hunyuan.py
| 1 | # Licensed under the TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT (the "License"); |
| 2 | # you may not use this file except in compliance with the License. |
| 3 | # You may obtain a copy of the License at |
| 4 | # |
| 5 | # https://github.com/Tencent-Hunyuan/HunyuanImage-3.0/blob/main/LICENSE |
| 6 | # |
| 7 | # Unless required by applicable law or agreed to in writing, software |
| 8 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 9 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 10 | # See the License for the specific language governing permissions and |
| 11 | # limitations under the License. |
| 12 | # ============================================================================== |
| 13 | |
| 14 | import math |
| 15 | import random |
| 16 | import re |
| 17 | import warnings |
| 18 | from dataclasses import dataclass |
| 19 | from typing import TYPE_CHECKING, List, Union, Optional, Dict, Any, Tuple, Callable |
| 20 | |
| 21 | import torch |
| 22 | import torch.nn.functional as F |
| 23 | import torch.utils.checkpoint |
| 24 | from einops import rearrange |
| 25 | from torch import Tensor |
| 26 | from torch import nn |
| 27 | from torch.cuda import nvtx |
| 28 | from transformers.activations import ACT2FN |
| 29 | from transformers.cache_utils import Cache, StaticCache |
| 30 | from transformers.generation.logits_process import LogitsProcessorList |
| 31 | from transformers.generation.stopping_criteria import StoppingCriteriaList |
| 32 | from transformers.generation.utils import GenerationMixin, GenerationConfig, ALL_CACHE_NAMES |
| 33 | from transformers.modeling_outputs import ( |
| 34 | BaseModelOutputWithPast, |
| 35 | CausalLMOutputWithPast, |
| 36 | ) |
| 37 | from transformers.modeling_utils import PreTrainedModel |
| 38 | from transformers.utils import ( |
| 39 | ModelOutput, |
| 40 | add_start_docstrings, |
| 41 | add_start_docstrings_to_model_forward, |
| 42 | is_flash_attn_2_available, |
| 43 | logging, |
| 44 | ) |
| 45 | |
| 46 | if TYPE_CHECKING: |
| 47 | from transformers.generation.streamers import BaseStreamer |
| 48 | |
| 49 | try: |
| 50 | import flashinfer |
| 51 | except Exception as e: |
| 52 | flashinfer = None |
| 53 | |
| 54 | from .autoencoder_kl_3d import AutoencoderKLConv3D |
| 55 | from .configuration_hunyuan import HunyuanImage3Config |
| 56 | from .hunyuan_image_3_pipeline import HunyuanImage3Text2ImagePipeline, FlowMatchDiscreteScheduler |
| 57 | from .image_processor import HunyuanImage3ImageProcessor |
| 58 | from .siglip2 import Siglip2VisionTransformer, LightProjector |
| 59 | from .tokenizer_wrapper import TokenizerWrapper, ImageInfo, JointImageInfo |
| 60 | from .system_prompt import get_system_prompt, t2i_system_prompts |
| 61 | |
| 62 | |
| 63 | logger = logging.get_logger(__name__) |
| 64 | |
| 65 | |
| 66 | if is_flash_attn_2_available(): |
| 67 | from flash_attn import flash_attn_func |
| 68 | |
| 69 | # Type aliases |
| 70 | BatchRaggedImages = Union[torch.Tensor, List[Union[torch.Tensor, List[torch.Tensor]]]] |
| 71 | BatchRaggedTensor = Union[torch.Tensor, List[torch.Tensor]] |
| 72 | |
| 73 | |
| 74 | _CONFIG_FOR_DOC = "HunyuanImage3Config" |
| 75 | |
| 76 | Hunyuan_START_DOCSTRING = r""" |
| 77 | This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the |
| 78 | library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads |
| 79 | etc.) |
| 80 | |
| 81 | This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. |
| 82 | Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage |
| 83 | and behavior. |
| 84 | |
| 85 | Parameters: |
| 86 | config ([`HunyuanImage3Config`]): |
| 87 | Model configuration class with all the parameters of the model. Initializing with a config file does not |
| 88 | load the weights associated with the model, only the configuration. Check out the |
| 89 | [`~PreTrainedModel.from_pretrained`] method to load the model weights. |
| 90 | """ |
| 91 | |
| 92 | # ======================================================= |
| 93 | # Helper Functions |
| 94 | # ======================================================= |
| 95 | |
| 96 | def default(val, d): |
| 97 | return val if val is not None else d |
| 98 | |
| 99 | |
| 100 | def to_device(data, device): |
| 101 | if device is None: |
| 102 | return data |
| 103 | if isinstance(data, torch.Tensor): |
| 104 | return data.to(device) |
| 105 | elif isinstance(data, list): |
| 106 | return [to_device(x, device) for x in data] |
| 107 | else: |
| 108 | return data |
| 109 | |
| 110 | |
| 111 | def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor: |
| 112 | """ |
| 113 | This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, |
| 114 | num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) |
| 115 | """ |
| 116 | batch, num_key_value_heads, slen, head_dim = hidden_states.shape |
| 117 | if n_rep == 1: |
| 118 | return hidden_states |
| 119 | hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim) |
| 120 | return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim) |
| 121 | |
| 122 | |
| 123 | def real_batched_index_select(t, dim, idx): |
| 124 | """ index_select for batched index and batched t """ |
| 125 | assert t.ndim >= 2 and idx.ndim >= 2, f"{t.ndim=} {idx.ndim=}" |
| 126 | assert len(t) == len(idx), f"{len(t)=} != {len(idx)=}" |
| 127 | return torch.stack([torch.index_select(t[i], dim - 1, idx[i]) for i in range(len(t))]) |
| 128 | |
| 129 | |
| 130 | # ======================================================= |
| 131 | # Module Functions |
| 132 | # ======================================================= |
| 133 | |
| 134 | def timestep_embedding(t, dim, max_period=10000): |
| 135 | """ |
| 136 | Create sinusoidal timestep embeddings. |
| 137 | |
| 138 | Args: |
| 139 | t (torch.Tensor): a 1-D Tensor of N indices, one per batch element. These may be fractional. |
| 140 | dim (int): the dimension of the output. |
| 141 | max_period (int): controls the minimum frequency of the embeddings. |
| 142 | |
| 143 | Returns: |
| 144 | embedding (torch.Tensor): An (N, D) Tensor of positional embeddings. |
| 145 | |
| 146 | .. ref_link: https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py |
| 147 | """ |
| 148 | half = dim // 2 |
| 149 | freqs = torch.exp( |
| 150 | -math.log(max_period) |
| 151 | * torch.arange(start=0, end=half, dtype=torch.float32) |
| 152 | / half |
| 153 | ).to(device=t.device) |
| 154 | args = t[:, None].float() * freqs[None] |
| 155 | embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1) |
| 156 | if dim % 2: |
| 157 | embedding = torch.cat( |
| 158 | [embedding, torch.zeros_like(embedding[:, :1])], dim=-1 |
| 159 | ) |
| 160 | return embedding |
| 161 | |
| 162 | |
| 163 | def conv_nd(dims, *args, **kwargs): |
| 164 | """ |
| 165 | Create a 1D, 2D, or 3D convolution module. |
| 166 | """ |
| 167 | if dims == 1: |
| 168 | return nn.Conv1d(*args, **kwargs) |
| 169 | elif dims == 2: |
| 170 | return nn.Conv2d(*args, **kwargs) |
| 171 | elif dims == 3: |
| 172 | return nn.Conv3d(*args, **kwargs) |
| 173 | raise ValueError(f"unsupported dimensions: {dims}") |
| 174 | |
| 175 | |
| 176 | def linear(*args, **kwargs): |
| 177 | """ |
| 178 | Create a linear module. |
| 179 | """ |
| 180 | return nn.Linear(*args, **kwargs) |
| 181 | |
| 182 | |
| 183 | def avg_pool_nd(dims, *args, **kwargs): |
| 184 | """ |
| 185 | Create a 1D, 2D, or 3D average pooling module. |
| 186 | """ |
