Spaces:
Runtime error
Runtime error
File size: 13,466 Bytes
62e9ca6 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 |
# --------------------------------------------------------
# Copyright (c) 2022 Microsoft
# Licensed under The MIT License [see LICENSE for details]
# Based on fairseq code bases
# https://github.com/facebookresearch/fairseq
# --------------------------------------------------------
"""
Modified from https://github.com/facebookresearch/fairseq/blob/main/fairseq/modules/transformer_layer.py
https://github.com/microsoft/SpeechT5/blob/main/Speech2C/speech2c/models/modules/transformer_decoder_layer.py
"""
from typing import Dict, List, Optional
import torch
from torch import Tensor
from fairseq.modules import LayerNorm
from speechlm.modules.multihead_attention import MultiheadAttention
from fairseq.modules.transformer_layer import TransformerEncoderLayerBase as FairseqTransformerEncoderLayerBase
from fairseq.modules.transformer_layer import TransformerDecoderLayerBase as FairseqTransformerDecoderLayerBase
class TransformerEncoderLayerBase(FairseqTransformerEncoderLayerBase):
"""Encoder layer block.
In the original paper each operation (multi-head attention or FFN) is
postprocessed with: `dropout -> add residual -> layernorm`. In the
tensor2tensor code they suggest that learning is more robust when
preprocessing each layer with layernorm and postprocessing with:
`dropout -> add residual`. We default to the approach in the paper, but the
tensor2tensor approach can be enabled by setting
*cfg.encoder.normalize_before* to ``True``.
Args:
args (argparse.Namespace): parsed command-line arguments
"""
def __init__(self, cfg, has_relative_attention_bias=False, scaling_for_att=1.0):
self.scaling_for_att = scaling_for_att
super().__init__(cfg)
if has_relative_attention_bias:
self.norm_k = LayerNorm(self.embed_dim // cfg.encoder.attention_heads)
def build_self_attention(self, embed_dim, cfg, scaling_for_att=1.0):
return MultiheadAttention(
embed_dim,
cfg.encoder.attention_heads,
dropout=cfg.attention_dropout,
self_attention=True,
q_noise=self.quant_noise,
qn_block_size=self.quant_noise_block_size,
scaling_for_att=self.scaling_for_att,
)
def forward(
self,
x,
encoder_padding_mask: Optional[Tensor],
attn_mask: Optional[Tensor] = None,
pos_bias=None,
):
"""
Args:
x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)`
encoder_padding_mask (ByteTensor): binary ByteTensor of shape
`(batch, seq_len)` where padding elements are indicated by ``1``.
attn_mask (ByteTensor): binary tensor of shape `(tgt_len, src_len)`,
where `tgt_len` is the length of output and `src_len` is the
length of input, though here both are equal to `seq_len`.
`attn_mask[tgt_i, src_j] = 1` means that when calculating the
embedding for `tgt_i`, we exclude (mask out) `src_j`. This is
useful for strided self-attention.
Returns:
encoded output of shape `(seq_len, batch, embed_dim)`
"""
# anything in original attn_mask = 1, becomes -1e8
# anything in original attn_mask = 0, becomes 0
# Note that we cannot use -inf here, because at some edge cases,
# the attention weight (before softmax) for some padded element in query
# will become -inf, which results in NaN in model parameters
if attn_mask is not None:
attn_mask = attn_mask.masked_fill(
attn_mask.to(torch.bool), -1e8 if x.dtype == torch.float32 else -1e4
)
residual = x
if self.normalize_before:
x = self.self_attn_layer_norm(x)
if pos_bias is not None:
pos_bias = self.norm_k(pos_bias)
x, _ = self.self_attn(
query=x,
key=x,
value=x,
key_padding_mask=encoder_padding_mask,
need_weights=False,
attn_mask=attn_mask,
position_bias=pos_bias,
)
x = self.dropout_module(x)
x = self.residual_connection(x, residual)
if not self.normalize_before:
x = self.self_attn_layer_norm(x)
residual = x
if self.normalize_before:
x = self.final_layer_norm(x)
x = self.activation_fn(self.fc1(x))
x = self.activation_dropout_module(x)
x = self.fc2(x)
x = self.dropout_module(x)
x = self.residual_connection(x, residual)
if not self.normalize_before:
x = self.final_layer_norm(x)
return x
class TransformerDecoderLayerBase(FairseqTransformerDecoderLayerBase):
"""Decoder layer block.
