FL2VA/video_vae/norm.py
| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # Torch-native normalization for the MiniMax H3 visual VAE. |
| 3 | import math |
| 4 | import os |
| 5 | |
| 6 | import torch |
| 7 | import torch.distributed as dist |
| 8 | import torch.nn as nn |
| 9 | import torch.nn.functional as F |
| 10 | |
| 11 | from .conv import SpatialParallelConv3d |
| 12 | from .parallel import all_reduce, get_parallel_state |
| 13 | |
| 14 | |
| 15 | def _validate_activation(activation): |
| 16 | valid_activations = {"identity", "silu", "relu"} |
| 17 | if activation not in valid_activations: |
| 18 | raise ValueError( |
| 19 | f"Unsupported activation: {activation}. Supported: {valid_activations}" |
| 20 | ) |
| 21 | |
| 22 | |
| 23 | def _apply_activation(x, activation): |
| 24 | _validate_activation(activation) |
| 25 | if activation == "identity": |
| 26 | return x |
| 27 | if activation == "silu": |
| 28 | return F.silu(x) |
| 29 | return F.relu(x) |
| 30 | |
| 31 | |
| 32 | def _merge_time_to_batch(x): |
| 33 | batch, channels, depth, height, width = x.shape |
| 34 | return ( |
| 35 | x.permute(0, 2, 1, 3, 4) |
| 36 | .contiguous() |
| 37 | .view(batch * depth, channels, 1, height, width) |
| 38 | ) |
| 39 | |
| 40 | |
| 41 | def _split_time_from_batch(x, batch): |
| 42 | batch_depth, channels, _, height, width = x.shape |
| 43 | depth = batch_depth // batch |
| 44 | return ( |
| 45 | x.view(batch, depth, channels, height, width) |
| 46 | .permute(0, 2, 1, 3, 4) |
| 47 | .contiguous() |
| 48 | ) |
| 49 | |
| 50 | |
| 51 | def fused_group_norm(x, num_groups, weight, bias, eps=1e-5, activation="silu"): |
| 52 | out = F.group_norm(x, num_groups, weight=weight, bias=bias, eps=eps) |
| 53 | return _apply_activation(out, activation) |
| 54 | |
| 55 | |
| 56 | def fused_spatial_norm( |
| 57 | f, |
| 58 | num_groups, |
| 59 | norm_weight, |
| 60 | norm_bias, |
| 61 | dynamic_scale, |
| 62 | dynamic_bias, |
| 63 | eps=1e-5, |
| 64 | activation="silu", |
| 65 | ): |
| 66 | norm_f = F.group_norm( |
| 67 | f, |
| 68 | num_groups, |
| 69 | weight=norm_weight, |
| 70 | bias=norm_bias, |
| 71 | eps=eps, |
| 72 | ) |
| 73 | out = norm_f * dynamic_scale + dynamic_bias |
| 74 | return _apply_activation(out, activation) |
| 75 | |
| 76 | |
| 77 | class DummyAffine(torch.nn.Module): |
| 78 | def __init__(self, num_channels, affine=True): |
| 79 | super().__init__() |
| 80 | if affine: |
| 81 | self.weight = torch.nn.Parameter(torch.ones(num_channels)) |
| 82 | self.bias = torch.nn.Parameter(torch.zeros(num_channels)) |
| 83 | else: |
| 84 | self.register_parameter("weight", None) |
| 85 | self.register_parameter("bias", None) |
| 86 | |
| 87 | def forward(self, input): |
| 88 | if self.weight is None: |
| 89 | return input |
| 90 | shape = [1, -1] + [1] * (input.dim() - 2) |
| 91 | return input * self.weight.view(*shape) + self.bias.view(*shape) |
| 92 | |
| 93 | |
| 94 | class FusedGroupNorm3D(torch.nn.Module): |
| 95 | """Compatibility wrapper implemented with native PyTorch ops.""" |
| 96 | |
| 97 | def __init__( |
| 98 | self, |
| 99 | num_groups, |
| 100 | num_channels, |
| 101 | eps=1e-5, |
| 102 | affine=True, |
| 103 | activation="silu", |
| 104 | cond_channels=None, |
| 105 | use_t_isolated_gn=False, |
| 106 | padding_mode="zeros", |
| 107 | padding_mode_t=None, |
| 108 | causal=True, |
| 109 | ): |
| 110 | super().__init__() |
| 111 | _validate_activation(activation) |
| 112 | self.num_groups = num_groups |
| 113 | self.num_channels = num_channels |
| 114 | self.eps = eps |
| 115 | self.affine = affine |
| 116 | self.activation = activation |
| 117 | self.use_t_isolated_gn = use_t_isolated_gn |
| 118 | |
| 119 | if cond_channels is not None: |
| 120 | self.use_spatial_affine = True |
| 121 | self.norm_layer = DummyAffine(num_channels, affine=affine) |
| 122 | self.conv_y = SpatialParallelConv3d( |