| 187 | if dims == 1: |
| 188 | return nn.AvgPool1d(*args, **kwargs) |
| 189 | elif dims == 2: |
| 190 | return nn.AvgPool2d(*args, **kwargs) |
| 191 | elif dims == 3: |
| 192 | return nn.AvgPool3d(*args, **kwargs) |
| 193 | raise ValueError(f"unsupported dimensions: {dims}") |
| 194 | |
| 195 | |
| 196 | def zero_module(module): |
| 197 | """ |
| 198 | Zero out the parameters of a module and return it. |
| 199 | """ |
| 200 | for p in module.parameters(): |
| 201 | p.detach().zero_() |
| 202 | return module |
| 203 | |
| 204 | |
| 205 | def normalization(channels, **kwargs): |
| 206 | """ |
| 207 | Make a standard normalization layer. |
| 208 | |
| 209 | :param channels: number of input channels. |
| 210 | :return: a nn.Module for normalization. |
| 211 | """ |
| 212 | return nn.GroupNorm(32, channels, **kwargs) |
| 213 | |
| 214 | |
| 215 | def topkgating( |
| 216 | logits: Tensor, |
| 217 | topk: int, |
| 218 | group_limited_greedy: bool = False, |
| 219 | n_group: int = None, |
| 220 | topk_group: int = None, |
| 221 | norm_topk_prob: bool = True, |
| 222 | routed_scaling_factor: float = 1.0, |
| 223 | capacity_factor: float = 1.0, |
| 224 | drop_tokens: bool = False, |
| 225 | ): |
| 226 | logits = logits.float() |
| 227 | gates = F.softmax(logits, dim=1) |
| 228 | |
| 229 | if group_limited_greedy: |
| 230 | group_shape = list(gates.shape[:-1]) + [n_group, gates.shape[-1] // n_group] |
| 231 | group_scores = ( |
| 232 | gates.reshape(group_shape).max(dim=-1).values |
| 233 | ) # [n, n_group] |
| 234 | group_idx = torch.topk( |
| 235 | group_scores, topk_group, dim=-1, sorted=False |
| 236 | )[ |
| 237 | 1 |
| 238 | ] # [n, top_k_group] |
| 239 | group_mask = torch.zeros_like(group_scores) # [n, n_group] |
| 240 | group_mask.scatter_(1, group_idx, 1) # [n, n_group] |
| 241 | score_mask = ( |
| 242 | group_mask.unsqueeze(-1) |
| 243 | .expand( |
| 244 | group_shape |
| 245 | ) |
| 246 | .reshape(list(gates.shape)) |
| 247 | ) # [n, e] |
| 248 | gates = gates.masked_fill(~score_mask.bool(), 0.0) |
| 249 | |
| 250 | num_experts = int(gates.shape[1]) |
| 251 | # Top-k router probability and corresponding expert indices for each token. |
| 252 | # Shape: [tokens_per_group, num_selected_experts]. |
| 253 | expert_gate, expert_index = torch.topk(gates, topk) |
| 254 | expert_mask = F.one_hot(expert_index, num_experts) |
| 255 | # For a given token, determine if it was routed to a given expert. |
| 256 | # Shape: [tokens_per_group, num_experts] |
| 257 | expert_mask_aux = expert_mask.max(dim=-2)[0] |
| 258 | tokens_per_group_and_expert = torch.mean(expert_mask_aux.float(), dim=-2) |
| 259 | router_prob_per_group_and_expert = torch.mean(gates.float(), dim=-2) |
| 260 | l_aux = num_experts ** 2 * torch.mean(tokens_per_group_and_expert * router_prob_per_group_and_expert) |
| 261 | |
| 262 | if drop_tokens: |
| 263 | expert_capacity = int(max(topk, topk * gates.shape[0] // gates.shape[1]) * capacity_factor) |
| 264 | else: |
| 265 | expert_index_flat = expert_index.flatten() |
| 266 | tokens_per_expert = torch.bincount(expert_index_flat, minlength=num_experts) |
| 267 | expert_capacity = torch.max(tokens_per_expert).item() |
| 268 | |
| 269 | if norm_topk_prob and topk > 1: |
| 270 | gates_s = torch.clamp( |
| 271 | torch.matmul(expert_mask.float(), gates.unsqueeze(-1)).sum(dim=1), min=torch.finfo(gates.dtype).eps |
| 272 | ) |
| 273 | router_probs = gates / gates_s |
| 274 | else: |
| 275 | router_probs = gates * routed_scaling_factor |
| 276 | # Make num_selected_experts the leading axis to ensure that top-1 choices |
| 277 | # have priority over top-2 choices, which have priority over top-3 choices, |
| 278 | # etc. |
| 279 | expert_index = torch.transpose(expert_index, 0, 1) |
| 280 | # Shape: [num_selected_experts * tokens_per_group] |
| 281 | expert_index = expert_index.reshape(-1) |
| 282 | |
| 283 | # Create mask out of indices. |
| 284 | # Shape: [tokens_per_group * num_selected_experts, num_experts]. |
| 285 | expert_mask = F.one_hot(expert_index, num_experts).to(torch.int32) |
| 286 | exp_counts = torch.sum(expert_mask, dim=0).detach() |
| 287 | |
| 288 | # Experts have a fixed capacity that we cannot exceed. A token's priority |
| 289 | # within the expert's buffer is given by the masked, cumulative capacity of |
| 290 | # its target expert. |
| 291 | # Shape: [tokens_per_group * num_selected_experts, num_experts]. |
| 292 | token_priority = torch.cumsum(expert_mask, dim=0) * expert_mask - 1 |
| 293 | # Shape: [num_selected_experts, tokens_per_group, num_experts]. |
| 294 | token_priority = token_priority.reshape((topk, -1, num_experts)) |
| 295 | # Shape: [tokens_per_group, num_selected_experts, num_experts]. |
| 296 | token_priority = torch.transpose(token_priority, 0, 1) |
| 297 | # For each token, across all selected experts, select the only non-negative |
| 298 | # (unmasked) priority. Now, for group G routing to expert E, token T has |
| 299 | # non-negative priority (i.e. token_priority[G,T,E] >= 0) if and only if E |
| 300 | # is its targeted expert. |
| 301 | # Shape: [tokens_per_group, num_experts]. |
| 302 | token_priority = torch.max(token_priority, dim=1)[0] |
| 303 | |
| 304 | # Token T can only be routed to expert E if its priority is positive and |
| 305 | # less than the expert capacity. One-hot matrix will ignore indices outside |
| 306 | # the range [0, expert_capacity). |
| 307 | # Shape: [tokens_per_group, num_experts, expert_capacity]. |
| 308 | valid_mask = torch.logical_and(token_priority >= 0, token_priority < expert_capacity) |
| 309 | token_priority = torch.masked_fill(token_priority, ~valid_mask, 0) |
| 310 | dispatch_mask = F.one_hot(token_priority, expert_capacity).to(torch.bool) |
| 311 | valid_mask = valid_mask.unsqueeze(-1).expand(-1, -1, expert_capacity) |
| 312 | dispatch_mask = torch.masked_fill(dispatch_mask, ~valid_mask, 0) |
| 313 | |
| 314 | # The combine array will be used for combining expert outputs, scaled by the |
| 315 | # router probabilities. Shape: [num_groups, tokens_per_group, num_experts, |
| 316 | # expert_capacity]. |
| 317 | combine_weights = torch.einsum("...te,...tec->...tec", router_probs, dispatch_mask) |
| 318 | exp_counts_capacity = torch.sum(dispatch_mask) |
| 319 | exp_capacity_rate = exp_counts_capacity / (logits.shape[0] * topk) |
| 320 | |
| 321 | return [l_aux, exp_capacity_rate], combine_weights, dispatch_mask, exp_counts |
| 322 | |
| 323 | |
| 324 | # ======================================================= |
| 325 | # Multi-Dimensional RoPE |
| 326 | # ======================================================= |
| 327 | |
| 328 | def _to_tuple(x, dim=2): |
| 329 | if isinstance(x, int): |
| 330 | return (x,) * dim |
| 331 | elif len(x) == dim: |
| 332 | return x |
| 333 | else: |
| 334 | raise ValueError(f"Expected length {dim} or int, but got {x}") |
| 335 | |
| 336 | |
| 337 | def get_meshgrid_nd(start, *args, dim=2): |
| 338 | """ |
| 339 | Get n-D meshgrid with start, stop and num. |
| 340 | |
| 341 | Args: |
| 342 | start (int or tuple): If len(args) == 0, start is num; If len(args) == 1, start is start, args[0] is stop, |
| 343 | step is 1; If len(args) == 2, start is start, args[0] is stop, args[1] is num. For n-dim, start/stop/num |
| 344 | should be int or n-tuple. If n-tuple is provided, the meshgrid will be stacked following the dim order in |
| 345 | n-tuples. |
| 346 | *args: See above. |
