File size: 10,492 Bytes
746496b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c80a9f3
 
746496b
 
 
 
c80a9f3
746496b
 
 
 
 
 
 
 
 
 
 
 
 
c80a9f3
746496b
 
 
 
 
 
c80a9f3
 
746496b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b38fe51
746496b
 
 
 
c80a9f3
 
746496b
c80a9f3
746496b
 
c80a9f3
 
746496b
 
c80a9f3
746496b
 
 
 
 
c80a9f3
746496b
 
 
 
 
 
 
 
 
c80a9f3
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
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
# From https://stackoverflow.com/a/23689767
# From https://github.com/pytorch/pytorch/issues/97899
# From https://github.com/facebookresearch/llama/blob/main/llama/model.py

import torch
from torch import nn

from xformers.ops import SwiGLU, memory_efficient_attention

from .rmsnorm import RMSNorm
from .rotary import precompute_freqs_cis, apply_rotary_emb

from transformers import PreTrainedModel, PretrainedConfig
from transformers.modeling_outputs import MaskedLMOutput

class DotDict(dict):
    """Dictionary that supports the dot notation to access attributes (similarly to HuggingFace)."""

    __getattr__ = dict.get
    __setattr__ = dict.__setitem__
    __delattr__ = dict.__delitem__

class AMPLIFYConfig(PretrainedConfig):
    model_type = "AMPLIFY"
    # All config parameters must have a default value.
    def __init__(
        self,
        hidden_size: int = 960,
        num_hidden_layers: int = 32,
        num_attention_heads: int = 15,
        intermediate_size: int = 3840,
        dropout_prob: float = 0,
        embedding_init_range: float = 0.02,
        decoder_init_range: float = 0.02,
        rms_norm: bool = True,
        norm_eps: float = 1e-05,
        hidden_act: str = "SwiGLU",
        layer_norm_after_embedding: bool = False,
        layer_norm_before_last_layer: bool = True,
        vocab_size: int = 27,
        ffn_bias: bool = False,
        att_bias: bool = False,
        pad_token_id: int = 0,
        max_length: int = 2048,
        **kwargs,
    ):
        super().__init__(**kwargs)
        
        self.hidden_size = hidden_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.intermediate_size = intermediate_size
        self.dropout_prob = dropout_prob
        self.embedding_init_range = embedding_init_range
        self.decoder_init_range = decoder_init_range
        self.rms_norm = rms_norm
        self.norm_eps = norm_eps
        self.hidden_act = hidden_act
        self.layer_norm_after_embedding = layer_norm_after_embedding
        self.layer_norm_before_last_layer = layer_norm_before_last_layer
        self.vocab_size = vocab_size
        self.ffn_bias = ffn_bias
        self.att_bias = att_bias
        self.pad_token_id = pad_token_id
        self.max_length = max_length
        

class EncoderBlock(nn.Module):
    """Transformer encoder block."""

    def __init__(self, config: AMPLIFYConfig):
        """Initialize a EncoderBlock.

        Args:
            hidden_size (int): _description_
            num_attention_heads (int): _description_
            intermediate_size (int, optional): _description_. Defaults to 2048.
            dropout_prob (float, optional): _description_. Defaults to 0.1.
            activation (str, optional): _description_. Defaults to "relu".
            rms_norm (bool, optional): _description_. Defaults to True.
            norm_eps (float, optional): _description_. Defaults to 1e-5.
            pad_token_id (int, optional): _description_. Defaults to 0.
            max_length (int, optional): _description_. Defaults to 2048.
            ffn_bias (bool, optional): _description_. Defaults to False.
            att_bias (bool, optional): _description_. Defaults to False.
        """
        super().__init__()

        self.config = config
        self.d_head = config.hidden_size // config.num_attention_heads

        # Attention
        self.q = nn.Linear(in_features=config.hidden_size, out_features=config.hidden_size, bias=config.att_bias)
        self.k = nn.Linear(in_features=config.hidden_size, out_features=config.hidden_size, bias=config.att_bias)
        self.v = nn.Linear(in_features=config.hidden_size, out_features=config.hidden_size, bias=config.att_bias)
        self.wo = nn.Linear(in_features=config.hidden_size, out_features=config.hidden_size, bias=config.att_bias)
        self.resid_dropout = nn.Dropout(config.dropout_prob)

        # Feedforward network
        match config.hidden_act.lower():
            case "swiglu":
                # To keep the number of parameters and the amount of computation constant, we reduce the number of
                # hidden units by a factor of 2/3 (https://arxiv.org/pdf/2002.05202.pdf) and make it a multiple of 8 to
                # avoid RuntimeError due to misaligned operand
                multiple_of = 8
                intermediate_size = int(2 * config.intermediate_size / 3)
                intermediate_size = multiple_of * ((intermediate_size + multiple_of - 1) // multiple_of)
                self.ffn = SwiGLU(config.hidden_size, intermediate_size, config.hidden_size, bias=config.ffn_bias)
            case "relu":
                self.ffn = nn.Sequential(
                    nn.Linear(config.hidden_size, config.intermediate_size, bias=config.ffn_bias),
                    nn.ReLU(),
                    nn.Linear(config.intermediate_size, config.hidden_size, bias=config.ffn_bias),
                )
            case "gelu":
                self.ffn = nn.Sequential(
                    nn.Linear(config.hidden_size, config.intermediate_size, bias=config.ffn_bias),
                    nn.GELU(),
                    nn.Linear(config.intermediate_size, config.hidden_size, bias=config.ffn_bias),
                )

