File size: 4,462 Bytes
b253e66
8443315
 
 
 
 
 
 
 
 
b253e66
8443315
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b253e66
 
a114a02
 
 
b253e66
a114a02
 
 
 
8443315
228bed3
 
 
 
535e574
 
 
fe36eff
f962dd0
 
 
 
c290138
 
8d171c2
 
b6ab215
dd5d2e0
bb204b7
 
 
 
 
 
 
 
b6ab215
 
bb204b7
b6ab215
bb204b7
e5222c4
 
 
535e574
 
 
 
 
 
 
bb204b7
 
 
 
 
 
 
535e574
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e5222c4
f962dd0
b6ab215
f962dd0
bb204b7
 
 
 
 
8443315
b6ab215
8443315
b253e66
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
131
from enum import Enum
from pathlib import Path

import streamlit as st
import streamlit.components.v1 as components
import torch
import torch.nn.functional as F
from transformers import AutoModelForCausalLM, AutoTokenizer, BatchEncoding

root_dir = Path(__file__).resolve().parent
highlighted_text_component = components.declare_component(
    "highlighted_text", path=root_dir / "highlighted_text" / "build"
)

def get_windows_batched(examples: BatchEncoding, window_len: int, stride: int = 1, pad_id: int = 0) -> BatchEncoding:
    return BatchEncoding({
        k: [
            t[i][j : j + window_len] + [
                pad_id if k == "input_ids" else 0
            ] * (j + window_len - len(t[i]))
            for i in range(len(examples["input_ids"]))
            for j in range(0, len(examples["input_ids"][i]) - 1, stride)
        ]
        for k, t in examples.items()
    })

BAD_CHAR = chr(0xfffd)

def ids_to_readable_tokens(tokenizer, ids, strip_whitespace=False):
    cur_ids = []
    result = []
    for idx in ids:
        cur_ids.append(idx)
        decoded = tokenizer.decode(cur_ids)
        if BAD_CHAR not in decoded:
            if strip_whitespace:
                decoded = decoded.strip()
            result.append(decoded)
            del cur_ids[:]
        else:
            result.append("")
    return result

st.header("Context length probing")

model_name = st.selectbox("Model", ["distilgpt2", "gpt2", "EleutherAI/gpt-neo-125m"])
metric_name = st.selectbox("Metric", ["KL divergence", "Cross entropy"], index=1)
window_len = st.select_slider(r"Window size ($c_\text{max}$)", options=[8, 16, 32, 64, 128, 256, 512, 1024], value=512)

text = st.text_area(
    "Input text",
    "The complex houses married and single soldiers and their families.",
)

if metric_name == "KL divergence":
    st.error("KL divergence is not supported yet. Stay tuned!", icon="😭")
    st.stop()

with st.spinner("Loading model…"):
    tokenizer = st.cache_resource(AutoTokenizer.from_pretrained, show_spinner=False)(model_name)
    model = st.cache_resource(AutoModelForCausalLM.from_pretrained, show_spinner=False)(model_name)

inputs = tokenizer([text])
[input_ids] = inputs["input_ids"]
window_len = min(window_len, len(input_ids))

if len(input_ids) < 2:
    st.error("Please enter at least 2 tokens.", icon="🚨")
    st.stop()

@st.cache_data(show_spinner=False)
@torch.inference_mode()
def get_logits(_model, _inputs, cache_key):
    del cache_key
    return _model(**_inputs).logits.to(torch.float16)

@st.cache_data(show_spinner=False)
@torch.inference_mode()
def run_context_length_probing(_model, _tokenizer, _inputs, window_len, cache_key):
    del cache_key

    inputs_sliding = get_windows_batched(
        _inputs,
        window_len=window_len,
        pad_id=_tokenizer.eos_token_id
    ).convert_to_tensors("pt")

    logits = []
    with st.spinner("Running model…"):
        batch_size = 8
        num_items = len(inputs_sliding["input_ids"])
        pbar = st.progress(0)
        for i in range(0, num_items, batch_size):
            pbar.progress(i / num_items, f"{i}/{num_items}")
            batch = {k: v[i:i + batch_size] for k, v in inputs_sliding.items()}
            logits.append(
                get_logits(
                    _model,
                    batch,
                    cache_key=(model_name, batch["input_ids"].cpu().numpy().tobytes())
                )
            )
        logits = torch.cat(logits, dim=0)
        pbar.empty()

    with st.spinner("Computing scores…"):
        logits = logits.permute(1, 0, 2)
        logits = F.pad(logits, (0, 0, 0, window_len, 0, 0), value=torch.nan)
        logits = logits.view(-1, logits.shape[-1])[:-window_len]
        logits = logits.view(window_len, len(input_ids) + window_len - 2, logits.shape[-1])

        scores = logits.to(torch.float32).log_softmax(dim=-1)
        scores = scores[:, torch.arange(len(input_ids[1:])), input_ids[1:]]
        scores = scores.diff(dim=0).transpose(0, 1)
        scores = scores.nan_to_num()
        scores /= scores.abs().max(dim=1, keepdim=True).values + 1e-9
        scores = scores.to(torch.float16)

    return scores

scores = run_context_length_probing(
    _model=model,
    _tokenizer=tokenizer,
    _inputs=inputs,
    window_len=window_len,
    cache_key=(model_name, text),
)
tokens = ids_to_readable_tokens(tokenizer, input_ids)

highlighted_text_component(tokens=tokens, scores=scores.tolist())