File size: 8,836 Bytes
fae01f4 | 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 | # SPDX-License-Identifier: Apache-2.0
# 3D causal CNN encoder for the MiniMax H3 visual VAE (inference-only bundle).
import os
import torch.nn as nn
import torch.nn.functional as F
from .attention import maybe_checkpoint
from .conv import SpatialParallelConv3d
from .norm import get_spatial_norm_3d
from .parallel import get_parallel_state, exchange_strides
from .norm import get_group_norm_3d
# ============================================================================
# 3D CNN Components
# ============================================================================
def norm_silu(x, norm, cond=None):
if cond is None:
return F.silu(norm(x))
else:
return F.silu(norm(x, cond))
class Downsample3D(nn.Module):
def __init__(
self,
in_channels,
out_channels,
time_stride=1,
space_stride=2,
padding_mode="zeros",
padding_mode_t=None,
causal=True,
):
super().__init__()
self.time_stride = time_stride
self.space_stride = space_stride
assert time_stride in [1, 2]
assert space_stride in [1, 2, 3]
self.conv = SpatialParallelConv3d(
in_channels,
out_channels,
kernel_size=3,
padding=(1, 0, 0),
stride=(time_stride, space_stride, space_stride),
padding_mode=padding_mode,
padding_mode_t=padding_mode_t,
causal=causal,
)
self.causal = self.conv.causal
self.pad_mode = self.conv.pad_mode
def forward(self, x):
if self.space_stride == 2:
if getattr(self.conv, "spatial_parallel", False):
state = get_parallel_state()
x = exchange_strides(
x,
self.pad_mode,
state["sp_rank"],
state["sp_size"],
state["sp_process_group"],
self.conv.chunk_dim,
)
else:
pad = (0, 1, 0, 1, 0, 0)
x = F.pad(x, pad, mode=self.pad_mode)
return self.conv(x)
class ResnetBlock3D(nn.Module):
def __init__(
self,
in_channels,
out_channels=None,
zq_ch=None,
padding_mode="zeros",
padding_mode_t=None,
causal=True,
use_t_isolated_gn=False,
):
super().__init__()
self.in_channels = in_channels
out_channels = in_channels if out_channels is None else out_channels
self.out_channels = out_channels
self.use_fused_norm = (
os.environ.get("MINIMAX_H3_USE_FUSED_NORM", "false").lower() == "true"
)
if zq_ch is None:
self.norm1 = get_group_norm_3d(in_channels, use_t_isolated_gn=use_t_isolated_gn)
self.norm2 = get_group_norm_3d(out_channels, use_t_isolated_gn=use_t_isolated_gn)
else:
self.norm1 = get_spatial_norm_3d(
in_channels,
zq_ch,
padding_mode=padding_mode,
padding_mode_t=padding_mode_t,
causal=causal,
use_t_isolated_gn=use_t_isolated_gn,
)
self.norm2 = get_spatial_norm_3d(
out_channels,
zq_ch,
padding_mode=padding_mode,
padding_mode_t=padding_mode_t,
causal=causal,
use_t_isolated_gn=use_t_isolated_gn,
)
self.conv1 = SpatialParallelConv3d(
in_channels,
out_channels,
kernel_size=3,
padding=1,
padding_mode=padding_mode,
padding_mode_t=padding_mode_t,
causal=causal,
)
self.conv2 = SpatialParallelConv3d(
out_channels,
out_channels,
kernel_size=3,
padding=1,
padding_mode=padding_mode,
padding_mode_t=padding_mode_t,
causal=causal,
)
if self.in_channels != self.out_channels:
self.nin_shortcut = SpatialParallelConv3d(
in_channels,
out_channels,
kernel_size=1,
padding_mode=padding_mode,
padding_mode_t=padding_mode_t,
causal=causal,
)
def forward(self, x, zq=None):
h = x
if self.use_fused_norm:
h = self.norm1(h, zq)
else:
h = norm_silu(h, self.norm1, zq)
h = self.conv1(h)
if self.use_fused_norm:
h = self.norm2(h, zq)
else:
h = norm_silu(h, self.norm2, zq)
h = self.conv2(h)
if self.in_channels != self.out_channels:
x = self.nin_shortcut(x)
return x + h
class EncoderFCN3D(nn.Module):
def __init__(
self,
ch,
ch_mult,
space_down,
time_down,
num_res_blocks,
in_channels,
z_channels,
double_z=False,
zq_ch=None,
padding_mode="zeros",
padding_mode_t=None,
causal=True,
use_t_isolated_gn=False,
):
super().__init__()
self.ch = ch
self.num_levels = len(ch_mult)
if isinstance(num_res_blocks, int):
self.num_res_blocks = [num_res_blocks] * self.num_levels
else:
self.num_res_blocks = num_res_blocks
self.space_down_factors = space_down
self.time_down_factors = time_down
self.in_channels = in_channels
self.use_fused_norm = (
os.environ.get("MINIMAX_H3_USE_FUSED_NORM", "false").lower() == "true"
)
block_mid = [ch * ch_mult[i] for i in range(self.num_levels)]
block_in = [block_mid[0]] + block_mid[:-1]
block_out = block_mid
conv_kwargs = dict(
padding_mode=padding_mode,
padding_mode_t=padding_mode_t,
causal=causal,
)
self.conv_in = SpatialParallelConv3d(
in_channels, block_in[0], kernel_size=3, padding=1, **conv_kwargs
)
self.down = nn.ModuleList()
for i_level in range(self.num_levels):
down = nn.Module()
down.block = nn.ModuleList()
for i in range(self.num_res_blocks[i_level]):
down.block.append(
ResnetBlock3D(
in_channels=block_in[i_level] if i == 0 else block_mid[i_level],
out_channels=block_mid[i_level],
zq_ch=zq_ch,
use_t_isolated_gn=use_t_isolated_gn,
**conv_kwargs,
)
)
if space_down[i_level] * time_down[i_level] > 1:
down.downsample = Downsample3D(
block_mid[i_level],
block_out[i_level],
time_stride=time_down[i_level],
space_stride=space_down[i_level],
**conv_kwargs,
)
else:
if block_out[i_level] != block_mid[i_level]:
down.downsample = SpatialParallelConv3d(
block_mid[i_level],
block_out[i_level],
kernel_size=1,
**conv_kwargs,
)
self.down.append(down)
if zq_ch is None:
self.norm_out = get_group_norm_3d(
block_out[-1], use_t_isolated_gn=use_t_isolated_gn
)
else:
self.norm_out = get_spatial_norm_3d(
block_out[-1],
zq_ch,
use_t_isolated_gn=use_t_isolated_gn,
**conv_kwargs,
)
self.conv_out = SpatialParallelConv3d(
block_out[-1],
2 * z_channels if double_z else z_channels,
kernel_size=3,
padding=1,
**conv_kwargs,
)
self.gradient_checkpointing = False
def _set_gradient_checkpointing(self, module, value=False):
if hasattr(module, "gradient_checkpointing"):
module.gradient_checkpointing = value
def forward(self, x, zq=None):
h = self.conv_in(x)
for i_level in range(self.num_levels):
for i_block in range(self.num_res_blocks[i_level]):
h = maybe_checkpoint(self, self.down[i_level].block[i_block], h, zq)
if hasattr(self.down[i_level], "downsample"):
h = self.down[i_level].downsample(h)
if self.use_fused_norm:
h = self.norm_out(h, zq)
else:
h = norm_silu(h, self.norm_out, zq)
h = self.conv_out(h)
return h
|