        self.attention_norm = RMSNorm(config.hidden_size, config.norm_eps) if config.rms_norm else nn.LayerNorm(config.hidden_size, config.norm_eps)
        self.ffn_norm = RMSNorm(config.hidden_size, config.norm_eps) if config.rms_norm else nn.LayerNorm(config.hidden_size, config.norm_eps)

        self.ffn_dropout = nn.Dropout(config.dropout_prob)

    def forward(self, x: torch.Tensor, attention_mask: torch.Tensor, freqs_cis: torch.Tensor, output_attentions: bool):
        attn, contact = self._att_block(self.attention_norm(x), attention_mask, freqs_cis, output_attentions)
        x = x + attn
        x = x + self._ff_block(self.ffn_norm(x))
        return x, contact

    def _att_block(self, x: torch.Tensor, attention_mask: torch.Tensor, freqs_cis: torch.Tensor, output_attentions: bool):
        batch_size, seq_len, _ = x.shape
        xq, xk, xv = self.q(x), self.k(x), self.v(x)

        # Reshape for rotary embeddings
        xq = xq.view(batch_size, seq_len, self.config.num_attention_heads, self.d_head)
        xk = xk.view(batch_size, seq_len, self.config.num_attention_heads, self.d_head)
        xv = xv.view(batch_size, seq_len, self.config.num_attention_heads, self.d_head)
        xq, xk = apply_rotary_emb(xq, xk, freqs_cis)

        attn = memory_efficient_attention(
            query=xq,
            key=xk,
            value=xv,
            attn_bias=attention_mask,
            p=self.config.dropout_prob if self.training else 0,
        )

        _attn = None
        if output_attentions:
            _attn = xq.permute(0, 2, 1, 3) @ xk.permute(0, 2, 3, 1) / (xq.size(-1) ** 0.5)
            if attention_mask is not None:
                _attn = _attn + attention_mask
            _attn = _attn.softmax(-1)

        return self.resid_dropout(self.wo(attn.view(batch_size, seq_len, self.config.num_attention_heads * self.d_head))), _attn

    def _ff_block(self, x: torch.Tensor):
        return self.ffn_dropout(self.ffn(x))


class AMPLIFYPreTrainedModel(PreTrainedModel):
    config_class = AMPLIFYConfig

    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            module.weight.data.uniform_(-self.config.decoder_init_range, self.config.decoder_init_range)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.uniform_(-self.config.embedding_init_range, self.config.embedding_init_range)


class AMPLIFY(AMPLIFYPreTrainedModel):
    """The main model class.

       Args:
          config (amplify.model.amplify.AMPLIFYConfig): model configuration, usually defined from the Hydra configuration.
    """
    def __init__(self, config: AMPLIFYConfig, **kwargs):
        super().__init__(config)

        self.config = config

        self.encoder = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)

        if config.layer_norm_after_embedding:
            self.layer_norm_1 = RMSNorm(config.hidden_size, config.norm_eps) if config.rms_norm else nn.LayerNorm(config.hidden_size, config.norm_eps)

        self.transformer_encoder = nn.ModuleList()
        for _ in range(config.num_hidden_layers):
            self.transformer_encoder.append(EncoderBlock(config))

        if config.layer_norm_before_last_layer:
            self.layer_norm_2 = RMSNorm(config.hidden_size, config.norm_eps) if config.rms_norm else nn.LayerNorm(config.hidden_size, config.norm_eps)

        self.decoder = nn.Linear(config.hidden_size, config.vocab_size)

        self.freqs_cis = precompute_freqs_cis(config.hidden_size // config.num_attention_heads, config.max_length)
        
        # Initialize weights and apply final processing
        self.post_init()

    def forward(self, input_ids, attention_mask=None, output_hidden_states=False, output_attentions=False, **kwargs):
        # Initialize
        hidden_states, attentions = [], []

        # Expand and repeat: (Batch, Length) -> (Batch, Heads, Length, Length)
        if attention_mask is not None and not torch.all(attention_mask == 0):
            attention_mask = attention_mask.unsqueeze(1).unsqueeze(1).repeat(1, self.config.num_attention_heads, attention_mask.size(-1), 1)
        else:
            attention_mask = None

        # RoPE
        self.freqs_cis = self.freqs_cis.to(input_ids.device, non_blocking=True)
        freqs_cis = self.freqs_cis[: input_ids.shape[1]]

        # Embedding
        x = self.encoder(input_ids)
        if self.config.layer_norm_after_embedding:
            x = self.layer_norm_1(x)

        # Transformer encoder
        for layer in self.transformer_encoder:
            x, attn = layer(x, attention_mask, freqs_cis, output_attentions)
            if output_hidden_states:
                hidden_states.append(x)
            if output_attentions:
                attentions.append(attn)

        # Classification head with layer norm
        logits = self.decoder(self.layer_norm_2(x) if self.config.layer_norm_before_last_layer else x)

        # Return logits or the output of the last hidden layer
        return MaskedLMOutput(logits=logits, hidden_states=hidden_states, attentions=attentions)