https://huggingface.co/meta-llama/Llama-2-7b-chat) by Meta,
this demo is governed by the original [license](https://huggingface.co/spaces/huggingface-projects/llama-2-7b-chat/blob/main/LICENSE.txt) and [acceptable use policy](https://huggingface.co/spaces/huggingface-projects/llama-2-7b-chat/blob/main/USE_POLICY.md).
"""

if not torch.cuda.is_available():
    DESCRIPTION += "\n<p>Running on CPU 🥶 This demo does not work on CPU.</p>"


if torch.cuda.is_available():
    model_id = "meta-llama/Llama-2-7b-chat-hf" # not gated version.
    model = AutoModelForCausalLM.from_pretrained(
        model_id, device_map="cuda", torch_dtype=torch.bfloat16
    )
    reft_model = ReftModel.load("pyvene/reft_goody2", model, from_huggingface_hub=True)
    reft_model.set_device("cuda")
    tokenizer = AutoTokenizer.from_pretrained(model_id)
    tokenizer.use_default_system_prompt = True

prompt_no_input_template = """<s>[INST] <<SYS>>
You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe.  Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.

If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.
<</SYS>>

%s [/INST]
"""

@spaces.GPU
def generate(
    message: str,
    chat_history: list[tuple[str, str]],
    max_new_tokens: int = 1024,
) -> Iterator[str]:

    # tokenize and prepare the input
    prompt = prompt_no_input_template % message
    prompt = tokenizer(prompt, return_tensors="pt").to(model.device)
    input_ids = prompt["input_ids"]
    attention_mask = prompt["attention_mask"]
    
    if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
        input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
        attention_mask = attention_mask[:, -MAX_INPUT_TOKEN_LENGTH:]
        gr.Warning(f"Trimmed input from conversation as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.")

    base_unit_location = input_ids.shape[-1] - 1  # last position
    
    streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
    generate_kwargs = {
        "base": {"input_ids": input_ids, "attention_mask": attention_mask},
        "unit_locations": {"sources->base": (None, [[[base_unit_location]]])},
        "max_new_tokens": max_new_tokens,
        "intervene_on_prompt": True,
        "streamer": streamer,
        "eos_token_id": tokenizer.eos_token_id,
        "early_stopping": True,
        "do_sample": False
    }

    t = Thread(target=reft_model.generate, kwargs=generate_kwargs)
    t.start()

    outputs = []
    for text in streamer:
        outputs.append(text)
        yield "".join(outputs)


chat_interface = gr.ChatInterface(
    fn=generate,
    additional_inputs=[
        gr.Slider(
            label="Max new tokens",
            minimum=1,
            maximum=MAX_MAX_NEW_TOKENS,
            step=1,
            value=DEFAULT_MAX_NEW_TOKENS,
        )
    ],
    stop_btn=None,
    examples=[
        ["What's 2+2?"],
        ["Why is the sky blue?"],
        ["What's Apple's stock price?"],
        ["Plan a family road trip to Austin"],
    ],
)

with gr.Blocks(css="style.css") as demo:
    gr.Markdown(DESCRIPTION)
    gr.DuplicateButton(value="Duplicate Space for private use", elem_id="duplicate-button")
    chat_interface.render()
    gr.Markdown(LICENSE)

if __name__ == "__main__":
    demo.queue(max_size=20).launch()