File size: 5,014 Bytes
254bbe9
 
64e5a86
 
 
254bbe9
 
 
 
 
 
 
 
 
64e5a86
 
 
 
 
 
 
 
 
254bbe9
 
 
 
 
 
64e5a86
 
 
 
 
 
 
 
 
254bbe9
64e5a86
 
254bbe9
 
 
 
 
 
5e49fae
 
254bbe9
5e49fae
254bbe9
64e5a86
 
 
 
 
254bbe9
64e5a86
 
 
 
 
254bbe9
64e5a86
5e49fae
 
 
 
 
 
64e5a86
5e49fae
 
 
 
 
 
64e5a86
 
 
6ecfb5c
5e49fae
6ecfb5c
64e5a86
 
5e49fae
64e5a86
 
 
 
 
5e49fae
 
 
 
 
 
 
254bbe9
5e49fae
 
 
64e5a86
7718a35
64e5a86
 
5e49fae
254bbe9
64e5a86
 
5e49fae
64e5a86
 
 
5e49fae
 
 
 
 
 
 
 
 
 
 
 
64e5a86
 
 
 
 
 
 
 
5e49fae
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
64e5a86
5e49fae
64e5a86
 
 
 
 
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
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
from typing import Tuple

import gradio as gr


def deepmind_flops(
    n_layer: int,
    d_model: int,
    d_ff: int,
    d_attn: int,
    n_ctx: int,
    n_vocab: int,
    n_heads: int,
) -> int:
    embeddings = 2 * n_ctx * n_vocab * d_model
    attn_qkv = 2 * n_ctx * 3 * d_model * (d_attn * n_heads)
    attn_logits = 2 * n_ctx * n_ctx * (d_attn * n_heads)
    attn_softmax = 3 * n_heads * n_ctx * n_ctx
    attn_reduce = 2 * n_ctx * n_ctx * (d_attn * n_heads)
    attn_project = 2 * n_ctx * (d_attn * n_heads) * d_model
    ff = 2 * n_ctx * (d_model * d_ff + d_model * d_ff)
    logits = 2 * n_ctx * d_model * n_vocab

    params = (
        embeddings / n_ctx / 2,
        (n_layer * (attn_qkv + attn_project + ff)) / n_ctx / 2,
        logits / n_ctx / 2,
    )

    return (
        embeddings,
        attn_qkv * n_layer,
        attn_logits * n_layer,
        attn_softmax * n_layer,
        attn_reduce * n_layer,
        attn_project * n_layer,
        ff * n_layer,
        logits,
    ), params


def calculator(
    n_layer: int,
    d_model: int,
    n_heads: int,
    n_vocab: int,
    ff_ratio: int,
    n_ctx: int,
    n_tokens: int,
    incl_embed: bool,
    fwd_only: bool,
) -> Tuple[int, int, int]:
    d_attn = d_model // n_heads
    if d_model % n_heads != 0:
        raise gr.Error("d_model must be divisible by n_heads")
    d_ff = d_model * ff_ratio

    flops_terms, params = deepmind_flops(
        n_layer, d_model, d_ff, d_attn, n_ctx, n_vocab, n_heads
    )

    if incl_embed:
        flops_per_sequence = sum(flops_terms)
        params = sum(params)
    else:
        flops_per_sequence = sum(flops_terms[1:])
        params = sum(params[1:])

    flops_per_token = flops_per_sequence / n_ctx

    n_tokens_flops = flops_per_token * n_tokens

    if not fwd_only:
        flops_per_sequence *= 3
        flops_per_token *= 3
        n_tokens_flops *= 3

    return params, flops_per_sequence, flops_per_token, n_tokens_flops


with gr.Blocks() as iface:
    gr.Markdown(
        "Calculate how many FLOPs a Transformer language model uses with the method described in [DeepMind's Chinchilla scaling law paper](https://arxiv.org/abs/2203.15556) (see Appendix F)."
    )
    with gr.Row():
        with gr.Column():
            gr.Markdown("#### Architecture details")
            n_layer = gr.Number(label="Number of layers (n_layer)")
            d_model = gr.Number(label="Model dimensions (d_model)")
            n_heads = gr.Number(label="Number of attention heads per layer (n_heads)")
            n_vocab = gr.Number(label="Vocabulary size (n_vocab)")
            ff_ratio = gr.Number(value=4, label="Feedforward ratio")
            gr.Markdown("#### Data details")
            n_ctx = gr.Number(label="Sequence length (n_ctx)")
            n_tokens = gr.Number(
                value=0,
                label="Total number of training tokens (n_tokens) (optional)",
            )
            gr.Markdown("#### Settings")
            incl_embed = gr.Checkbox(value=True, label="Include embeddings")
            fwd_only = gr.Checkbox(
                value=False, label="Calculate FLOPs for only forward pass"
            )

            btn = gr.Button(value="Enter", variant="primary")

        with gr.Column():
            gr.Markdown("#### Output")
            params = gr.Number(label="Model parameters")
            flops_per_sequence = gr.Number(label="FLOPs per sequence")
            flops_per_token = gr.Number(label="FLOPs per token")
            n_tokens_flops = gr.Number(label="Total FLOPs for n_tokens")

    btn.click(
        calculator,
        inputs=[
            n_layer,
            d_model,
            n_heads,
            n_vocab,
            ff_ratio,
            n_ctx,
            n_tokens,
            incl_embed,
            fwd_only,
        ],
        outputs=[params, flops_per_sequence, flops_per_token, n_tokens_flops],
    )

    gr.Markdown("### GPT-3 model family examples")
    gr.Markdown(
        "In order are the 125M, 350M, 1.3B, 2.7B, 6.7B, 13B, 30B, 66B, and 175B parameter variants."
    )
    gr.Examples(
        [
            [12, 768, 12, 50257, 4, 4096, 0, True, False],
            [24, 1024, 16, 50257, 4, 4096, 0, True, False],
            [24, 2048, 32, 50257, 4, 4096, 0, True, False],
            [32, 2560, 32, 50257, 4, 4096, 0, True, False],
            [32, 4096, 32, 50257, 4, 4096, 0, True, False],
            [40, 5120, 40, 50257, 4, 4096, 0, True, False],
            [48, 7168, 56, 50257, 4, 4096, 0, True, False],
            [64, 9216, 72, 50257, 4, 4096, 0, True, False],
            [96, 12288, 96, 50257, 4, 4096, 0, True, False],
        ],
        [
            n_layer,
            d_model,
            n_heads,
            n_vocab,
            ff_ratio,
            n_ctx,
            n_tokens,
            incl_embed,
            fwd_only,
        ],
        [params, flops_per_sequence, flops_per_token, n_tokens_flops],
        calculator,
        cache_examples=False,
    )

iface.launch()