| 123 | cond_channels, |
| 124 | num_channels, |
| 125 | kernel_size=1, |
| 126 | padding_mode=padding_mode, |
| 127 | padding_mode_t=padding_mode_t, |
| 128 | causal=causal, |
| 129 | ) |
| 130 | self.conv_b = SpatialParallelConv3d( |
| 131 | cond_channels, |
| 132 | num_channels, |
| 133 | kernel_size=1, |
| 134 | padding_mode=padding_mode, |
| 135 | padding_mode_t=padding_mode_t, |
| 136 | causal=causal, |
| 137 | ) |
| 138 | else: |
| 139 | self.use_spatial_affine = False |
| 140 | if self.affine: |
| 141 | self.weight = torch.nn.Parameter(torch.ones(num_channels)) |
| 142 | self.bias = torch.nn.Parameter(torch.zeros(num_channels)) |
| 143 | else: |
| 144 | self.register_parameter("weight", None) |
| 145 | self.register_parameter("bias", None) |
| 146 | |
| 147 | def forward(self, f, cond=None): |
| 148 | need_reshape = self.use_t_isolated_gn and f.dim() == 5 |
| 149 | batch = f.shape[0] if need_reshape else None |
| 150 | f_size = f.shape[-3:] |
| 151 | if need_reshape: |
| 152 | f = _merge_time_to_batch(f) |
| 153 | |
| 154 | if self.use_spatial_affine: |
| 155 | scale = self.conv_y(cond) |
| 156 | bias = self.conv_b(cond) |
| 157 | if math.prod(scale.shape[-3:]) * math.prod(bias.shape[-3:]) > 1: |
| 158 | scale = F.interpolate(scale, size=f_size, mode="nearest") |
| 159 | bias = F.interpolate(bias, size=f_size, mode="nearest") |
| 160 | if need_reshape: |
| 161 | scale = _merge_time_to_batch(scale) |
| 162 | bias = _merge_time_to_batch(bias) |
| 163 | out = fused_spatial_norm( |
| 164 | f, |
| 165 | self.num_groups, |
| 166 | self.norm_layer.weight, |
| 167 | self.norm_layer.bias, |
| 168 | scale, |
| 169 | bias, |
| 170 | self.eps, |
| 171 | self.activation, |
| 172 | ) |
| 173 | else: |
| 174 | if cond is not None: |
| 175 | raise NotImplementedError("Dynamic affine is not defined") |
| 176 | weight = self.weight if self.affine else None |
| 177 | bias = self.bias if self.affine else None |
| 178 | out = fused_group_norm( |
| 179 | f, self.num_groups, weight, bias, self.eps, self.activation |
| 180 | ) |
| 181 | |
| 182 | if need_reshape: |
| 183 | out = _split_time_from_batch(out, batch) |
| 184 | return out |
| 185 | |
| 186 | |
| 187 | class SpatialParallelGroupNorm(nn.GroupNorm): |
| 188 | def __init__( |
| 189 | self, |
| 190 | *args, |
| 191 | **kwargs, |
| 192 | ): |
| 193 | super().__init__(*args, **kwargs) |
| 194 | self.spatial_parallel = False |
| 195 | |
| 196 | def _compute_stats(self, input): |
| 197 | batch, channels = input.shape[0], input.shape[1] |
| 198 | spatial_dims = input.shape[2:] |
| 199 | spatial_size = math.prod(spatial_dims) |
| 200 | |
| 201 | groups = self.num_groups |
| 202 | x = input.reshape(batch, groups, channels // groups, -1).to(torch.float32) |
| 203 | |
| 204 | local_sum = x.sum(dim=(2, 3)) |
| 205 | local_square_sum = (x * x).sum(dim=(2, 3)) |
| 206 | local_n = (channels // groups) * spatial_size |
| 207 | local_n_tensor = torch.full_like(local_sum, float(local_n)) |
| 208 | |
| 209 | stats = torch.stack([local_sum, local_square_sum, local_n_tensor], dim=0) |
| 210 | |
| 211 | local_process_group = get_parallel_state()["local_process_group"] |
| 212 | stats = all_reduce(stats, dist.ReduceOp.SUM, local_process_group) |
| 213 | |
| 214 | total_sum = stats[0] |
| 215 | total_square_sum = stats[1] |
| 216 | total_n = stats[2] |
| 217 | |
| 218 | mean = total_sum / total_n |
| 219 | var = (total_square_sum / total_n) - mean**2 |
| 220 | return mean, var |
| 221 | |
| 222 | def forward(self, input): |
| 223 | if not self.spatial_parallel: |
| 224 | return nn.GroupNorm.forward(self, input) |
| 225 | |
| 226 | batch, channels = input.shape[0], input.shape[1] |
| 227 | orig_shape = input.shape |
| 228 | |
| 229 | mean, var = self._compute_stats(input) |
| 230 | x = input.reshape(batch, self.num_groups, channels // self.num_groups, -1) |
| 231 | |
| 232 | mean = mean.unsqueeze(-1).unsqueeze(-1) |