| 347 | dim (int): Dimension of the meshgrid. Defaults to 2. |
| 348 | |
| 349 | Returns: |
| 350 | grid (np.ndarray): [dim, ...] |
| 351 | """ |
| 352 | if len(args) == 0: |
| 353 | # start is grid_size |
| 354 | num = _to_tuple(start, dim=dim) |
| 355 | start = (0,) * dim |
| 356 | stop = num |
| 357 | elif len(args) == 1: |
| 358 | # start is start, args[0] is stop, step is 1 |
| 359 | start = _to_tuple(start, dim=dim) |
| 360 | stop = _to_tuple(args[0], dim=dim) |
| 361 | num = [stop[i] - start[i] for i in range(dim)] |
| 362 | # assert num are all integers |
| 363 | num_int = [int(x) for x in num] |
| 364 | assert (torch.tensor(num) == torch.tensor(num_int)).all(), f"num should be int, but got {num}" |
| 365 | num = num_int |
| 366 | elif len(args) == 2: |
| 367 | # start is start, args[0] is stop, args[1] is num |
| 368 | start = _to_tuple(start, dim=dim) # Left-Top eg: 12,0 |
| 369 | stop = _to_tuple(args[0], dim=dim) # Right-Bottom eg: 20,32 |
| 370 | num = _to_tuple(args[1], dim=dim) # Target Size eg: 32,124 |
| 371 | else: |
| 372 | raise ValueError(f"len(args) should be 0, 1 or 2, but got {len(args)}") |
| 373 | |
| 374 | # PyTorch implement of np.linspace(start[i], stop[i], num[i], endpoint=False) |
| 375 | axis_grid = [] |
| 376 | for i in range(dim): |
| 377 | a, b, n = start[i], stop[i], num[i] |
| 378 | g = torch.linspace(a, b, n + 1, dtype=torch.float32)[:n] |
| 379 | axis_grid.append(g) |
| 380 | grid = torch.meshgrid(*axis_grid, indexing="ij") # dim x [H, W] |
| 381 | grid = torch.stack(grid, dim=0) # [dim, H, W] |
| 382 | |
| 383 | return grid |
| 384 | |
| 385 | |
| 386 | def build_2d_rope( |
| 387 | seq_len: int, n_elem: int, image_infos: Optional[List[Tuple[slice, Tuple[int, int]]]] = None, |
| 388 | device: Optional[torch.device] = None, base: int = 10000, base_rescale_factor: float = 1.0, |
| 389 | return_all_pos: bool = False, |
| 390 | ): |
| 391 | """ |
| 392 | Reference: https://kexue.fm/archives/10352 |
| 393 | |
| 394 | Start from 1, we have |
| 395 | beta_y = L + (wh - h)/2 |
| 396 | beta_x = L + (wh - w)/2 |
| 397 | |
| 398 | Returns |
| 399 | ------- |
| 400 | cos: torch.Tensor with shape of [seq_len, n_elem] |
| 401 | sin: torch.Tensor with shape of [seq_len, n_elem] |
| 402 | """ |
| 403 | assert n_elem % 4 == 0, f"n_elem must be divisible by 4, but got {n_elem}." |
| 404 | |
| 405 | # theta |
| 406 | if base_rescale_factor != 1.0: |
| 407 | base *= base_rescale_factor ** (n_elem / (n_elem - 2)) |
| 408 | theta = 1.0 / (base ** (torch.arange(0, n_elem, 2, device=device).float() / n_elem)) |
| 409 | theta = theta.reshape(1, n_elem // 4, 2) # [1, half_d, 2] |
| 410 | |
| 411 | # position indices |
| 412 | if image_infos is None: |
| 413 | image_infos = [] |
| 414 | |
| 415 | image_infos_list = [image_infos] |
| 416 | sample_seq_lens = [seq_len] |
| 417 | |
| 418 | # Prepare position indices for each sample |
| 419 | x_sections = [] |
| 420 | y_sections = [] |
| 421 | for sample_id, sample_image_infos in enumerate(image_infos_list): |
| 422 | last_pos = 0 |
| 423 | for sec_slice, (h, w) in sample_image_infos: |
| 424 | L = sec_slice.start # start from 0, so image_slice.start is just L |
| 425 | # previous text |
| 426 | if last_pos < L: |
| 427 | y_sections.append(torch.arange(last_pos, L)) |
| 428 | x_sections.append(torch.arange(last_pos, L)) |
| 429 | elif h is None: |
| 430 | # Interleave data has overlapped positions for <boi> <size> <ratio> <timestep> <eoi> tokens. |
| 431 | y_sections.append(torch.arange(sec_slice.start, sec_slice.stop)) |
| 432 | x_sections.append(torch.arange(sec_slice.start, sec_slice.stop)) |
| 433 | continue |
| 434 | else: |
| 435 | # Interleave data has overlapped positions for noised image and the successive clean image, |
| 436 | # leading to last_pos (= last text end L + noise w * h) > L (last text end L). |
| 437 | pass |
| 438 | # current image |
| 439 | beta_y = L + (w * h - h) / 2 |
| 440 | beta_x = L + (w * h - w) / 2 |
| 441 | grid = get_meshgrid_nd((beta_y, beta_x), (beta_y + h, beta_x + w)) # [2, h, w] |
| 442 | grid = grid.reshape(2, -1) # (y, x) |
| 443 | y_sections.append(grid[0]) |
| 444 | x_sections.append(grid[1]) |
| 445 | # step |
| 446 | last_pos = L + w * h |
| 447 | # final text |
| 448 | y_sections.append(torch.arange(last_pos, sample_seq_lens[sample_id])) |
| 449 | x_sections.append(torch.arange(last_pos, sample_seq_lens[sample_id])) |
| 450 | |
| 451 | x_pos = torch.cat(x_sections).long() |
| 452 | y_pos = torch.cat(y_sections).long() |
| 453 | # If there are overlap positions, we need to remove them. |
| 454 | x_pos = x_pos[:seq_len] |
| 455 | y_pos = y_pos[:seq_len] |
| 456 | all_pos = torch.stack((y_pos, x_pos), dim=1).unsqueeze(1).to(device) # [seq_len, 1, 2] |
| 457 | |
| 458 | # calc rope |
| 459 | idx_theta = (all_pos * theta).reshape(all_pos.shape[0], n_elem // 2).repeat(1, 2) |
| 460 | |
| 461 | cos = torch.cos(idx_theta) |
| 462 | sin = torch.sin(idx_theta) |
| 463 | |
| 464 | if return_all_pos: |
| 465 | return cos, sin, all_pos |
| 466 | |
| 467 | return cos, sin |
| 468 | |
| 469 | |
| 470 | def build_batch_2d_rope( |
| 471 | seq_len: int, n_elem: int, image_infos: Optional[List[List[Tuple[slice, Tuple[int, int]]]]] = None, |
| 472 | device: Optional[torch.device] = None, base: int = 10000, base_rescale_factor: float = 1.0, |
| 473 | return_all_pos: bool = False, |
| 474 | ): |
| 475 | cos_list, sin_list, all_pos_list = [], [], [] |
| 476 | if image_infos is None: |
| 477 | image_infos = [None] |
| 478 | for i, image_info in enumerate(image_infos): |
| 479 | res = build_2d_rope( |
| 480 | seq_len, n_elem, image_infos=image_info, device=device, |
| 481 | base=base, base_rescale_factor=base_rescale_factor, |
| 482 | return_all_pos=return_all_pos, |
| 483 | ) |
| 484 | if return_all_pos: |
| 485 | cos, sin, all_pos = res |
| 486 | else: |
| 487 | cos, sin = res |
| 488 | all_pos = None |
| 489 | cos_list.append(cos) |
| 490 | sin_list.append(sin) |
| 491 | all_pos_list.append(all_pos) |
| 492 | |
| 493 | stacked_cos = torch.stack(cos_list, dim=0) |
| 494 | stacked_sin = torch.stack(sin_list, dim=0) |
| 495 | |
| 496 | if return_all_pos: |
| 497 | return stacked_cos, stacked_sin, all_pos_list |
| 498 | |
| 499 | return stacked_cos, stacked_sin |
| 500 | |
| 501 | |
| 502 | def rotate_half(x): |
| 503 | """Rotates half the hidden dims of the input.""" |
| 504 | x1 = x[..., : x.shape[-1] // 2] |
| 505 | x2 = x[..., x.shape[-1] // 2:] |
| 506 | return torch.cat((-x2, x1), dim=-1) |
| 507 | |
| 508 | |
| 509 | def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1): |
| 510 | """Applies Rotary Position Embedding to the query and key tensors. |
| 511 | |
| 512 | Args: |
| 513 | q (`torch.Tensor`): The query tensor. |
| 514 | k (`torch.Tensor`): The key tensor. |
| 515 | cos (`torch.Tensor`): The cosine part of the rotary embedding. |
| 516 | sin (`torch.Tensor`): The sine part of the rotary embedding. |
| 517 | position_ids (`torch.Tensor`): |
| 518 | The position indices of the tokens corresponding to the query and key tensors. For example, this can be |
| 519 | used to pass shifted position ids when working with a KV-cache. |
| 520 | unsqueeze_dim (`int`, *optional*, defaults to 1): |
| 521 | The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and |
| 522 | sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note |
| 523 | that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and |
