File size: 17,843 Bytes
7f316fe | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 | # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the CC-by-NC license found in the
# LICENSE file in the root directory of this source tree.
# Part of this implementation is adapted from https://github.com/facebookresearch/DiT
# which is released under NonCommercial-4.0 license
# Part of this implementation is adapted from https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
# which is released under MIT license
# Part of this implementation is adapted from https://github.com/louaaron/Score-Entropy-Discrete-Diffusion
# which is released under MIT license
import math
from typing import Optional, Tuple
import torch
import torch.nn.functional as F
from torch.nn.attention.flex_attention import flex_attention
from einops import rearrange
from omegaconf import OmegaConf
from omegaconf.dictconfig import DictConfig
from torch import nn, Tensor
from transformers import EsmModel
from . import rotary
import pdb
def lengths_to_offsets(lengths: torch.Tensor) -> torch.Tensor:
# lengths: (B,) long -> offsets: (B+1,) long
return F.pad(lengths.cumsum(0), (1, 0))
def build_seq_ids(lengths: torch.Tensor, device=None) -> torch.Tensor:
# lengths: (B,) -> seq_ids: (T,) where T=sum(lengths)
device = device or lengths.device
return torch.repeat_interleave(torch.arange(lengths.numel(), device=device), lengths)
def make_score_mod_for_intra_sequence_only(lengths: torch.Tensor):
seq_ids = build_seq_ids(lengths, device=lengths.device) # (T,)
def score_mod(scores, b, h, q_idx, k_idx):
# scores: (..., q_block, k_block)
same = (seq_ids[q_idx] == seq_ids[k_idx]) # bool, broadcastable to scores
# Set cross-sequence scores to a very negative value *without* in-place ops
neg_large = torch.finfo(scores.dtype).min # dtype-safe "−large" (no -inf issues)
return torch.where(same, scores, torch.full_like(scores, neg_large))
return score_mod
# def make_score_mod_for_intra_sequence_only(lengths: torch.Tensor):
# """
# Returns a score_mod callback for FlexAttention that sets logits to -inf
# when query and key belong to different sequences (ragged, no padding).
# """
# seq_ids = build_seq_ids(lengths, device=lengths.device) # (T,)
# def score_mod(scores, b, h, q_idx, k_idx):
# # scores: (..., q_block, k_block); q_idx/k_idx: flat token indices in [0..T-1]
# original_type = scores.dtype
# scores = scores.float()
# same = (seq_ids[q_idx] == seq_ids[k_idx]).to(scores.dtype)
# scores += (same - 1) * 1e9 # subtract a large number for cross-seq pairs
# return scores.to(original_type)
# return score_mod
def bias_dropout_add_scale(
x: Tensor, scale: Tensor, residual: Optional[Tensor], prob: float, training: bool
) -> Tensor:
return residual + scale * F.dropout(x, p=prob, training=training)
def modulate(x: Tensor, shift: Tensor, scale: Tensor) -> Tensor:
return x * (1 + scale) + shift
class LayerNorm(nn.Module):
def __init__(self, dim: int):
super().__init__()
self.weight = nn.Parameter(torch.ones([dim]))
self.dim = dim
def forward(self, x: Tensor) -> Tensor:
with torch.amp.autocast("cuda", enabled=False):
y = F.layer_norm(x.float(), [self.dim])
if y.dim() == 3: # (B,S,H)
scale = self.weight[None, None, :]
elif y.dim() == 2: # (T,H)
scale = self.weight[None, :]
else:
raise ValueError(f"LayerNorm expects 2D/3D, got {y.shape}")
return y * scale
class TimestepEmbedder(nn.Module):
"""
Embeds scalar timesteps into vector representations.
"""
def __init__(self, hidden_size: int, frequency_embedding_size: int = 256):
super().__init__()
self.mlp = nn.Sequential(
nn.Linear(frequency_embedding_size, hidden_size, bias=True),
nn.SiLU(),
nn.Linear(hidden_size, hidden_size, bias=True),
)
self.frequency_embedding_size = frequency_embedding_size
@staticmethod
def timestep_embedding(time: Tensor, dim: int, max_period: int = 10000) -> Tensor:
"""
Create sinusoidal timestep embeddings.
:param t: a 1-D Tensor of N indices, one per batch element.
These may be fractional.
:param dim: the dimension of the output.
:param max_period: controls the minimum frequency of the embeddings.
:return: an (N, D) Tensor of positional embeddings.