In the original paper each operation (multi-head attention, encoder
attention or FFN) is postprocessed with: `dropout -> add residual ->
layernorm`. In the tensor2tensor code they suggest that learning is more
robust when preprocessing each layer with layernorm and postprocessing with:
`dropout -> add residual`. We default to the approach in the paper, but the
tensor2tensor approach can be enabled by setting
*cfg.decoder.normalize_before* to ``True``.
Args:
args (argparse.Namespace): parsed command-line arguments
no_encoder_attn (bool, optional): whether to attend to encoder outputs
(default: False).
"""
def __init__(
self, cfg, no_encoder_attn=False, add_bias_kv=False, add_zero_attn=False, has_relative_attention_bias=False, scaling_for_att=1.0,
):
self.scaling_for_att = scaling_for_att
super().__init__(cfg,
no_encoder_attn,
add_bias_kv,
add_zero_attn,
)
if has_relative_attention_bias:
self.norm_k = LayerNorm(self.embed_dim // cfg.decoder.attention_heads)
def build_self_attention(
self, embed_dim, cfg, add_bias_kv=False, add_zero_attn=False
):
return MultiheadAttention(
embed_dim,
cfg.decoder.attention_heads,
dropout=cfg.attention_dropout,
add_bias_kv=add_bias_kv,
add_zero_attn=add_zero_attn,
self_attention=not cfg.cross_self_attention,
q_noise=self.quant_noise,
qn_block_size=self.quant_noise_block_size,
scaling_for_att=self.scaling_for_att,
)
def build_encoder_attention(self, embed_dim, cfg):
return MultiheadAttention(
embed_dim,
cfg.decoder.attention_heads,
kdim=cfg.encoder.embed_dim,
vdim=cfg.encoder.embed_dim,
dropout=cfg.attention_dropout,
encoder_decoder_attention=True,
q_noise=self.quant_noise,
qn_block_size=self.quant_noise_block_size,
scaling_for_att=self.scaling_for_att,
)
def forward(
self,
x,
encoder_out: Optional[torch.Tensor] = None,
encoder_padding_mask: Optional[torch.Tensor] = None,
incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None,
prev_self_attn_state: Optional[List[torch.Tensor]] = None,
prev_attn_state: Optional[List[torch.Tensor]] = None,
self_attn_mask: Optional[torch.Tensor] = None,
self_attn_padding_mask: Optional[torch.Tensor] = None,
need_attn: bool = False,
need_head_weights: bool = False,
pos_bias=None,
):
"""
Args:
x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)`
encoder_padding_mask (ByteTensor, optional): binary
ByteTensor of shape `(batch, src_len)` where padding
elements are indicated by ``1``.
need_attn (bool, optional): return attention weights
need_head_weights (bool, optional): return attention weights
for each head (default: return average over heads).
