File size: 9,139 Bytes
04ffec9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
"""
Much of this code is adapted from Andrej Karpathy's NanoGPT
(https://github.com/karpathy/nanoGPT)
"""
import math
from dataclasses import dataclass

import torch
import torch.nn as nn
from torch.nn import functional as F

class LayerNorm(nn.Module):
    """ LayerNorm but with an optional bias. PyTorch doesn't support simply bias=False """

    def __init__(self, ndim, bias):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(ndim))
        self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None

    def forward(self, input):
        return F.layer_norm(input, self.weight.shape, self.weight, self.bias, 1e-5)

class CausalSelfAttention(nn.Module):

    def __init__(self, config):
        super().__init__()
        assert config.n_embd % config.n_head == 0
        # key, query, value projections for all heads, but in a batch
        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
        # output projection
        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
        # regularization
        self.attn_dropout = nn.Dropout(config.dropout)
        self.resid_dropout = nn.Dropout(config.dropout)
        self.n_head = config.n_head
        self.n_embd = config.n_embd
        self.dropout = config.dropout
        # flash attention make GPU go brrrrr but support is only in PyTorch nightly and still a bit scary
        self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention')
        if not self.flash:
            # print("WARNING: using slow attention. Flash Attention atm needs PyTorch nightly and dropout=0.0")
            # causal mask to ensure that attention is only applied to the left in the input sequence
            self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size))
                                        .view(1, 1, config.block_size, config.block_size))

    def forward(self, x, past_kv=None, use_cache=False):
        B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)

        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
        q, k ,v  = self.c_attn(x).split(self.n_embd, dim=2)
        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)

        if past_kv is not None:
            past_key = past_kv[0]
            past_value = past_kv[1]
            k = torch.cat((past_key, k), dim=-2)
            v = torch.cat((past_value, v), dim=-2)

        FULL_T = k.shape[-2]

        if use_cache is True:
            present = (k, v)
        else:
            present = None

        # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
        if self.flash:
            # efficient attention using Flash Attention CUDA kernels
            if past_kv is not None:
                # When `past_kv` is provided, we're doing incremental decoding and `q.shape[2] == 1`: q only contains
                # the query for the last token. scaled_dot_product_attention interprets this as the first token in the
                # sequence, so if is_causal=True it will mask out all attention from it. This is not what we want, so 
                # to work around this we set is_causal=False.
                is_causal = False
            else:
                is_causal = True

            y = torch.nn.functional.scaled_dot_product_attention(q, k, v, dropout_p=self.dropout, is_causal=is_causal)
        else:
            # manual implementation of attention
            att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
            att = att.masked_fill(self.bias[:,:,FULL_T-T:FULL_T,:FULL_T] == 0, float('-inf'))
            att = F.softmax(att, dim=-1)
            att = self.attn_dropout(att)
            y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
        y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side

        # output projection
        y = self.resid_dropout(self.c_proj(y))
        return (y, present)

class MLP(nn.Module):

    def __init__(self, config):
        super().__init__()
        self.c_fc    = nn.Linear(config.n_embd, 4 * config.n_embd, bias=config.bias)
        self.c_proj  = nn.Linear(4 * config.n_embd, config.n_embd, bias=config.bias)
        self.dropout = nn.Dropout(config.dropout)
        self.gelu = nn.GELU()

    def forward(self, x):
        x = self.c_fc(x)
        x = self.gelu(x)
        x = self.c_proj(x)
        x = self.dropout(x)
        return x

class Block(nn.Module):

    def __init__(self, config, layer_idx):
        super().__init__()
        self.ln_1 = LayerNorm(config.n_embd, bias=config.bias)
        self.attn = CausalSelfAttention(config)
        self.ln_2 = LayerNorm(config.n_embd, bias=config.bias)
        self.mlp = MLP(config)
        self.layer_idx = layer_idx

    def forward(self, x, past_kv=None, use_cache=False):
        attn_output, prev_kvs = self.attn(self.ln_1(x), past_kv=past_kv, use_cache=use_cache)
        x = x + attn_output
        x = x + self.mlp(self.ln_2(x))
        return (x, prev_kvs)

@dataclass
class GPTConfig:
    block_size: int = 1024
    input_vocab_size: int = 10_048
    output_vocab_size: int = 10_048
    n_layer: int = 12
    n_head: int = 12
    n_embd: int = 768
    dropout: float = 0.0
    bias: bool = True # True: bias in Linears and LayerNorms, like GPT-2. False: a bit better and faster

class GPT(nn.Module):

    def __init__(self, config):
        super().__init__()
        assert config.input_vocab_size is not None
        assert config.output_vocab_size is not None
        assert config.block_size is not None
        self.config = config

        self.transformer = nn.ModuleDict(dict(
            wte = nn.Embedding(config.input_vocab_size, config.n_embd),
            wpe = nn.Embedding(config.block_size, config.n_embd),
            drop = nn.Dropout(config.dropout),
            h = nn.ModuleList([Block(config, idx) for idx in range(config.n_layer)]),
            ln_f = LayerNorm(config.n_embd, bias=config.bias),
        ))
        self.lm_head = nn.Linear(config.n_embd, config.output_vocab_size, bias=False)

    def get_num_params(self, non_embedding=True):
        """
        Return the number of parameters in the model.
        For non-embedding count (default), the position embeddings get subtracted.
        The token embeddings would too, except due to the parameter sharing these
        params are actually used as weights in the final layer, so we include them.
        """
        n_params = sum(p.numel() for p in self.parameters())
        if non_embedding:
            n_params -= self.transformer.wte.weight.numel()
            n_params -= self.transformer.wpe.weight.numel()
        return n_params

    def forward(self, idx, merge_context=False, past_kv=None, position_ids=None, use_cache=False):
        device = idx.device
        b, t = idx.size()
        if past_kv is not None:
            assert t == 1
            tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)
        else:
            if merge_context:
                assert(idx.shape[1] >= 256+256+1)
                t = idx.shape[1] - 256
            else:
                assert t <= self.config.block_size, f"Cannot forward sequence of length {t}, block size is only {self.config.block_size}"

            # forward the GPT model itself
            if merge_context:
                tok_emb = torch.cat([
                    self.transformer.wte(idx[:,:256]) + self.transformer.wte(idx[:,256:256+256]),
                    self.transformer.wte(idx[:,256+256:])
                ], dim=1)
            else:
                tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)

        if past_kv is None:
            past_length = 0
            past_kv = tuple([None] * len(self.transformer.h))
        else:
            past_length = past_kv[0][0].size(-2)

        if position_ids is None:
            position_ids = torch.arange(past_length, t + past_length, dtype=torch.long, device=device)
            position_ids = position_ids.unsqueeze(0) # shape (1, t)
            assert position_ids.shape == (1, t)

        pos_emb = self.transformer.wpe(position_ids) # position embeddings of shape (1, t, n_embd)

        x = self.transformer.drop(tok_emb + pos_emb)

        new_kv = () if use_cache else None

        for i, (block, past_layer_kv) in enumerate(zip(self.transformer.h, past_kv)):
            x, kv = block(x, past_kv=past_layer_kv, use_cache=use_cache)

            if use_cache:
                new_kv = new_kv + (kv,)

        x = self.transformer.ln_f(x)

        # inference-time mini-optimization: only forward the lm_head on the very last position
        logits = self.lm_head(x[:, [-1], :]) # note: using list [-1] to preserve the time dim

        return (logits, new_kv)