"""
half = dim // 2
freqs = torch.exp(
-math.log(max_period)
* torch.arange(start=0, end=half, dtype=torch.float32)
/ half
).to(device=time.device)
args = time[:, None].float() * freqs[None]
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
if dim % 2:
embedding = torch.cat(
[embedding, torch.zeros_like(embedding[:, :1])], dim=-1
)
return embedding
def forward(self, time: Tensor) -> Tensor:
t_freq = self.timestep_embedding(time=time, dim=self.frequency_embedding_size)
t_emb = self.mlp(t_freq)
return t_emb
class DDiTBlock(nn.Module):
def __init__(
self,
dim: int,
n_heads: int,
cond_dim: int,
mlp_ratio: int = 4,
dropout: float = 0.1,
):
super().__init__()
assert dim % n_heads == 0, "dim must be devisable by n_heads"
self.n_heads = n_heads
self.dim = dim
self.dropout = dropout
self.head_dim = self.dim // self.n_heads
self.norm1 = LayerNorm(dim=dim)
self.qw = nn.Linear(dim, dim, bias=False)
self.kw = nn.Linear(dim, dim, bias=False)
self.vw = nn.Linear(dim, dim, bias=False)
self.attn_out = nn.Linear(dim, dim, bias=False)
self.dropout1 = nn.Dropout(dropout)
self.norm2 = LayerNorm(dim=dim)
self.mlp = nn.Sequential(
nn.Linear(dim, mlp_ratio * dim, bias=True),
nn.GELU(approximate="tanh"),
nn.Linear(mlp_ratio * dim, dim, bias=True),
)
self.adaLN_modulation = nn.Linear(cond_dim, 6 * dim, bias=True)
self.adaLN_modulation.weight.data.zero_()
self.adaLN_modulation.bias.data.zero_()
def forward(
self,
x: Tensor, # (T, H) flat tokens; T = sum(lengths)
rotary_caches: tuple, # (cos, sin)
c: Tensor # (B, cond_dim)
) -> Tensor:
(
shift_msa,
scale_msa,
gate_msa,
shift_mlp,
scale_mlp,
gate_mlp,
) = self.adaLN_modulation(c)[:, None].chunk(6, dim=2)
seq_ids = build_seq_ids(lengths, device=x.device) # (T,)
shift_msa_b = shift_msa.squeeze(1)
scale_msa_b = scale_msa.squeeze(1)
gate_msa_b = gate_msa.squeeze(1)
shift_mlp_b = shift_mlp.squeeze(1)
scale_mlp_b = scale_mlp.squeeze(1)
gate_mlp_b = gate_mlp.squeeze(1)
# Token-wise (T,H)
shift_msa_t = shift_msa_b[seq_ids]
scale_msa_t = scale_msa_b[seq_ids]
gate_msa_t = gate_msa_b[seq_ids]
shift_mlp_t = shift_mlp_b[seq_ids]
scale_mlp_t = scale_mlp_b[seq_ids]
gate_mlp_t = gate_mlp_b[seq_ids]
x_skip = x
x = modulate(self.norm1(x), shift=shift_msa_t, scale=scale_msa_t) # (T, H)
q = self.qw(x); k = self.kw(x); v = self.vw(x) # (T, H)
T = x.shape[0]
q = q.view(T, self.n_heads, self.head_dim)
k = k.view(T, self.n_heads, self.head_dim)
v = v.view(T, self.n_heads, self.head_dim)
with torch.amp.autocast("cuda", enabled=False):
##################
# TODO:
##################
q = q.to(dtype); k = k.to(dtype)
# fold heads into batch
q = q.transpose(0, 1).contiguous() # (Hh, T, Dh)
k = k.transpose(0, 1).contiguous()
v = v.transpose(0, 1).contiguous()
q = q.unsqueeze(0) # -> (1, Hh, T, Dh)
k = k.unsqueeze(0) # -> (1, Hh, T, Dh)
v = v.unsqueeze(0) # -> (1, Hh, T, Dh)
score_mod = make_score_mod_for_intra_sequence_only(lengths)
attn_out = flex_attention(q, k, v, score_mod=score_mod) # (Hh, T, Dh)
attn_out = attn_out.squeeze(0) # -> (Hh, T, Dh)
x = attn_out.transpose(0, 1).contiguous().view(T, self.dim) # (T,H)
x = bias_dropout_add_scale(self.attn_out(x), gate_msa_t, x_skip, self.dropout, self.training)
x = bias_dropout_add_scale(
self.mlp(modulate(self.norm2(x), shift=shift_mlp_t, scale=scale_mlp_t)),
gate_mlp_t, x, self.dropout, self.training
)
return x
class DDitFinalLayer(nn.Module):
def __init__(self, hidden_size: int, out_channels: int, cond_dim: int):
super().__init__()