| 524 | k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes |
| 525 | cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have |
| 526 | the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. |
| 527 | Returns: |
| 528 | `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. |
| 529 | """ |
| 530 | if position_ids is not None: |
| 531 | cos = cos[position_ids] |
| 532 | sin = sin[position_ids] |
| 533 | |
| 534 | cos = cos.unsqueeze(unsqueeze_dim) |
| 535 | sin = sin.unsqueeze(unsqueeze_dim) |
| 536 | |
| 537 | q_embed = (q * cos) + (rotate_half(q) * sin) |
| 538 | k_embed = (k * cos) + (rotate_half(k) * sin) |
| 539 | return q_embed, k_embed |
| 540 | |
| 541 | |
| 542 | # ======================================================= |
| 543 | # Modules for Image Generation |
| 544 | # ======================================================= |
| 545 | |
| 546 | class TimestepEmbedder(nn.Module): |
| 547 | """ |
| 548 | Embeds scalar timesteps into vector representations. |
| 549 | """ |
| 550 | def __init__(self, |
| 551 | hidden_size, |
| 552 | act_layer=nn.GELU, |
| 553 | frequency_embedding_size=256, |
| 554 | max_period=10000, |
| 555 | out_size=None, |
| 556 | dtype=None, |
| 557 | device=None |
| 558 | ): |
| 559 | factory_kwargs = {'dtype': dtype, 'device': device} |
| 560 | super().__init__() |
| 561 | self.frequency_embedding_size = frequency_embedding_size |
| 562 | self.max_period = max_period |
| 563 | if out_size is None: |
| 564 | out_size = hidden_size |
| 565 | |
| 566 | self.mlp = nn.Sequential( |
| 567 | nn.Linear(frequency_embedding_size, hidden_size, bias=True, **factory_kwargs), |
| 568 | act_layer(), |
| 569 | nn.Linear(hidden_size, out_size, bias=True, **factory_kwargs), |
| 570 | ) |
| 571 | nn.init.normal_(self.mlp[0].weight, std=0.02) |
| 572 | nn.init.normal_(self.mlp[2].weight, std=0.02) |
| 573 | |
| 574 | def forward(self, t): |
| 575 | t_freq = timestep_embedding(t, self.frequency_embedding_size, self.max_period).type(self.mlp[0].weight.dtype) |
| 576 | t_emb = self.mlp(t_freq) |
| 577 | return t_emb |
| 578 | |
| 579 | |
| 580 | class Upsample(nn.Module): |
| 581 | """ |
| 582 | An upsampling layer with an optional convolution. |
| 583 | |
| 584 | :param channels: channels in the inputs and outputs. |
| 585 | :param use_conv: a bool determining if a convolution is applied. |
| 586 | :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then |
| 587 | upsampling occurs in the inner-two dimensions. |
| 588 | """ |
| 589 | |
| 590 | def __init__(self, channels, use_conv, dims=2, out_channels=None, device=None, dtype=None): |
| 591 | factory_kwargs = {'device': device, 'dtype': dtype} |
| 592 | super().__init__() |
| 593 | self.channels = channels |
| 594 | self.out_channels = out_channels or channels |
| 595 | self.use_conv = use_conv |
| 596 | self.dims = dims |
| 597 | if use_conv: |
| 598 | self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=1, **factory_kwargs) |
| 599 | |
| 600 | def forward(self, x): |
| 601 | assert x.shape[1] == self.channels |
| 602 | if self.dims == 3: |
| 603 | x = F.interpolate( |
| 604 | x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest" |
| 605 | ) |
| 606 | else: |
| 607 | x = F.interpolate(x, scale_factor=2, mode="nearest") |
| 608 | if self.use_conv: |
| 609 | x = self.conv(x) |
| 610 | return x |
| 611 | |
| 612 | |
| 613 | class Downsample(nn.Module): |
| 614 | """ |
| 615 | A downsampling layer with an optional convolution. |
| 616 | |
| 617 | :param channels: channels in the inputs and outputs. |
| 618 | :param use_conv: a bool determining if a convolution is applied. |
| 619 | :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then |
| 620 | downsampling occurs in the inner-two dimensions. |
| 621 | """ |
| 622 | |
| 623 | def __init__(self, channels, use_conv, dims=2, out_channels=None, device=None, dtype=None): |
| 624 | factory_kwargs = {'device': device, 'dtype': dtype} |
| 625 | super().__init__() |
| 626 | self.channels = channels |
| 627 | self.out_channels = out_channels or channels |
| 628 | self.use_conv = use_conv |
| 629 | self.dims = dims |
| 630 | stride = 2 if dims != 3 else (1, 2, 2) |
| 631 | if use_conv: |
| 632 | self.op = conv_nd( |
| 633 | dims, self.channels, self.out_channels, 3, stride=stride, padding=1, **factory_kwargs |
| 634 | ) |
| 635 | else: |
| 636 | assert self.channels == self.out_channels |
| 637 | self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride) |
| 638 | |
| 639 | def forward(self, x): |
| 640 | assert x.shape[1] == self.channels |
| 641 | return self.op(x) |
| 642 | |
| 643 | |
| 644 | class ResBlock(nn.Module): |
| 645 | """ |
| 646 | A residual block that can optionally change the number of channels. |
| 647 | |
| 648 | :param in_channels: the number of input channels. |
| 649 | :param emb_channels: the number of timestep embedding channels. |
| 650 | :param dropout: the rate of dropout. |
| 651 | :param out_channels: if specified, the number of out channels. |
| 652 | :param use_conv: if True and out_channels is specified, use a spatial |
| 653 | convolution instead of a smaller 1x1 convolution to change the |
| 654 | channels in the skip connection. |
| 655 | :param dims: determines if the signal is 1D, 2D, or 3D. |
| 656 | :param up: if True, use this block for upsampling. |
| 657 | :param down: if True, use this block for downsampling. |
| 658 | """ |
| 659 | |
| 660 | def __init__( |
| 661 | self, |
| 662 | in_channels, |
| 663 | emb_channels, |
| 664 | out_channels=None, |
| 665 | dropout=0.0, |
| 666 | use_conv=False, |
| 667 | dims=2, |
| 668 | up=False, |
| 669 | down=False, |
| 670 | device=None, |
| 671 | dtype=None, |
| 672 | ): |
| 673 | factory_kwargs = {'dtype': dtype, 'device': device} |
| 674 | super().__init__() |
| 675 | self.in_channels = in_channels |
| 676 | self.dropout = dropout |
| 677 | self.out_channels = out_channels or self.in_channels |
| 678 | self.use_conv = use_conv |
| 679 | |
| 680 | self.in_layers = nn.Sequential( |
| 681 | normalization(self.in_channels, **factory_kwargs), |
| 682 | nn.SiLU(), |
| 683 | conv_nd(dims, self.in_channels, self.out_channels, 3, padding=1, **factory_kwargs), |
| 684 | ) |
| 685 | |
| 686 | self.updown = up or down |
| 687 | |
| 688 | if up: |
| 689 | self.h_upd = Upsample(self.in_channels, False, dims, **factory_kwargs) |
| 690 | self.x_upd = Upsample(self.in_channels, False, dims, **factory_kwargs) |
| 691 | elif down: |
| 692 | self.h_upd = Downsample(self.in_channels, False, dims, **factory_kwargs) |
| 693 | self.x_upd = Downsample(self.in_channels, False, dims, **factory_kwargs) |
| 694 | else: |
| 695 | self.h_upd = self.x_upd = nn.Identity() |
| 696 | |
| 697 | self.emb_layers = nn.Sequential( |
| 698 | nn.SiLU(), |
| 699 | linear(emb_channels, 2 * self.out_channels, **factory_kwargs) |
| 700 | ) |
| 701 | |
| 702 | self.out_layers = nn.Sequential( |
| 703 | normalization(self.out_channels, **factory_kwargs), |
| 704 | nn.SiLU(), |
| 705 | nn.Dropout(p=dropout), |
| 706 | zero_module( |
| 707 | conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1, **factory_kwargs) |
| 708 | ), |
| 709 | ) |
| 710 | |
| 711 | if self.out_channels == self.in_channels: |
| 712 | self.skip_connection = nn.Identity() |
| 713 | elif use_conv: |
| 714 | self.skip_connection = conv_nd( |
| 715 | dims, self.in_channels, self.out_channels, 3, padding=1, **factory_kwargs |
| 716 | ) |