Returns:
encoded output of shape `(seq_len, batch, embed_dim)`
"""
if need_head_weights:
need_attn = True
residual = x
if self.normalize_before:
x = self.self_attn_layer_norm(x)
if pos_bias is not None:
pos_bias = self.norm_k(pos_bias)
if prev_self_attn_state is not None:
prev_key, prev_value = prev_self_attn_state[:2]
saved_state: Dict[str, Optional[Tensor]] = {
"prev_key": prev_key,
"prev_value": prev_value,
}
if len(prev_self_attn_state) >= 3:
saved_state["prev_key_padding_mask"] = prev_self_attn_state[2]
assert incremental_state is not None
self.self_attn._set_input_buffer(incremental_state, saved_state)
_self_attn_input_buffer = self.self_attn._get_input_buffer(incremental_state)
if self.cross_self_attention and not (
incremental_state is not None
and _self_attn_input_buffer is not None
and "prev_key" in _self_attn_input_buffer
):
if self_attn_mask is not None:
assert encoder_out is not None
self_attn_mask = torch.cat(
(x.new_zeros(x.size(0), encoder_out.size(0)), self_attn_mask), dim=1
)
if self_attn_padding_mask is not None:
if encoder_padding_mask is None:
assert encoder_out is not None
encoder_padding_mask = self_attn_padding_mask.new_zeros(
encoder_out.size(1), encoder_out.size(0)
)
self_attn_padding_mask = torch.cat(
(encoder_padding_mask, self_attn_padding_mask), dim=1
)
assert encoder_out is not None
y = torch.cat((encoder_out, x), dim=0)
else:
y = x
x, attn = self.self_attn(
query=x,
key=y,
value=y,
key_padding_mask=self_attn_padding_mask,
incremental_state=incremental_state,
need_weights=False,
attn_mask=self_attn_mask,
position_bias=pos_bias,
)
if self.c_attn is not None:
tgt_len, bsz = x.size(0), x.size(1)
x = x.view(tgt_len, bsz, self.nh, self.head_dim)
x = torch.einsum("tbhd,h->tbhd", x, self.c_attn)
x = x.reshape(tgt_len, bsz, self.embed_dim)
if self.attn_ln is not None:
x = self.attn_ln(x)
x = self.dropout_module(x)
x = self.residual_connection(x, residual)
if not self.normalize_before:
x = self.self_attn_layer_norm(x)
if self.encoder_attn is not None and encoder_out is not None:
residual = x
if self.normalize_before:
x = self.encoder_attn_layer_norm(x)
if prev_attn_state is not None:
prev_key, prev_value = prev_attn_state[:2]
saved_state: Dict[str, Optional[Tensor]] = {
"prev_key": prev_key,
"prev_value": prev_value,
}
if len(prev_attn_state) >= 3:
saved_state["prev_key_padding_mask"] = prev_attn_state[2]
assert incremental_state is not None
self.encoder_attn._set_input_buffer(incremental_state, saved_state)
x, attn = self.encoder_attn(
query=x,
key=encoder_out,
value=encoder_out,
key_padding_mask=encoder_padding_mask,
incremental_state=incremental_state,
static_kv=True,
need_weights=need_attn or (not self.training and self.need_attn),
need_head_weights=need_head_weights,
)
x = self.dropout_module(x)
x = self.residual_connection(x, residual)
if not self.normalize_before:
x = self.encoder_attn_layer_norm(x)
residual = x
if self.normalize_before:
x = self.final_layer_norm(x)
x = self.activation_fn(self.fc1(x))
x = self.activation_dropout_module(x)
if self.ffn_layernorm is not None:
x = self.ffn_layernorm(x)
x = self.fc2(x)
x = self.dropout_module(x)
if self.w_resid is not None:
residual = torch.mul(self.w_resid, residual)
x = self.residual_connection(x, residual)
if not self.normalize_before:
x = self.final_layer_norm(x)
if self.onnx_trace and incremental_state is not None:
saved_state = self.self_attn._get_input_buffer(incremental_state)
assert saved_state is not None
if self_attn_padding_mask is not None:
self_attn_state = [
saved_state["prev_key"],
saved_state["prev_value"],
saved_state["prev_key_padding_mask"],
]
else:
self_attn_state = [saved_state["prev_key"], saved_state["prev_value"]]
return x, attn, self_attn_state
return x, attn, None
def make_generation_fast_(self, need_attn: bool = False, **kwargs):
self.need_attn = need_attn
|