self.norm_final = LayerNorm(hidden_size)
self.linear = nn.Linear(hidden_size, out_channels)
self.linear.weight.data.zero_()
self.linear.bias.data.zero_()
self.adaLN_modulation = nn.Linear(cond_dim, 2 * hidden_size, bias=True)
self.adaLN_modulation.weight.data.zero_()
self.adaLN_modulation.bias.data.zero_()
def forward(self, x: Tensor, c: Tensor) -> Tensor:
shift, scale = self.adaLN_modulation(c)[:, None].chunk(2, dim=2)
x = modulate(x=self.norm_final(x), shift=shift, scale=scale)
x = self.linear(x)
return x
# -------------------------------------------------------------
# NEW: small helper to keep λ ≥ 0
class Positive(nn.Module):
def __init__(self):
super().__init__()
self.bias = nn.Parameter(torch.tensor(0.0))
def forward(self, x: Tensor) -> Tensor:
return F.softplus(x + self.bias)
# NEW: project token hidden states to "slot" hidden states (n+1)
class SlotProjector(nn.Module):
"""
Builds n+1 slot states from n token states using learnable boundaries.
slot i uses left h_{i-1} and right h_i (with learned BOS/EOS).
"""
def __init__(self, hidden_size: int):
super().__init__()
self.bos = nn.Parameter(torch.zeros(hidden_size))
self.eos = nn.Parameter(torch.zeros(hidden_size))
self.proj = nn.Linear(2 * hidden_size, hidden_size, bias=True)
self.act = nn.GELU()
def forward(self, h_tok: Tensor) -> Tensor:
# h_tok: (B, n, H)
B, n, H = h_tok.shape
bos = self.bos.expand(B, 1, H) # (B, 1, H)
eos = self.eos.expand(B, 1, H) # (B, 1, H)
# left/right neighbors for the n+1 between-token slots
left = torch.cat([bos, h_tok], dim=1) # (B, n+1, H)
right = torch.cat([h_tok, eos], dim=1) # (B, n+1, H)
slots = torch.cat([left, right], dim=-1) # (B, n+1, 2H)
return self.act(self.proj(slots)) # (B, n+1, H)
# NEW: multi-head output layer for edit flows
class EditFlowsHead(nn.Module):
"""
Produces:
- λ_ins: (B, n+1), Q_ins: (B, n+1, V)
- λ_del: (B, n), λ_sub: (B, n), Q_sub: (B, n, V)
"""
def __init__(self, hidden_size: int, vocab_size: int, cond_dim: int):
super().__init__()
self.norm = LayerNorm(hidden_size)
self.cond = nn.Linear(cond_dim, 2 * hidden_size, bias=True)
self.cond.weight.data.zero_()
self.cond.bias.data.zero_()
# token-position heads (n)
self.lambda_del = nn.Linear(hidden_size, 1, bias=True)
self.lambda_sub = nn.Linear(hidden_size, 1, bias=True)
self.q_sub = nn.Linear(hidden_size, vocab_size, bias=True)
# slot-position heads (n+1)
self.slot_proj = SlotProjector(hidden_size)
self.lambda_ins = nn.Linear(hidden_size, 1, bias=True)
self.q_ins = nn.Linear(hidden_size, vocab_size, bias=True)
self.to_positive = Positive()
# init: keep outputs small at start
for m in [self.lambda_del, self.lambda_sub, self.lambda_ins, self.q_sub, self.q_ins]:
nn.init.zeros_(m.weight); nn.init.zeros_(m.bias)
def forward(self, h_tok: Tensor, c: Tensor) -> tuple[Tensor, Tensor, Tensor, Tensor, Tensor]:
# FiLM-style modulation like your DDitFinalLayer
shift, scale = self.cond(c)[:, None].chunk(2, dim=-1)
h_tok = modulate(self.norm(h_tok), shift=shift, scale=scale) # (B, n, H)
# token positions (n)
lam_del = self.to_positive(self.lambda_del(h_tok)).squeeze(-1) # (B, n)
lam_sub = self.to_positive(self.lambda_sub(h_tok)).squeeze(-1) # (B, n)
q_sub = self.q_sub(h_tok) # (B, n, V) -- removed softmax here, may need to adjust tensor shape
# slot positions (n+1)
h_slot = self.slot_proj(h_tok) # (B, n+1, H)
lam_ins = self.to_positive(self.lambda_ins(h_slot)).squeeze(-1) # (B, n+1)