| 717 | else: |
| 718 | self.skip_connection = conv_nd(dims, self.in_channels, self.out_channels, 1, **factory_kwargs) |
| 719 | |
| 720 | def forward(self, x, emb): |
| 721 | if self.updown: |
| 722 | in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1] |
| 723 | h = in_rest(x) |
| 724 | h = self.h_upd(h) |
| 725 | x = self.x_upd(x) |
| 726 | h = in_conv(h) |
| 727 | else: |
| 728 | h = self.in_layers(x) |
| 729 | |
| 730 | emb_out = self.emb_layers(emb) |
| 731 | while len(emb_out.shape) < len(h.shape): |
| 732 | emb_out = emb_out[..., None] |
| 733 | |
| 734 | # Adaptive Group Normalization |
| 735 | out_norm, out_rest = self.out_layers[0], self.out_layers[1:] |
| 736 | scale, shift = torch.chunk(emb_out, 2, dim=1) |
| 737 | h = out_norm(h) * (1. + scale) + shift |
| 738 | h = out_rest(h) |
| 739 | |
| 740 | return self.skip_connection(x) + h |
| 741 | |
| 742 | |
| 743 | class UNetDown(nn.Module): |
| 744 | """ |
| 745 | patch_size: one of [1, 2 ,4 ,8] |
| 746 | in_channels: vae latent dim |
| 747 | hidden_channels: hidden dim for reducing parameters |
| 748 | out_channels: transformer model dim |
| 749 | """ |
| 750 | def __init__(self, patch_size, in_channels, emb_channels, hidden_channels, out_channels, |
| 751 | dropout=0.0, device=None, dtype=None): |
| 752 | factory_kwargs = {'dtype': dtype, 'device': device} |
| 753 | super().__init__() |
| 754 | |
| 755 | self.patch_size = patch_size |
| 756 | assert self.patch_size in [1, 2, 4, 8] |
| 757 | |
| 758 | self.model = nn.ModuleList( |
| 759 | [conv_nd( |
| 760 | 2, |
| 761 | in_channels=in_channels, |
| 762 | out_channels=hidden_channels, |
| 763 | kernel_size=3, |
| 764 | padding=1, |
| 765 | **factory_kwargs |
| 766 | )] |
| 767 | ) |
| 768 | |
| 769 | if self.patch_size == 1: |
| 770 | self.model.append(ResBlock( |
| 771 | in_channels=hidden_channels, |
| 772 | emb_channels=emb_channels, |
| 773 | out_channels=out_channels, |
| 774 | dropout=dropout, |
| 775 | **factory_kwargs |
| 776 | )) |
| 777 | else: |
| 778 | for i in range(self.patch_size // 2): |
| 779 | self.model.append(ResBlock( |
| 780 | in_channels=hidden_channels, |
| 781 | emb_channels=emb_channels, |
| 782 | out_channels=hidden_channels if (i + 1) * 2 != self.patch_size else out_channels, |
| 783 | dropout=dropout, |
| 784 | down=True, |
| 785 | **factory_kwargs |
| 786 | )) |
| 787 | |
| 788 | def forward(self, x, t): |
| 789 | assert x.shape[2] % self.patch_size == 0 and x.shape[3] % self.patch_size == 0 |
| 790 | for module in self.model: |
| 791 | if isinstance(module, ResBlock): |
| 792 | x = module(x, t) |
| 793 | else: |
| 794 | x = module(x) |
| 795 | _, _, token_h, token_w = x.shape |
| 796 | x = rearrange(x, 'b c h w -> b (h w) c') |
| 797 | return x, token_h, token_w |
| 798 | |
| 799 | |
| 800 | class UNetUp(nn.Module): |
| 801 | """ |
| 802 | patch_size: one of [1, 2 ,4 ,8] |
| 803 | in_channels: transformer model dim |
| 804 | hidden_channels: hidden dim for reducing parameters |
| 805 | out_channels: vae latent dim |
| 806 | """ |
| 807 | def __init__(self, patch_size, in_channels, emb_channels, hidden_channels, out_channels, |
| 808 | dropout=0.0, device=None, dtype=None, out_norm=False): |
| 809 | factory_kwargs = {'dtype': dtype, 'device': device} |
| 810 | super().__init__() |
| 811 | |
| 812 | self.patch_size = patch_size |
| 813 | assert self.patch_size in [1, 2, 4, 8] |
| 814 | |
| 815 | self.model = nn.ModuleList() |
| 816 | |
| 817 | if self.patch_size == 1: |
| 818 | self.model.append(ResBlock( |
| 819 | in_channels=in_channels, |
| 820 | emb_channels=emb_channels, |
| 821 | out_channels=hidden_channels, |
| 822 | dropout=dropout, |
| 823 | **factory_kwargs |
| 824 | )) |
| 825 | else: |
| 826 | for i in range(self.patch_size // 2): |
| 827 | self.model.append(ResBlock( |
| 828 | in_channels=in_channels if i == 0 else hidden_channels, |
| 829 | emb_channels=emb_channels, |
| 830 | out_channels=hidden_channels, |
| 831 | dropout=dropout, |
| 832 | up=True, |
| 833 | **factory_kwargs |
| 834 | )) |
| 835 | |
| 836 | if out_norm: |
| 837 | self.model.append(nn.Sequential( |
| 838 | normalization(hidden_channels, **factory_kwargs), |
| 839 | nn.SiLU(), |
| 840 | conv_nd( |
| 841 | 2, |
| 842 | in_channels=hidden_channels, |
| 843 | out_channels=out_channels, |
| 844 | kernel_size=3, |
| 845 | padding=1, |
| 846 | **factory_kwargs |
| 847 | ), |
| 848 | )) |
| 849 | else: |
| 850 | self.model.append(conv_nd( |
| 851 | 2, |
| 852 | in_channels=hidden_channels, |
| 853 | out_channels=out_channels, |
| 854 | kernel_size=3, |
| 855 | padding=1, |
| 856 | **factory_kwargs |
| 857 | )) |
| 858 | |
| 859 | # batch_size, seq_len, model_dim |
| 860 | def forward(self, x, t, token_h, token_w): |
| 861 | x = rearrange(x, 'b (h w) c -> b c h w', h=token_h, w=token_w) |
| 862 | for module in self.model: |
| 863 | if isinstance(module, ResBlock): |
| 864 | x = module(x, t) |
| 865 | else: |
| 866 | x = module(x) |
| 867 | return x |
| 868 | |
| 869 | |
| 870 | # ======================================================= |
| 871 | # Modules for Transformer Backbone |
| 872 | # ======================================================= |
| 873 | |
| 874 | @dataclass |
| 875 | class CausalMMOutputWithPast(CausalLMOutputWithPast): |
| 876 | diffusion_prediction: Optional[torch.Tensor] = None |
| 877 | |
| 878 | |
| 879 | class HunyuanStaticCache(StaticCache): |
| 880 | """ |
| 881 | A custom static cache for multi-modal models that supports dynamic extension of the cache |
| 882 | and inplace updates of the cache. |
| 883 | |
| 884 | This cache supports batch cache_position updates. |
| 885 | """ |
| 886 | def __init__(self, *args, **kwargs): |
| 887 | self.dynamic = kwargs.pop("dynamic", False) |
| 888 | super().__init__(*args, **kwargs) |
| 889 | |
| 890 | def update( |
| 891 | self, |
| 892 | key_states: torch.Tensor, |
| 893 | value_states: torch.Tensor, |
| 894 | layer_idx: int, |
| 895 | cache_kwargs: Optional[Dict[str, Any]] = None, |
| 896 | ) -> Tuple[torch.Tensor, torch.Tensor]: |
| 897 | """ |
| 898 | Updates the cache with the new `key_states` and `value_states` for the layer `layer_idx`. |
| 899 | It is VERY important to index using a tensor, otherwise you introduce a copy to the device. |
| 900 | |
| 901 | Parameters: |
| 902 | key_states (`torch.Tensor`): |
| 903 | The new key states to cache. |
| 904 | value_states (`torch.Tensor`): |
| 905 | The new value states to cache. |
| 906 | layer_idx (`int`): |
| 907 | The index of the layer to cache the states for. |
| 908 | cache_kwargs (`Dict[str, Any]`, `optional`): |
| 909 | Additional arguments for the cache subclass. The `StaticCache` needs the `cache_position` input |
| 910 | to know how where to write in the cache. |
| 911 | |
| 912 | Return: |
| 913 | A tuple containing the updated key and value states. |
| 914 | """ |
| 915 | cache_position = cache_kwargs.get("cache_position") |
| 916 | if hasattr(self, "key_cache") and hasattr(self, "value_cache"): |
| 917 | if self.key_cache[layer_idx].device != key_states.device: |
| 918 | self.key_cache[layer_idx] = self.key_cache[layer_idx].to(key_states.device) |
| 919 | self.value_cache[layer_idx] = self.value_cache[layer_idx].to(value_states.device) |
| 920 | k_out = self.key_cache[layer_idx] |
| 921 | v_out = self.value_cache[layer_idx] |
| 922 | key_states = key_states.to(k_out.dtype) |
| 923 | value_states = value_states.to(v_out.dtype) |
