File size: 13,490 Bytes
f30f923 | 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 | # SPDX-License-Identifier: Apache-2.0
# ViT3D decoder for the MiniMax H3 visual VAE (inference-only bundle).
import torch
import torch.nn as nn
import torch.distributed as dist
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.modeling_utils import ModelMixin
from diffusers.utils import logging
from .attention import maybe_checkpoint
from .base_module import TransformerBlock, RotaryEmbeddingND
from .flash import make_block_causal_mask_mod
from .func import create_token_ids
from .parallel import get_subseq, gather_subseq, get_parallel_state
logger = logging.get_logger(__name__)
def _linear_with_module_dtype(linear, tensor, out_dtype=None):
weight = getattr(linear, "weight", None)
target_dtype = getattr(weight, "dtype", tensor.dtype)
output = linear(tensor.to(target_dtype))
if out_dtype is not None and output.dtype != out_dtype:
output = output.to(out_dtype)
return output
def _make_seq_len_mask_mod(seq_len, base_mask_mod=None):
if base_mask_mod is None:
def mask_mod(b, h, q_idx, kv_idx, seqlen_info, aux_tensors):
return (q_idx < seq_len) & (kv_idx < seq_len)
mask_mod.block_sparse_cache_key = ("seq_len", seq_len)
return mask_mod
def mask_mod(b, h, q_idx, kv_idx, seqlen_info, aux_tensors):
return (
(q_idx < seq_len)
& (kv_idx < seq_len)
& base_mask_mod(b, h, q_idx, kv_idx, seqlen_info, aux_tensors)
)
base_cache_key = getattr(base_mask_mod, "block_sparse_cache_key", None)
if base_cache_key is not None:
mask_mod.block_sparse_cache_key = ("seq_len", seq_len, base_cache_key)
if hasattr(base_mask_mod, "use_fast_sampling"):
mask_mod.use_fast_sampling = base_mask_mod.use_fast_sampling
return mask_mod
def _pack_tensors_3d(tensors, patch_size, patch_size_t):
batch_size, num_channels_tensors, temporal, height, width = tensors.shape
tensors = tensors.view(
batch_size,
num_channels_tensors,
temporal // patch_size_t,
patch_size_t,
height // patch_size,
patch_size,
width // patch_size,
patch_size,
)
tensors = tensors.permute(0, 2, 4, 6, 1, 3, 5, 7)
tensors = tensors.reshape(
batch_size,
(temporal // patch_size_t) * (height // patch_size) * (width // patch_size),
num_channels_tensors * patch_size_t * patch_size * patch_size,
)
return tensors
def _unpack_tensors_3d(tensors, patch_size, patch_size_t, temporal, height, width):
batch_size, num_patches, channels = tensors.shape
num_channels_tensors = channels // (patch_size_t * patch_size * patch_size)
tensors = tensors.view(
batch_size,
temporal // patch_size_t,
height // patch_size,
width // patch_size,
num_channels_tensors,
patch_size_t,
patch_size,
patch_size,
)
tensors = tensors.permute(0, 4, 1, 5, 2, 6, 3, 7).contiguous()
tensors = tensors.reshape(batch_size, num_channels_tensors, temporal, height, width)
return tensors
class ViTBase(ModelMixin, ConfigMixin):
"""Base class for ViT Encoder and Decoder with common functionality."""