q_ins = self.q_ins(h_slot) # (B, n+1, V) -- removed softmax here, may need to adjust tensor shape
return lam_ins, q_ins, lam_del, lam_sub, q_sub
class Transformer(nn.Module):
def __init__(self, vocab_size: int, masked: bool, config: DictConfig):
super().__init__()
if isinstance(config, dict):
config = OmegaConf.create(config)
self.config = config
self.vocab_size = vocab_size
self.pad_id = getattr(config, "pad_id", 0)
add_token = 1 if masked else 0 # keep if you need a mask token elsewhere
# ESM-2 embedding approach (similar to gpm_model.py)
esm_model_name = getattr(config, "esm_model_name", "facebook/esm2_t33_650M_UR50D")
freeze_esm = getattr(config, "freeze_esm", True)
if esm_model_name is not None:
self.tok_embedder = EsmModel.from_pretrained(esm_model_name)
tok_embed_dim = self.tok_embedder.config.hidden_size
if freeze_esm:
for param in self.tok_embedder.parameters():
param.requires_grad = False
self.tok_embedder.eval()
# Project from ESM hidden size to model hidden size
self.tok_embed_to_hidden = nn.Linear(tok_embed_dim, config.hidden_size)
self.vocab_embed = None # Not needed when using ESM
else:
# Fallback to embedding layer if ESM is not used
self.tok_embedder = None
self.tok_embed_to_hidden = None # Not needed when using standard embedding
self.vocab_embed = nn.Embedding(self.vocab_size + add_token, config.hidden_size)
self.time_embedding = TimestepEmbedder(hidden_size=config.cond_dim)
self.rotary_emb = rotary.Rotary(dim=config.hidden_size // config.n_heads)
self.blocks = nn.ModuleList(
[
DDiTBlock(
dim=config.hidden_size,
n_heads=config.n_heads,
cond_dim=config.cond_dim,
dropout=config.dropout,
)
for _ in range(config.n_blocks)
]
)
# CHANGED: use EditFlowsHead instead of DDitFinalLayer
self.output_layer = EditFlowsHead(
hidden_size=config.hidden_size,
vocab_size=vocab_size + add_token,
cond_dim=config.cond_dim,
)
def _embed_ragged(self, x_t, mask):
# Handle device - use ESM device or vocab_embed device
if self.tok_embedder is not None:
device = next(self.tok_embedder.parameters()).device
else:
device = self.vocab_embed.weight.device
if self.tok_embedder is not None:
x_t_emb = self.tok_embedder(x_t, mask).last_hidden_state
x = self.tok_embed_to_hidden(x_t_emb)
else:
# Fallback to standard embedding
x = self.vocab_embed(x_t) # (T, H)
# positions from RoPE helper:
B, S = x_t.shape
lengths = torch.full((B,), S, device=device, dtype=torch.long)
positions = self.rotary_emb.positions_like(lengths) # (T,)
return x, lengths, positions
def forward(self, x_t, time: torch.Tensor):
"""
x_t: List[LongTensor] of variable lengths OR a (B,S) tensor
time: (B,) float in [0,1]
returns lists (ragged): lam_ins, q_ins, lam_del, lam_sub, q_sub
"""
# 0) pad x_t
x_t = torch.nn.utils.rnn.pad_sequence(x_t, batch_first=True, padding_value=self.pad_id)
mask = (x_t != self.pad_id).long()
# 1) embed ragged
x_emb = self._embed_ragged(x_t, mask)
# 2) time conditioning
B = x_t.shape[0]
t_emb = F.silu(self.time_embedding(time)) # (B, cond_dim)
# 3) rotary caches (you can keep your existing cache logic)
###############
# TODO:
###############
# 4) transformer blocks
with torch.amp.autocast("cuda", dtype=torch.bfloat16):
for blk in self.blocks:
x = blk(
x=x_emb,
mask=mask,
rotary_caches=rotary_caches,
c=t_emb,
)
#############
# TODO:
#############
lam_ins, q_ins, lam_del, lam_sub, q_sub = zip(*outs)
# 6) return ragged lists (keep padding outside the model if needed)
return list(lam_ins), list(q_ins), list(lam_del), list(lam_sub), list(q_sub)
|