| 924 | else: |
| 925 | if self.layers[layer_idx].keys is None: |
| 926 | self.layers[layer_idx].lazy_initialization(key_states) |
| 927 | k_out = self.layers[layer_idx].keys |
| 928 | v_out = self.layers[layer_idx].values |
| 929 | |
| 930 | if cache_position is None: |
| 931 | k_out.copy_(key_states) |
| 932 | v_out.copy_(value_states) |
| 933 | else: |
| 934 | # Note: here we use `tensor.index_copy_(dim, index, tensor)` that is equivalent to |
| 935 | # `tensor[:, :, index] = tensor`, but the first one is compile-friendly and it does explicitly an in-place |
| 936 | # operation, that avoids copies and uses less memory. |
| 937 | if cache_position.dim() == 1: |
| 938 | k_out.index_copy_(2, cache_position, key_states) |
| 939 | v_out.index_copy_(2, cache_position, value_states) |
| 940 | |
| 941 | if self.dynamic: |
| 942 | end = cache_position[-1].item() + 1 |
| 943 | k_out = k_out[:, :, :end] |
| 944 | v_out = v_out[:, :, :end] |
| 945 | else: |
| 946 | assert cache_position.dim() == 2, f"multiple batch dims not yet {cache_position.shape=}" |
| 947 | batch_size, idx_size = cache_position.shape |
| 948 | assert batch_size == k_out.size(0) |
| 949 | assert batch_size == v_out.size(0) |
| 950 | assert batch_size == key_states.size(0) |
| 951 | assert batch_size == value_states.size(0) |
| 952 | for i in range(batch_size): |
| 953 | unbatched_dim = 1 |
| 954 | k_out[i].index_copy_(unbatched_dim, cache_position[i], key_states[i]) |
| 955 | v_out[i].index_copy_(unbatched_dim, cache_position[i], value_states[i]) |
| 956 | |
| 957 | if self.dynamic: |
| 958 | assert len(cache_position) == 1 |
| 959 | end = cache_position[0, -1].item() + 1 |
| 960 | k_out = k_out[:, :, :end] |
| 961 | v_out = v_out[:, :, :end] |
| 962 | |
| 963 | return k_out, v_out |
| 964 | |
| 965 | |
| 966 | class HunyuanRMSNorm(nn.Module): |
| 967 | def __init__(self, hidden_size, eps=1e-6): |
| 968 | """ |
| 969 | HunyuanRMSNorm is equivalent to T5LayerNorm |
| 970 | """ |
| 971 | super().__init__() |
| 972 | self.weight = nn.Parameter(torch.ones(hidden_size)) |
| 973 | self.variance_epsilon = eps |
| 974 | |
| 975 | def forward(self, hidden_states): |
| 976 | input_dtype = hidden_states.dtype |
| 977 | hidden_states = hidden_states.to(torch.float32) |
| 978 | variance = hidden_states.pow(2).mean(-1, keepdim=True) |
| 979 | hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) |
| 980 | return self.weight * hidden_states.to(input_dtype) |
| 981 | |
| 982 | |
| 983 | class HunyuanMLP(nn.Module): |
| 984 | def __init__(self, config: HunyuanImage3Config, layer_idx=None, is_shared_mlp=False, is_moe=False): |
| 985 | super().__init__() |
| 986 | self.config = config |
| 987 | self.layer_idx = layer_idx |
| 988 | self.hidden_size = config.hidden_size |
| 989 | self.hidden_act = config.hidden_act |
| 990 | |
| 991 | self.intermediate_size = config.intermediate_size |
| 992 | if is_shared_mlp or is_moe: |
| 993 | # 如果是 moe 的话，优先用 moe_intermediate_size |
| 994 | if config.moe_intermediate_size is not None: |
| 995 | self.intermediate_size = config.moe_intermediate_size \ |
| 996 | if isinstance(config.moe_intermediate_size, int) else config.moe_intermediate_size[layer_idx] |
| 997 | |
| 998 | if is_shared_mlp: |
| 999 | num_shared_expert = config.num_shared_expert \ |
| 1000 | if isinstance(config.num_shared_expert, int) else config.num_shared_expert[layer_idx] |
| 1001 | self.intermediate_size *= num_shared_expert |
| 1002 | |
| 1003 | self.act_fn = ACT2FN[config.hidden_act] |
| 1004 | if self.hidden_act == "silu": |
| 1005 | self.intermediate_size *= 2 # SwiGLU |
| 1006 | self.gate_and_up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias) |
| 1007 | self.down_proj = nn.Linear(self.intermediate_size // 2, self.hidden_size, bias=config.mlp_bias) |
| 1008 | elif self.hidden_act == "gelu": |
| 1009 | self.gate_and_up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias) |
| 1010 | self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias) |
| 1011 | else: |
| 1012 | assert False, "other hidden_act are not supported" |
| 1013 | |
| 1014 | def forward(self, x): |
| 1015 | if self.hidden_act == "silu": |
| 1016 | gate_and_up_proj = self.gate_and_up_proj(x) |
| 1017 | x1, x2 = gate_and_up_proj.chunk(2, dim=2) |
| 1018 | down_proj = self.down_proj(x1 * self.act_fn(x2)) |
| 1019 | return down_proj |
| 1020 | elif self.hidden_act == "gelu": |
| 1021 | intermediate = self.gate_and_up_proj(x) |
| 1022 | intermediate = self.act_fn(intermediate) |
| 1023 | output = self.down_proj(intermediate) |
| 1024 | return output |
| 1025 | else: |
| 1026 | assert False, "other hidden_act are not supported" |
| 1027 | |
| 1028 | |
| 1029 | class HunyuanTopKGate(nn.Module): |
| 1030 | def __init__(self, config: HunyuanImage3Config, layer_idx: Optional[int] = None): |
| 1031 | super().__init__() |
| 1032 | self.config = config |
| 1033 | self.layer_idx = layer_idx |
| 1034 | self.moe_topk = config.moe_topk if isinstance(config.moe_topk, int) else config.moe_topk[layer_idx] |
| 1035 | self.drop_tokens = config.moe_drop_tokens |
| 1036 | self.min_capacity = 8 |
| 1037 | self.random_routing_dropped_token = config.moe_random_routing_dropped_token |
| 1038 | num_experts = config.num_experts if isinstance(config.num_experts, int) else config.num_experts[layer_idx] |
| 1039 | self.wg = nn.Linear(config.hidden_size, num_experts, bias=False, dtype=torch.float32) |
| 1040 | |
| 1041 | # DeepSeek gating args |
| 1042 | self.routed_scaling_factor = config.routed_scaling_factor |
| 1043 | self.n_group = config.n_group |
| 1044 | self.topk_group = config.topk_group |
| 1045 | self.norm_topk_prob = config.norm_topk_prob |
| 1046 | self.group_limited_greedy = config.group_limited_greedy |
| 1047 | |
| 1048 | def forward(self, hidden_states, topk_impl='default'): |
| 1049 | bsz, seq_len, hidden_size = hidden_states.shape |
| 1050 | hidden_states = hidden_states.reshape(-1, hidden_size) |
| 1051 | if self.wg.weight.dtype == torch.float32: |
| 1052 | hidden_states = hidden_states.float() |
| 1053 | logits = self.wg(hidden_states) |
| 1054 | if topk_impl == 'default': |
| 1055 | gate_output = topkgating(logits, self.moe_topk, group_limited_greedy=self.group_limited_greedy, |
| 1056 | n_group=self.n_group, topk_group=self.topk_group, |
| 1057 | norm_topk_prob=self.norm_topk_prob, |
| 1058 | routed_scaling_factor=self.routed_scaling_factor, |
| 1059 | capacity_factor=self.config.capacity_factor, |
| 1060 | drop_tokens=self.drop_tokens) |
| 1061 | elif topk_impl == 'easy': |
| 1062 | gate_output = self.easy_topk(logits, self.moe_topk) |
| 1063 | else: |
| 1064 | raise ValueError(f"Unsupported topk_impl: {topk_impl}") |
| 1065 | |
| 1066 | return gate_output |
| 1067 | |
| 1068 | @staticmethod |
| 1069 | def easy_topk(logits, moe_topk): |
| 1070 | gates = F.softmax(logits, dim=1) |
| 1071 | topk_weight_1, expert_index = torch.topk(gates, moe_topk) |
| 1072 | weight_sums = topk_weight_1.sum(dim=1, keepdim=True) |
| 1073 | weight_sums = torch.clamp(weight_sums, min=1e-8) |
| 1074 | topk_weight = topk_weight_1 / weight_sums |
| 1075 | |
| 1076 | return topk_weight, expert_index |
| 1077 | |
| 1078 | |
| 1079 | class HunyuanMoE(nn.Module): |
| 1080 | def __init__(self, config: HunyuanImage3Config, layer_idx: Optional[int] = None): |
| 1081 | super().__init__() |
| 1082 | self.config = config |
| 1083 | self.layer_idx = layer_idx |
| 1084 | self.moe_topk = config.moe_topk |
