File size: 11,167 Bytes
ae8ade0 | 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 | """One-block feature Predictor used by the initialization sweep."""
from __future__ import annotations
import copy
from typing import Literal
import torch
from torch import nn
from torch.utils.checkpoint import checkpoint
from wan.modules.causal_model import CausalWanAttentionBlock
InitializationMethod = Literal[
"teacher_full",
"random_full",
"teacher_identity",
"random_identity",
"full_zero",
]
GateMode = Literal["baseline", "learned", "constant"]
class PreviousFeatureGate(nn.Module):
"""Per-token scalar reliability gate for the normalized previous feature."""
def __init__(
self,
dim: int,
hidden_dim: int = 128,
initial_bias: float = 4.6,
) -> None:
super().__init__()
self.net = nn.Sequential(
nn.Linear(2 * dim, hidden_dim),
nn.SiLU(),
nn.Linear(hidden_dim, 1),
)
nn.init.zeros_(self.net[-1].weight)
nn.init.constant_(self.net[-1].bias, initial_bias)
def forward(
self, anchor_normalized: torch.Tensor, previous_normalized: torch.Tensor
) -> torch.Tensor:
return torch.sigmoid(
self.net(torch.cat([anchor_normalized, previous_normalized], dim=-1))
)
class TripleFeatureFusion(nn.Module):
"""Fuse target latent tokens, anchor hidden, and previous-chunk hidden."""
def __init__(
self,
dim: int = 1536,
gate_mode: GateMode = "baseline",
gate_hidden_dim: int = 128,
gate_initial_bias: float = 4.6,
gate_floor: float = 0.0,
constant_gate: float = 1.0,
) -> None:
super().__init__()
if gate_mode not in {"baseline", "learned", "constant"}:
raise ValueError(f"Unknown gate mode: {gate_mode}")
if not 0.0 <= constant_gate <= 1.0:
raise ValueError("constant_gate must be in [0, 1]")
if not 0.0 <= gate_floor < 1.0:
raise ValueError("gate_floor must be in [0, 1)")
self.current_norm = nn.LayerNorm(dim, eps=1e-6)
self.anchor_norm = nn.LayerNorm(dim, eps=1e-6)
self.previous_norm = nn.LayerNorm(dim, eps=1e-6)
self.gate_mode = gate_mode
self.constant_gate = float(constant_gate)
self.gate_floor = float(gate_floor)
self.gate_override: float | None = None
self.gate = (
PreviousFeatureGate(dim, gate_hidden_dim, gate_initial_bias)
if gate_mode == "learned"
else None
)
self.last_gate: torch.Tensor | None = None
self.proj_in = nn.Linear(3 * dim, 2 * dim)
self.activation = nn.SiLU()
self.proj_out = nn.Linear(2 * dim, dim)
def forward(
self,
current: torch.Tensor,
anchor: torch.Tensor,
previous: torch.Tensor,
) -> torch.Tensor:
if current.shape != anchor.shape or anchor.shape != previous.shape:
raise ValueError(
"TripleFeatureFusion inputs must have identical shapes: "
f"{current.shape}, {anchor.shape}, {previous.shape}"
)
anchor_normalized = self.anchor_norm(anchor)
previous_normalized = self.previous_norm(previous)
if self.gate_override is not None:
gate = previous_normalized.new_full(
(*previous_normalized.shape[:-1], 1), self.gate_override
)
elif self.gate_mode == "learned":
assert self.gate is not None
raw_gate = self.gate(anchor_normalized, previous_normalized)
gate = self.gate_floor + (1.0 - self.gate_floor) * raw_gate
elif self.gate_mode == "constant":
gate = previous_normalized.new_full(
(*previous_normalized.shape[:-1], 1), self.constant_gate
)
else:
gate = previous_normalized.new_ones(
*previous_normalized.shape[:-1], 1
)
# Gate after normalization. A positive per-token scalar applied before
# LayerNorm would be normalized away and could not test reliability.
self.last_gate = gate.detach()
return self.proj_out(
self.activation(
self.proj_in(
torch.cat(
[
self.current_norm(current),
anchor_normalized,
gate * previous_normalized,
],
dim=-1,
)
)
)
)
def _new_random_block(
teacher_block: CausalWanAttentionBlock,
) -> CausalWanAttentionBlock:
block = CausalWanAttentionBlock(
cross_attn_type="t2v_cross_attn",
dim=teacher_block.dim,
ffn_dim=teacher_block.ffn_dim,
num_heads=teacher_block.num_heads,
local_attn_size=teacher_block.local_attn_size,
sink_size=teacher_block.self_attn.sink_size,
qk_norm=teacher_block.qk_norm,
cross_attn_norm=teacher_block.cross_attn_norm,
eps=teacher_block.eps,
)
# This matches CausalWanModel.init_weights rather than PyTorch Linear's
# Kaiming default.
for module in block.modules():
if isinstance(module, nn.Linear):
nn.init.xavier_uniform_(module.weight)
if module.bias is not None:
nn.init.zeros_(module.bias)
return block
def _zero_residual_branch_outputs(block: CausalWanAttentionBlock) -> None:
for projection in (
block.self_attn.o,
block.cross_attn.o,
block.ffn[2],
):
nn.init.zeros_(projection.weight)
if projection.bias is not None:
nn.init.zeros_(projection.bias)
def initialize_predictor_block(
teacher_block: CausalWanAttentionBlock,
method: InitializationMethod,
) -> CausalWanAttentionBlock:
"""Create a trainable FP32 block under one controlled initialization."""
