Spaces:
Running
on
Zero
Running
on
Zero
File size: 7,842 Bytes
bcb05d1 |
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 |
from typing import *
import torch
import torch.nn as nn
from ..basic import SparseTensor
from ..attention import SparseMultiHeadAttention, SerializeMode, SpatialSparseAttention
from ...norm import LayerNorm32
from .blocks import SparseFeedForwardNet
class ModulatedSparseTransformerBlock(nn.Module):
"""
Sparse Transformer block (MSA + FFN) with adaptive layer norm conditioning.
"""
def __init__(
self,
channels: int,
num_heads: int,
mlp_ratio: float = 4.0,
attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "full",
window_size: Optional[int] = None,
shift_sequence: Optional[int] = None,
shift_window: Optional[Tuple[int, int, int]] = None,
serialize_mode: Optional[SerializeMode] = None,
use_checkpoint: bool = False,
use_rope: bool = False,
qk_rms_norm: bool = False,
qkv_bias: bool = True,
share_mod: bool = False,
):
super().__init__()
self.use_checkpoint = use_checkpoint
self.share_mod = share_mod
self.norm1 = LayerNorm32(channels, elementwise_affine=False, eps=1e-6)
self.norm2 = LayerNorm32(channels, elementwise_affine=False, eps=1e-6)
self.attn = SparseMultiHeadAttention(
channels,
num_heads=num_heads,
attn_mode=attn_mode,
window_size=window_size,
shift_sequence=shift_sequence,
shift_window=shift_window,
serialize_mode=serialize_mode,
qkv_bias=qkv_bias,
use_rope=use_rope,
qk_rms_norm=qk_rms_norm,
)
self.mlp = SparseFeedForwardNet(
channels,
mlp_ratio=mlp_ratio,
)
if not share_mod:
self.adaLN_modulation = nn.Sequential(
nn.SiLU(),
nn.Linear(channels, 6 * channels, bias=True)
)
def _forward(self, x: SparseTensor, mod: torch.Tensor) -> SparseTensor:
if self.share_mod:
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(6, dim=1)
else:
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(mod).chunk(6, dim=1)
h = x.replace(self.norm1(x.feats))
h = h * (1 + scale_msa) + shift_msa
h = self.attn(h)
h = h * gate_msa
x = x + h
h = x.replace(self.norm2(x.feats))
h = h * (1 + scale_mlp) + shift_mlp
h = self.mlp(h)
h = h * gate_mlp
x = x + h
return x
def forward(self, x: SparseTensor, mod: torch.Tensor) -> SparseTensor:
if self.use_checkpoint:
return torch.utils.checkpoint.checkpoint(self._forward, x, mod, use_reentrant=False)
else:
return self._forward(x, mod)
class ModulatedSparseTransformerCrossBlock(nn.Module):
"""
Sparse Transformer cross-attention block (MSA + MCA + FFN) with adaptive layer norm conditioning.
"""
def __init__(
self,
channels: int,
ctx_channels: int,
num_heads: int,
mlp_ratio: float = 4.0,
attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "full",
window_size: Optional[int] = None,
shift_sequence: Optional[int] = None,
shift_window: Optional[Tuple[int, int, int]] = None,
serialize_mode: Optional[SerializeMode] = None,
use_checkpoint: bool = False,
compression_version: str = "v2",
use_rope: bool = False,
qk_rms_norm: bool = False,
qk_rms_norm_cross: bool = False,
qkv_bias: bool = True,
share_mod: bool = False,
use_ssa: bool = True,
num_kv_heads: int = 2,
compression_block_size: int = 4,
selection_block_size: int = 8,
topk: int = 8,
resolution: int = 64,
):
super().__init__()
self.use_checkpoint = use_checkpoint
self.share_mod = share_mod
self.norm1 = LayerNorm32(channels, elementwise_affine=False, eps=1e-6)
self.norm2 = LayerNorm32(channels, elementwise_affine=True, eps=1e-6)
self.norm3 = LayerNorm32(channels, elementwise_affine=False, eps=1e-6)
if use_ssa:
self.self_attn = SpatialSparseAttention(
channels,
num_q_heads=num_heads,
num_kv_heads=num_kv_heads,
head_dim=channels//num_heads,
compression_block_size=compression_block_size,
compression_version=compression_version,
selection_block_size=selection_block_size,
topk=topk,
window_size=window_size,
shift_window=shift_window,
resolution=resolution,
)
else:
self.self_attn = SparseMultiHeadAttention(
channels,
num_heads=num_heads,
type="self",
attn_mode=attn_mode,
window_size=window_size,
shift_sequence=shift_sequence,
shift_window=shift_window,
serialize_mode=serialize_mode,
qkv_bias=qkv_bias,
use_rope=use_rope,
qk_rms_norm=qk_rms_norm,
)
self.cross_attn = SparseMultiHeadAttention(
channels,
ctx_channels=ctx_channels,
num_heads=num_heads,
type="cross",
attn_mode="full",
qkv_bias=qkv_bias,
qk_rms_norm=qk_rms_norm_cross,
)
self.mlp = SparseFeedForwardNet(
channels,
mlp_ratio=mlp_ratio,
)
if not share_mod:
self.adaLN_modulation = nn.Sequential(
nn.SiLU(),
nn.Linear(channels, 6 * channels, bias=True)
)
def _forward(self, x: SparseTensor, mod: torch.Tensor, context: torch.Tensor) -> SparseTensor:
if self.share_mod:
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(6, dim=1)
else:
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(mod).chunk(6, dim=1)
h = x.replace(self.norm1(x.feats))
feats_h = h.feats
layouts = h.layout
ada_r1 = []
for i in range(len(layouts)):
ada_r1.append(feats_h[layouts[i]] * (1 + scale_msa[i:i+1]) + shift_msa[i:i+1])
h = h.replace(torch.cat(ada_r1, dim=0))
h = self.self_attn(h)
feats_h = h.feats
layouts = h.layout
ada_r2 = []
for i in range(len(layouts)):
ada_r2.append(feats_h[layouts[i]] * gate_msa[i:i+1])
h = h.replace(torch.cat(ada_r2, dim=0))
x = x + h
h = x.replace(self.norm2(x.feats))
h = self.cross_attn(h, context)
x = x + h
h = x.replace(self.norm3(x.feats))
feats_h = h.feats
layouts = h.layout
ada_r3 = []
for i in range(len(layouts)):
ada_r3.append(feats_h[layouts[i]] * (1 + scale_mlp[i:i+1]) + shift_mlp[i:i+1])
h = h.replace(torch.cat(ada_r3, dim=0))
h = self.mlp(h)
feats_h = h.feats
layouts = h.layout
ada_r4 = []
for i in range(len(layouts)):
ada_r4.append(feats_h[layouts[i]] * gate_mlp[i:i+1])
h = h.replace(torch.cat(ada_r4, dim=0))
x = x + h
return x
def forward(self, x: SparseTensor, mod: torch.Tensor, context: torch.Tensor) -> SparseTensor:
if self.use_checkpoint:
return torch.utils.checkpoint.checkpoint(self._forward, x, mod, context, use_reentrant=False)
else:
return self._forward(x, mod, context)
|