| 1085 | self.num_experts = config.num_experts if isinstance(config.num_experts, int) else config.num_experts[layer_idx] |
| 1086 | if config.use_mixed_mlp_moe: |
| 1087 | self.shared_mlp = HunyuanMLP(config, layer_idx=layer_idx, is_shared_mlp=True) |
| 1088 | self.gate = HunyuanTopKGate(config, layer_idx=layer_idx) |
| 1089 | self.experts = nn.ModuleList( |
| 1090 | [HunyuanMLP(config, layer_idx=layer_idx, is_shared_mlp=False, is_moe=True) for _ in range(self.num_experts)] |
| 1091 | ) |
| 1092 | |
| 1093 | self._moe_impl = config.moe_impl |
| 1094 | # For FlashInfer |
| 1095 | self.moe_weight = None |
| 1096 | self.moe_weight_2 = None |
| 1097 | self._weights_initialized = False |
| 1098 | |
| 1099 | @property |
| 1100 | def moe_impl(self): |
| 1101 | return self._moe_impl |
| 1102 | |
| 1103 | @moe_impl.setter |
| 1104 | def moe_impl(self, value): |
| 1105 | self._moe_impl = value |
| 1106 | if self._moe_impl == "flashinfer": |
| 1107 | assert flashinfer is not None, "When using fused_moe, flashinfer must be installed." |
| 1108 | |
| 1109 | def forward(self, hidden_states): |
| 1110 | torch.cuda.set_device(hidden_states.device.index) |
| 1111 | bsz, seq_len, hidden_size = hidden_states.shape |
| 1112 | |
| 1113 | if self.config.use_mixed_mlp_moe: |
| 1114 | hidden_states_mlp = self.shared_mlp(hidden_states) |
| 1115 | |
| 1116 | reshaped_input = hidden_states.reshape(-1, hidden_size) # [bsz*seq_len, hidden_size] |
| 1117 | |
| 1118 | with nvtx.range("MoE"): |
| 1119 | if self._moe_impl == "flashinfer": |
| 1120 | # Get expert weights |
| 1121 | if not self._weights_initialized: |
| 1122 | self._initialize_weights_on_device(hidden_states.device) |
| 1123 | topk_weight, topk_index = self.gate(hidden_states, topk_impl='easy') |
| 1124 | |
| 1125 | combined_output = torch.zeros_like(reshaped_input) |
| 1126 | _ = flashinfer.fused_moe.cutlass_fused_moe( # noqa |
| 1127 | reshaped_input.contiguous(), |
| 1128 | topk_index.to(torch.int).contiguous(), |
| 1129 | topk_weight.to(torch.float).contiguous(), |
| 1130 | self.moe_weight, |
| 1131 | self.moe_weight_2, |
| 1132 | torch.bfloat16, |
| 1133 | output=combined_output, |
| 1134 | quant_scales=None, |
| 1135 | ) |
| 1136 | else: |
| 1137 | # Original implementation - fallback for compatibility |
| 1138 | l_moe, combine_weights, dispatch_mask, exp_counts = self.gate(hidden_states, topk_impl='default') |
| 1139 | dispatched_input = torch.einsum("sec,sm->ecm", dispatch_mask.type_as(hidden_states), reshaped_input) |
| 1140 | chunks = dispatched_input.chunk(self.num_experts, dim=0) |
| 1141 | expert_outputs = [] |
| 1142 | for chunk, expert in zip(chunks, self.experts): |
| 1143 | expert_outputs.append(expert(chunk)) |
| 1144 | |
| 1145 | expert_output = torch.cat(expert_outputs, dim=0) |
| 1146 | combined_output = torch.einsum("sec,ecm->sm", combine_weights.type_as(hidden_states), expert_output) |
| 1147 | |
| 1148 | combined_output = combined_output.reshape(bsz, seq_len, hidden_size) |
| 1149 | |
| 1150 | if self.config.use_mixed_mlp_moe: |
| 1151 | output = hidden_states_mlp + combined_output # noqa |
| 1152 | else: |
| 1153 | output = combined_output |
| 1154 | |
| 1155 | return output |
| 1156 | |
| 1157 | def _initialize_weights_on_device(self, device): |
| 1158 | expert_weights_gate_up = [] |
| 1159 | expert_weights_down = [] |
| 1160 | |
| 1161 | for expert in self.experts: |
| 1162 | expert.to(device) |
| 1163 | expert_weights_gate_up.append(expert.gate_and_up_proj.weight.to(device)) |
| 1164 | expert_weights_down.append(expert.down_proj.weight.to(device)) |
| 1165 | |
| 1166 | self.moe_weight = torch.stack(expert_weights_gate_up).contiguous() |
| 1167 | self.moe_weight_2 = torch.stack(expert_weights_down).contiguous() |
| 1168 | # empty the expert weights |
| 1169 | for expert in self.experts: |
| 1170 | expert.gate_and_up_proj.weight.data = torch.empty(0, device=device) |
| 1171 | if expert.gate_and_up_proj.bias is not None: |
| 1172 | expert.gate_and_up_proj.bias.data = torch.empty(0, device=device) |
| 1173 | expert.down_proj.weight.data = torch.empty(0, device=device) |
| 1174 | if expert.down_proj.bias is not None: |
| 1175 | expert.down_proj.bias.data = torch.empty(0, device=device) |
| 1176 | |
| 1177 | self._weights_initialized = True |
| 1178 | |
| 1179 | |
| 1180 | class HunyuanImage3SDPAAttention(nn.Module): |
| 1181 | """PyTorch SDPA attention implementation using torch.nn.functional.scaled_dot_product_attention""" |
| 1182 | |
| 1183 | def __init__(self, config: HunyuanImage3Config, layer_idx: int): |
| 1184 | super().__init__() |
| 1185 | self.config = config |
| 1186 | self.layer_idx = layer_idx |
| 1187 | self.attention_type = 'self' |
| 1188 | |
| 1189 | self.attention_dropout = config.attention_dropout |
| 1190 | self.hidden_size = config.hidden_size |
| 1191 | self.num_heads = config.num_attention_heads |
| 1192 | # self.head_dim = self.hidden_size // self.num_heads |
| 1193 | self.head_dim = config.attention_head_dim |
| 1194 | self.num_key_value_heads = config.num_key_value_heads if config.num_key_value_heads else self.num_heads |
| 1195 | self.num_key_value_groups = self.num_heads // self.num_key_value_heads |
| 1196 | self.max_position_embeddings = config.max_position_embeddings |
| 1197 | self.rope_theta = config.rope_theta |
| 1198 | self.is_causal = True |
| 1199 | self.use_qk_norm = config.use_qk_norm |
| 1200 | self.use_rotary_pos_emb = config.use_rotary_pos_emb |
| 1201 | self.hidden_size_q = self.head_dim * self.num_heads |
| 1202 | self.hidden_size_kv = self.head_dim * self.num_key_value_heads |
| 1203 | |
| 1204 | # define layers |
| 1205 | self.qkv_proj = nn.Linear( |
| 1206 | self.hidden_size, |
| 1207 | self.hidden_size_q + 2 * self.hidden_size_kv, |
| 1208 | bias=config.attention_bias |
| 1209 | ) |
| 1210 | self.o_proj = nn.Linear(self.hidden_size_q, self.hidden_size, bias=config.attention_bias) |
| 1211 | |
| 1212 | if self.use_qk_norm: |
| 1213 | self.query_layernorm = HunyuanRMSNorm(self.head_dim, eps=config.rms_norm_eps) |
| 1214 | self.key_layernorm = HunyuanRMSNorm(self.head_dim, eps=config.rms_norm_eps) |
| 1215 | |
| 1216 | if self.use_rotary_pos_emb: |
| 1217 | self._init_rope() |
| 1218 | |
| 1219 | def _init_rope(self): |
| 1220 | scaling_type = self.config.rope_scaling["type"] |
| 1221 | if scaling_type == "custom": |
| 1222 | # Using custom rotary embedding |
| 1223 | self.rotary_emb = None |
| 1224 | else: |
| 1225 | raise ValueError(f"Unknown RoPE scaling type {scaling_type}") |
| 1226 | |
| 1227 | def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int): |
| 1228 | return tensor.reshape(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous() |
| 1229 | |
| 1230 | def forward( |
| 1231 | self, |
| 1232 | hidden_states: torch.Tensor, |
| 1233 | attention_mask: Optional[torch.Tensor] = None, |
| 1234 | position_ids: Optional[torch.LongTensor] = None, |
| 1235 | past_key_value: Optional[Cache] = None, |
| 1236 | output_attentions: bool = False, |
| 1237 | use_cache: Optional[bool] = False, |
| 1238 | custom_pos_emb: Optional[Tuple[torch.FloatTensor]] = None, |
| 1239 | **kwargs, |
| 1240 | ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]: |
| 1241 | if output_attentions: |
| 1242 | raise NotImplementedError( |
| 1243 | 'HunyuanImage3Model is using HunyuanImage3SDPAAttention,' |
| 1244 | 'but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`.' |