_supports_gradient_checkpointing = True
_no_split_modules = ["TransformerBlock"]
gradient_checkpointing_mode = "full"
def _set_gradient_checkpointing(self, module, value=False):
if hasattr(module, "gradient_checkpointing"):
module.gradient_checkpointing = value
def set_spatial_parallel(self, enabled):
self.spatial_parallel = enabled
if hasattr(self, "transformer_blocks"):
for block in self.transformer_blocks:
block.attn.spatial_parallel = enabled
def _init_weights(self):
def basic_init(m):
if isinstance(m, nn.Linear):
nn.init.xavier_uniform_(m.weight)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
self.apply(basic_init)
def init_mask_config(self, dim, is_3d=False):
self._mask_dim = dim
self._mask_is_3d = is_3d
self.register_buffer("mask_token", torch.zeros(1, 1, dim))
def set_mask_config(self, mask_config):
self.mask_prob = mask_config.get("mask_prob", 0.0)
self.mask_enabled = self.mask_prob > 0
self.mask_style = mask_config.get("mask_style", "replace")
if self.mask_enabled and self.mask_style == "drop" and self.mask_prob < 1.0:
logger.warning("mask_style='drop' with mask_prob < 1.0")
if self._mask_is_3d:
self.temporal_scale_range = mask_config.get("temporal_scale_range", (0.3, 0.5))
self.spatial_scale_range = mask_config.get("spatial_scale_range", (0.1, 0.25))
self.min_mask_ratio = mask_config.get("min_mask_ratio", 0.75)
self.max_mask_ratio = mask_config.get("max_mask_ratio", 0.95)
else:
self.spatial_scale_range = mask_config.get("spatial_scale_range", (0.15, 0.15))
self.min_mask_ratio = mask_config.get("min_mask_ratio", 0.5)
self.max_mask_ratio = mask_config.get("max_mask_ratio", 0.75)
self.aspect_ratio_range = mask_config.get("aspect_ratio_range", (0.75, 1.5))
self.max_retries = mask_config.get("max_retries", 100)
if self.mask_enabled and self.mask_style == "drop" and getattr(self, "t_causal", False):
logger.warning("mask_style='drop' with t_causal may cause issues")
if self.mask_enabled and "mask_token" in self._buffers:
del self._buffers["mask_token"]
self.mask_token = nn.Parameter(torch.randn(1, 1, self._mask_dim) * 0.02)
def init_suffix_tokens(self, dim, num_register_tokens, has_cls_token=True):
self.num_register_tokens = num_register_tokens
if num_register_tokens > 0:
self.register_tokens = nn.Parameter(torch.randn(1, num_register_tokens, dim) * 0.02)
else:
self.register_tokens = None
if has_cls_token:
self.cls_token = nn.Parameter(torch.randn(1, 1, dim) * 0.02)
def apply_mask_preprocess(self, hidden_states, img_ids, patch_dims, num_suffix):
if self.training and self.mask_enabled:
raise NotImplementedError(
"mask modeling is not supported in this inference-only bundle"
)
return hidden_states, img_ids
def forward_transformer_blocks(self, hidden_states, rotary_pos_emb, pack_info=None):
if pack_info is None:
pack_info = {}
for block in self.transformer_blocks:
hidden_states = maybe_checkpoint(
self, block, hidden_states, rotary_pos_emb, pack_info
)
return hidden_states
def _pad_for_sp(self, hidden_states, img_ids, pack_info=None):
if pack_info is None:
pack_info = {}
if not self.spatial_parallel:
return hidden_states, img_ids, pack_info, 0
seq_len = hidden_states.shape[1]
sp_size = get_parallel_state().get("sp_size", 1)
pad_len = (-seq_len) % sp_size
if pad_len == 0:
return hidden_states, img_ids, pack_info, 0
hidden_states = torch.nn.functional.pad(hidden_states, (0, 0, 0, pad_len))
img_ids = torch.nn.functional.pad(img_ids, (0, 0, 0, pad_len))
pack_info = dict(pack_info)
base_mask_mod = pack_info.get("mask_mod")
pack_info["mask_mod"] = _make_seq_len_mask_mod(seq_len, base_mask_mod)
pack_info.pop("block_sparse", None)
return hidden_states, img_ids, pack_info, pad_len
@staticmethod
def _unpad_for_sp(hidden_states, pad_len):
if pad_len == 0:
return hidden_states
return hidden_states[:, :-pad_len, :]
def apply_mask_postprocess(self, hidden_states, num_patches):
if self.training and self.mask_enabled and self.mask_style == "drop":
raise NotImplementedError(
"mask modeling is not supported in this inference-only bundle"
)
return hidden_states
class ViT3DDecoder(ViTBase):
"""Vision Transformer Video Decoder using TransformerBlock."""