if method in {"teacher_full", "teacher_identity", "full_zero"}:
block = copy.deepcopy(teacher_block).float()
elif method in {"random_full", "random_identity"}:
block = _new_random_block(teacher_block).float()
else:
raise ValueError(f"Unknown initialization method: {method}")
if method in {"teacher_identity", "random_identity"}:
_zero_residual_branch_outputs(block)
elif method == "full_zero":
for parameter in block.parameters():
nn.init.zeros_(parameter)
return block
class SingleBlockPredictor(nn.Module):
"""Predict final hidden as an anchor residual through one causal Wan block."""
def __init__(
self,
block: CausalWanAttentionBlock,
dim: int = 1536,
gradient_checkpointing: bool = True,
gate_mode: GateMode = "baseline",
gate_hidden_dim: int = 128,
gate_initial_bias: float = 4.6,
gate_floor: float = 0.0,
constant_gate: float = 1.0,
) -> None:
super().__init__()
self.fusion = TripleFeatureFusion(
dim,
gate_mode=gate_mode,
gate_hidden_dim=gate_hidden_dim,
gate_initial_bias=gate_initial_bias,
gate_floor=gate_floor,
constant_gate=constant_gate,
)
self.block = block
self.residual_out = nn.Linear(dim, dim)
nn.init.zeros_(self.residual_out.weight)
nn.init.zeros_(self.residual_out.bias)
self.gradient_checkpointing = gradient_checkpointing
def fusion_parameters(self) -> list[nn.Parameter]:
return list(self.fusion.parameters()) + list(self.residual_out.parameters())
def gate_parameters(self) -> list[nn.Parameter]:
return list(self.fusion.gate.parameters()) if self.fusion.gate else []
def fusion_parameters_without_gate(self) -> list[nn.Parameter]:
gate_ids = {id(parameter) for parameter in self.gate_parameters()}
return [
parameter for parameter in self.fusion_parameters()
if id(parameter) not in gate_ids
]
def block_parameters(self) -> list[nn.Parameter]:
return list(self.block.parameters())
def set_block_trainable(self, enabled: bool) -> None:
self.block.requires_grad_(enabled)
def forward(
self,
*,
current_tokens: torch.Tensor,
anchor_hidden: torch.Tensor,
previous_hidden: torch.Tensor,
timestep_modulation: torch.Tensor,
grid_sizes: torch.Tensor,
freqs: torch.Tensor,
history_k: torch.Tensor,
history_v: torch.Tensor,
cross_k: torch.Tensor,
cross_v: torch.Tensor,
current_start: int,
) -> torch.Tensor:
fused = self.fusion(current_tokens, anchor_hidden, previous_hidden)
batch, sequence, dim = fused.shape
history_length = history_k.shape[1]
if history_k.shape[0] != batch:
raise ValueError(
f"history_k {history_k.shape} incompatible with "
f"batch={batch}"
)
if history_v.shape != history_k.shape:
raise ValueError("History K/V shapes differ")
seq_lens = torch.full(
(batch,),
sequence,
dtype=torch.long,
device="cpu",
)
context = fused.new_zeros(batch, 1, dim)
def run_block(value: torch.Tensor) -> torch.Tensor:
# A fresh cache is required for checkpoint recomputation because
# CausalWanSelfAttention updates cache indices and current K/V.
heads = history_k.shape[2]
head_dim = history_k.shape[3]
current_k = history_k.new_empty(batch, sequence, heads, head_dim)
current_v = history_v.new_empty(batch, sequence, heads, head_dim)
kv_cache = {
"k": torch.cat([history_k, current_k], dim=1),
"v": torch.cat([history_v, current_v], dim=1),
"global_end_index": torch.tensor(
[current_start], device=value.device, dtype=torch.long
),
"local_end_index": torch.tensor(
[history_length], device=value.device, dtype=torch.long
),
}
crossattn_cache = {
"k": cross_k,
"v": cross_v,
"is_init": True,
}
return self.block(
value,
e=timestep_modulation,
seq_lens=seq_lens,
grid_sizes=grid_sizes,
freqs=freqs,
context=context,
context_lens=None,
block_mask=None,
kv_cache=kv_cache,
crossattn_cache=crossattn_cache,
current_start=current_start,
cache_start=None,
)
if self.training and self.gradient_checkpointing:
transformed = checkpoint(
run_block,
fused,
use_reentrant=False,
)
else:
transformed = run_block(fused)
return anchor_hidden + self.residual_out(transformed)
|