File size: 14,140 Bytes
2650f1f b8defd9 2650f1f 99d97ed 2650f1f e8bdeb2 2650f1f 99d97ed 2650f1f 99d97ed 2650f1f |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 |
# 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 yaml
import safetensors
import torch
from torch import nn
from torch.nn.functional import scaled_dot_product_attention
from flash_attn.flash_attn_interface import flash_attn_varlen_func
from xformers.ops import SwiGLU
from .rmsnorm import RMSNorm
from .rotary import precompute_freqs_cis, apply_rotary_emb
from .tokenizer import ProteinTokenizer
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
act = config.hidden_act.lower()
if act == "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)
elif act == "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),
)
elif act == "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),
)
else:
raise ValueError(f"Unsupported hidden_act: {config.hidden_act}")
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,
pad_mask: torch.Tensor,
freqs_cis: torch.Tensor,
output_attentions: bool,
max_seqlen: int = None,
cu_seqlens: torch.Tensor = None,
):
attn, contact = self._att_block(self.attention_norm(x), pad_mask, freqs_cis, output_attentions, max_seqlen, cu_seqlens)
x = x + attn
x = x + self._ff_block(self.ffn_norm(x))
return x, contact
def _att_block(
self,
x: torch.Tensor,
pad_mask: torch.Tensor,
freqs_cis: torch.Tensor,
output_attentions: bool,
max_seqlen: int = None,
cu_seqlens: torch.Tensor = None,
):
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 block
attn_weights = None
# Flash attention if the tensors are packed
if cu_seqlens is not None:
attn = flash_attn_varlen_func(
q=xq.squeeze(0),
k=xk.squeeze(0),
v=xv.squeeze(0),
cu_seqlens_q=cu_seqlens,
cu_seqlens_k=cu_seqlens,
max_seqlen_q=max_seqlen,
max_seqlen_k=max_seqlen,
dropout_p=0.0,
causal=False,
)
# Eager attention if attention weights are needed in the output
elif output_attentions:
attn_weights = xq.permute(0, 2, 1, 3) @ xk.permute(0, 2, 3, 1) / (xq.size(-1) ** 0.5)
if pad_mask is not None:
attn_weights = attn_weights + pad_mask.type(attn_weights.dtype)
attn_weights = attn_weights.softmax(-1)
attn = attn_weights @ xv.permute(0, 2, 1, 3)
attn = attn.transpose(1, 2)
# SDPA will pick an appropriate backend otherwise
else:
attn = scaled_dot_product_attention(
query=xq.transpose(1, 2),
key=xk.transpose(1, 2),
value=xv.transpose(1, 2),
attn_mask=pad_mask,
dropout_p=0,
).transpose(1, 2)
attn_scores = self.wo(attn.reshape(batch_size, seq_len, self.config.num_attention_heads * self.d_head))
return (self.resid_dropout(attn_scores), attn_weights)
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)
freqs_cis = precompute_freqs_cis(config.hidden_size // config.num_attention_heads, config.max_length)
# Ensures freqs_cis is moved to the same devices as the model. Non-persistent buffers are not saved in the state_dict.
self.register_buffer("freqs_cis", freqs_cis, persistent=False)
# Initialize weights and apply final processing
self.post_init()
@classmethod
def load(cls, checkpoint_path: str, config_path: str):
with open(config_path, "r") as file:
cfg = yaml.safe_load(file)
model = AMPLIFY(AMPLIFYConfig(**cfg["model"], **cfg["tokenizer"]))
if checkpoint_path.endswith(".safetensors"):
state_dict = safetensors.torch.load_file(checkpoint_path)
elif checkpoint_path.endswith(".pt"):
state_dict = torch.load(checkpoint_path)
else:
raise ValueError(f"Expected checkpoint to be a `.pt` or `.safetensors` file.")
model.load_state_dict(state_dict)
cfg["tokenizer"]["vocab_path"] = "/home/mila/l/lola.lebreton/AMPLIFY/conf/tokenizer/amplify_vocab.txt"
tokenizer = ProteinTokenizer(**cfg["tokenizer"], model_max_length=cfg["trainer"]["train"]["max_length"])
return model, tokenizer
def forward(
self,
src,
position_ids: torch.Tensor = None,
max_seqlen: int = None,
cu_seqlens: torch.Tensor = None,
pad_mask=None,
output_hidden_states=False,
output_attentions=False,
):
# Initialize
hidden_states, attentions = [], []
# We will output all the hidden_states that have an index higher than output_hidden_index
if type(output_hidden_states) == bool and not output_hidden_states:
output_hidden_index = self.config.num_hidden_layers + 1
elif type(output_hidden_states) == int:
output_hidden_index = output_hidden_states
else:
output_hidden_index = 0
# Expand and repeat: (Batch, Length) -> (Batch, Heads, Length, Length)
if pad_mask is not None:
pad_mask = pad_mask.unsqueeze(1).unsqueeze(1).repeat(1, self.config.num_attention_heads, pad_mask.size(-1), 1)
if output_attentions:
pad_mask = torch.where(pad_mask == 1, float(0.0), float("-inf"))
# Checks to be done if inputs are packed sequences
if cu_seqlens is not None:
assert not output_attentions, "Output attentions is not supported when sequences are packed."
assert max_seqlen is not None, "Missing max_seqlen. It must be provided when cu_seqlens are not None."
assert src.shape[0] == 1, "Cumulative sequence lengths are provided but src are not packed."
assert src.is_cuda, "Packing uses an implementation of flash-attention and is only supported on GPU."
# RoPE
if position_ids is not None:
freqs_cis = self.freqs_cis[position_ids]
else:
freqs_cis = self.freqs_cis[: src.shape[1]].unsqueeze(0).repeat(src.shape[0], 1, 1)
# Embedding
x = self.encoder(src)
if self.config.layer_norm_after_embedding:
x = self.layer_norm_1(x)
# Transformer encoder
for idx, layer in enumerate(self.transformer_encoder):
x, attn = layer(x, pad_mask, freqs_cis, output_attentions, max_seqlen, cu_seqlens)
if idx >= output_hidden_index:
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)
|