| 233 | var = var.unsqueeze(-1).unsqueeze(-1) |
| 234 | x = (x - mean) / torch.sqrt(var + self.eps) |
| 235 | x = x.reshape(orig_shape) |
| 236 | |
| 237 | if self.affine: |
| 238 | shape = [1, -1] + [1] * (len(orig_shape) - 2) |
| 239 | x *= self.weight.view(*shape) |
| 240 | x += self.bias.view(*shape) |
| 241 | |
| 242 | return x |
| 243 | |
| 244 | |
| 245 | class TemporalIsolatedSpatialParallelGroupNorm(SpatialParallelGroupNorm): |
| 246 | def forward(self, input): |
| 247 | if input.dim() == 5: |
| 248 | batch = input.shape[0] |
| 249 | input = _merge_time_to_batch(input) |
| 250 | output = super().forward(input) |
| 251 | return _split_time_from_batch(output, batch) |
| 252 | return super().forward(input) |
| 253 | |
| 254 | |
| 255 | |
| 256 | |
| 257 | |
| 258 | |
| 259 | |
| 260 | |
| 261 | class SpatialNorm3D(nn.Module): |
| 262 | def __init__( |
| 263 | self, |
| 264 | f_channels, |
| 265 | zq_channels, |
| 266 | padding_mode="zeros", |
| 267 | padding_mode_t=None, |
| 268 | causal=True, |
| 269 | use_t_isolated_gn=False, |
| 270 | ): |
| 271 | super().__init__() |
| 272 | norm_cls = ( |
| 273 | TemporalIsolatedSpatialParallelGroupNorm |
| 274 | if use_t_isolated_gn |
| 275 | else SpatialParallelGroupNorm |
| 276 | ) |
| 277 | self.norm_layer = norm_cls( |
| 278 | num_groups=32, num_channels=f_channels, eps=1e-6, affine=True |
| 279 | ) |
| 280 | |
| 281 | self.conv_y = SpatialParallelConv3d( |
| 282 | zq_channels, |
| 283 | f_channels, |
| 284 | kernel_size=1, |
| 285 | padding_mode=padding_mode, |
| 286 | padding_mode_t=padding_mode_t, |
| 287 | causal=causal, |
| 288 | ) |
| 289 | self.conv_b = SpatialParallelConv3d( |
| 290 | zq_channels, |
| 291 | f_channels, |
| 292 | kernel_size=1, |
| 293 | padding_mode=padding_mode, |
| 294 | padding_mode_t=padding_mode_t, |
| 295 | causal=causal, |
| 296 | ) |
| 297 | |
| 298 | def forward(self, f, zq): |
| 299 | f_size = f.shape[-3:] |
| 300 | norm_f = self.norm_layer(f) |
| 301 | scale = self.conv_y(zq) |
| 302 | bias = self.conv_b(zq) |
| 303 | |
| 304 | if math.prod(scale.shape[-3:]) * math.prod(bias.shape[-3:]) > 1: |
| 305 | scale = F.interpolate(scale, size=f_size, mode="nearest") |
| 306 | bias = F.interpolate(bias, size=f_size, mode="nearest") |
| 307 | |
| 308 | return norm_f * scale + bias |
| 309 | |
| 310 | |
| 311 | def get_spatial_norm_3d( |
| 312 | num_channels, |
| 313 | cond_channels, |
| 314 | *, |
| 315 | padding_mode="zeros", |
| 316 | padding_mode_t=None, |
| 317 | causal=True, |
| 318 | use_t_isolated_gn=False, |
| 319 | ): |
| 320 | if os.environ.get("MINIMAX_H3_USE_FUSED_NORM", "false").lower() == "true": |
| 321 | return FusedGroupNorm3D( |
| 322 | num_groups=32, |
| 323 | num_channels=num_channels, |
| 324 | eps=1e-6, |
| 325 | affine=True, |
| 326 | cond_channels=cond_channels, |
| 327 | use_t_isolated_gn=use_t_isolated_gn, |
| 328 | padding_mode=padding_mode, |
| 329 | padding_mode_t=padding_mode_t, |
| 330 | causal=causal, |
| 331 | ) |
| 332 | return SpatialNorm3D( |
| 333 | num_channels, |
| 334 | cond_channels, |
| 335 | padding_mode=padding_mode, |
| 336 | padding_mode_t=padding_mode_t, |
| 337 | causal=causal, |
| 338 | use_t_isolated_gn=use_t_isolated_gn, |
| 339 | ) |
| 340 | |
| 341 | |
| 342 | def get_group_norm_3d(num_channels, use_t_isolated_gn=False): |
| 343 | if os.environ.get("MINIMAX_H3_USE_FUSED_NORM", "false").lower() == "true": |
| 344 | return FusedGroupNorm3D( |
| 345 | num_groups=32, |
| 346 | num_channels=num_channels, |
| 347 | eps=1e-6, |
| 348 | affine=True, |
| 349 | use_t_isolated_gn=use_t_isolated_gn, |
| 350 | ) |
| 351 | |
| 352 | norm_cls = ( |
| 353 | TemporalIsolatedSpatialParallelGroupNorm |
| 354 | if use_t_isolated_gn |
| 355 | else SpatialParallelGroupNorm |
| 356 | ) |
| 357 | return norm_cls(num_groups=32, num_channels=num_channels, eps=1e-6, affine=True) |
| 358 | |