@register_to_config
def __init__(
self,
patch_size: int = 16,
patch_size_t: int = 4,
t_causal: bool = False,
in_channels: int = 16,
out_channels: int = 3,
num_layers: int = 24,
heads: int = 16,
dim_head: int = 64,
norm_type: str = "layer_norm",
norm_affine: bool = True,
qk_norm_type: str = None,
qk_norm_affine: bool = False,
ffn_activation_fn: str = "gelu",
ffn_use_gated: bool = False,
rope_theta: float = 100.0,
rope_dim_ratio: float = 1.0,
bias: bool = True,
eps: float = 1e-5,
num_register_tokens: int = 4,
mask_config: dict = {},
**kwargs,
):
super().__init__()
dim = heads * dim_head
rope_apply_dim = int(dim_head * rope_dim_ratio)
self.pos_embed = RotaryEmbeddingND(rope_apply_dim, rope_theta, n_dim=3, use_angle=True)
self.x_embedder = nn.Linear(in_channels, dim)
self.init_suffix_tokens(dim, num_register_tokens, has_cls_token=False)
self.t_causal = t_causal
self.transformer_blocks = nn.ModuleList(
[
TransformerBlock(
heads=heads,
dim_head=dim_head,
norm_type=norm_type,
norm_affine=norm_affine,
qk_norm_type=qk_norm_type,
qk_norm_affine=qk_norm_affine,
ffn_activation_fn=ffn_activation_fn,
ffn_use_gated=ffn_use_gated,
bias=bias,
eps=eps,
**kwargs,
)
for _ in range(num_layers)
]
)
self.spatial_parallel = False
for block in self.transformer_blocks:
block.attn.spatial_parallel = False
self.norm_out = nn.LayerNorm(dim, elementwise_affine=norm_affine, eps=eps)
patch_dim = out_channels * patch_size_t * patch_size * patch_size
self.proj_out = nn.Linear(dim, patch_dim)
self.init_mask_config(dim, is_3d=True)
self.set_mask_config(mask_config)
self._init_weights()
self.gradient_checkpointing = False
if len(kwargs) > 0 and (not dist.is_initialized() or dist.get_rank() == 0):
logger.warning(f"Unused kwargs: {kwargs}")
def forward(self, x: torch.Tensor) -> torch.Tensor:
self.loss_info = {}
B, C, latent_T, latent_H, latent_W = x.shape
patch_size = self.config.patch_size
patch_size_t = self.config.patch_size_t
num_suffix = 1 + self.num_register_tokens
hidden_states = _pack_tensors_3d(x, 1, 1)
latent_size = (latent_T, latent_H, latent_W)
with torch.autocast("cuda", enabled=False):
hidden_states = _linear_with_module_dtype(self.x_embedder, hidden_states, hidden_states.dtype)
num_patches = hidden_states.shape[1]
tokens = [hidden_states]
if self.register_tokens is not None:
register_tokens = self.register_tokens.expand(B, -1, -1)
tokens.append(register_tokens)
cls_token = torch.zeros_like(hidden_states[:, 0:1, :])
tokens.append(cls_token)
hidden_states = torch.cat(tokens, dim=1)
patch_dims = [latent_T, latent_H, latent_W]
img_ids = create_token_ids(latent_size, x.device, x.dtype).expand(B, -1, -1)
suffix_ids = torch.zeros((B, num_suffix, 3), device=x.device, dtype=img_ids.dtype)
img_ids = torch.cat([img_ids, suffix_ids], dim=1)
hidden_states, img_ids = self.apply_mask_preprocess(hidden_states, img_ids, patch_dims, num_suffix)
pack_info = {}
if self.t_causal:
spatial_size = latent_H * latent_W
mask_mod = make_block_causal_mask_mod(
num_tokens=num_patches,
block_size=spatial_size,
suffix=True,
)
pack_info["mask_mod"] = mask_mod
hidden_states, img_ids, pack_info, sp_pad_len = self._pad_for_sp(hidden_states, img_ids, pack_info)
rotary_pos_emb = self.pos_embed(img_ids)
if self.spatial_parallel:
hidden_states = get_subseq(hidden_states)
for block in self.transformer_blocks:
hidden_states = maybe_checkpoint(
self, block, hidden_states, rotary_pos_emb, pack_info
)
if self.spatial_parallel:
hidden_states = gather_subseq(hidden_states)
hidden_states = self._unpad_for_sp(hidden_states, sp_pad_len)
hidden_states = self.norm_out(hidden_states)
hidden_states = self.apply_mask_postprocess(hidden_states, num_patches)
with torch.autocast("cuda", enabled=False):
output = _linear_with_module_dtype(self.proj_out, hidden_states, hidden_states.dtype)
output = output[:, :num_patches, :]
video_t = latent_size[0] * patch_size_t
video_h = latent_size[1] * patch_size
video_w = latent_size[2] * patch_size
output = _unpack_tensors_3d(output, patch_size, patch_size_t, video_t, video_h, video_w)
return output
|