| 1245 | ) |
| 1246 | |
| 1247 | bsz, q_len, _ = hidden_states.size() |
| 1248 | |
| 1249 | qkv_states = self.qkv_proj(hidden_states) |
| 1250 | qkv_states = qkv_states.reshape(bsz, q_len, self.num_key_value_heads, self.num_key_value_groups + 2, |
| 1251 | self.head_dim) |
| 1252 | query_states, key_states, value_states = torch.split(qkv_states, [self.num_key_value_groups, 1, 1], dim=3) |
| 1253 | |
| 1254 | query_states = query_states.reshape(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) |
| 1255 | key_states = key_states.reshape(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) |
| 1256 | value_states = value_states.reshape(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) |
| 1257 | |
| 1258 | if self.use_rotary_pos_emb: |
| 1259 | cos, sin = custom_pos_emb |
| 1260 | query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) |
| 1261 | |
| 1262 | if self.use_qk_norm: |
| 1263 | query_states = self.query_layernorm(query_states) |
| 1264 | key_states = self.key_layernorm(key_states) |
| 1265 | |
| 1266 | query_states = query_states.to(value_states.dtype) |
| 1267 | key_states = key_states.to(value_states.dtype) |
| 1268 | |
| 1269 | if past_key_value is not None: |
| 1270 | cache_kwargs = {"cache_position": position_ids} |
| 1271 | key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) |
| 1272 | query_states = query_states.to(key_states.dtype) |
| 1273 | |
| 1274 | key_states = repeat_kv(key_states, self.num_key_value_groups) |
| 1275 | value_states = repeat_kv(value_states, self.num_key_value_groups) |
| 1276 | |
| 1277 | # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with |
| 1278 | # custom attn_mask, |
| 1279 | # Reference: https://github.com/pytorch/pytorch/issues/112577. |
| 1280 | if query_states.device.type == "cuda" and attention_mask is not None: |
| 1281 | query_states = query_states.contiguous() |
| 1282 | key_states = key_states.contiguous() |
| 1283 | value_states = value_states.contiguous() |
| 1284 | |
| 1285 | attn_output = torch.nn.functional.scaled_dot_product_attention( |
| 1286 | query_states, key_states, value_states, attn_mask=attention_mask, dropout_p=0.0 |
| 1287 | ) |
| 1288 | attn_output = attn_output.transpose(1, 2).contiguous() |
| 1289 | |
| 1290 | attn_output = attn_output.reshape(bsz, q_len, -1) |
| 1291 | |
| 1292 | attn_output = self.o_proj(attn_output) |
| 1293 | |
| 1294 | return attn_output, None, past_key_value |
| 1295 | |
| 1296 | |
| 1297 | class HunyuanImage3FlashAttention2(HunyuanImage3SDPAAttention): |
| 1298 | |
| 1299 | def forward( |
| 1300 | self, |
| 1301 | hidden_states: torch.Tensor, |
| 1302 | attention_mask: Optional[torch.Tensor] = None, |
| 1303 | position_ids: Optional[torch.LongTensor] = None, |
| 1304 | past_key_value: Optional[Cache] = None, |
| 1305 | output_attentions: bool = False, |
| 1306 | use_cache: Optional[bool] = False, |
| 1307 | custom_pos_emb: Optional[Tuple[torch.FloatTensor]] = None, |
| 1308 | **kwargs, |
| 1309 | ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]: |
| 1310 | if output_attentions: |
| 1311 | return super().forward( |
| 1312 | hidden_states=hidden_states, |
| 1313 | attention_mask=attention_mask, |
| 1314 | position_ids=position_ids, |
| 1315 | past_key_value=past_key_value, |
| 1316 | output_attentions=output_attentions, |
| 1317 | ) |
| 1318 | |
| 1319 | bsz, q_len, _ = hidden_states.size() |
| 1320 | |
| 1321 | qkv_states = self.qkv_proj(hidden_states) |
| 1322 | qkv_states = qkv_states.reshape(bsz, q_len, self.num_key_value_heads, self.num_key_value_groups + 2, |
| 1323 | self.head_dim) |
| 1324 | query_states, key_states, value_states = torch.split(qkv_states, [self.num_key_value_groups, 1, 1], dim=3) |
| 1325 | |
| 1326 | query_states = query_states.reshape(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) |
| 1327 | key_states = key_states.reshape(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) |
| 1328 | value_states = value_states.reshape(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) |
| 1329 | |
| 1330 | if self.use_rotary_pos_emb: |
| 1331 | cos, sin = custom_pos_emb |
| 1332 | query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) |
| 1333 | |
| 1334 | if self.use_qk_norm: |
| 1335 | query_states = self.query_layernorm(query_states) |
| 1336 | key_states = self.key_layernorm(key_states) |
| 1337 | |
| 1338 | query_states = query_states.to(value_states.dtype) |
| 1339 | key_states = key_states.to(value_states.dtype) |
| 1340 | |
| 1341 | if past_key_value is not None: |
| 1342 | cache_kwargs = {"cache_position": position_ids} |
| 1343 | key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) |
| 1344 | |
| 1345 | key_states = repeat_kv(key_states, self.num_key_value_groups) |
| 1346 | value_states = repeat_kv(value_states, self.num_key_value_groups) |
| 1347 | |
| 1348 | # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with |
| 1349 | # custom attn_mask, |
| 1350 | # Reference: https://github.com/pytorch/pytorch/issues/112577. |
| 1351 | if query_states.device.type == "cuda" and attention_mask is not None: |
| 1352 | query_states = query_states.contiguous() |
| 1353 | key_states = key_states.contiguous() |
| 1354 | value_states = value_states.contiguous() |
| 1355 | |
| 1356 | target_dtype = key_states.dtype if key_states.dtype in [torch.bfloat16, torch.float16] else torch.bfloat16 |
| 1357 | |
| 1358 | q_fa = query_states.to(target_dtype).transpose(1, 2).contiguous() |
| 1359 | k_fa = key_states.to(target_dtype).transpose(1, 2).contiguous() |
| 1360 | v_fa = value_states.to(target_dtype).transpose(1, 2).contiguous() |
| 1361 | |
| 1362 | mode = kwargs.get("mode", "gen_text") |
| 1363 | # For gen_text and gen_image, we need to handle the attention differently |
| 1364 | with nvtx.range("attention"): |
| 1365 | if mode == "gen_text": |
| 1366 | if attention_mask is None: |
| 1367 | attn_output = flash_attn_func(q_fa, k_fa, v_fa, causal=False) # decode attention |
| 1368 | else: |
| 1369 | attn_output = flash_attn_func(q_fa, k_fa, v_fa, causal=True) # prefill attention |
| 1370 | else: # image attention |
| 1371 | gen_timestep_scatter_index: Optional[torch.Tensor] = kwargs.get("gen_timestep_scatter_index", None) |
| 1372 | assert gen_timestep_scatter_index is not None, \ |
| 1373 | "When gen_image, `gen_timestep_scatter_index` must be provided." |
| 1374 | # TODO: batchify |
| 1375 | timestep_index = gen_timestep_scatter_index[0, 0].item() |
| 1376 | # When image generation, different attention implementations for the first step and the following steps |
| 1377 | # help to improve the inference speed. |
| 1378 | first_step = kwargs.get("first_step", None) |
| 1379 | if first_step is None: |
| 1380 | raise ValueError("When gen_image, `first_step` must be provided.") |
| 1381 | if first_step: |
| 1382 | casual_len = timestep_index + 1 |
| 1383 | text_query_states = q_fa[:, :casual_len, :, :] |
| 1384 | text_key_states = k_fa[:, :casual_len, :, :] |
| 1385 | text_value_states = v_fa[:, :casual_len, :, :] |
| 1386 | text_attn_output = flash_attn_func( |
| 1387 | text_query_states, text_key_states, text_value_states, causal=True) |
| 1388 | image_query_states = q_fa[:, casual_len:, :, :] |
| 1389 | image_attn_output = flash_attn_func(image_query_states, k_fa, v_fa, causal=False) |
| 1390 | attn_output = torch.cat((text_attn_output, image_attn_output), dim=1) |
| 1391 | else: |
| 1392 | casual_len = timestep_index + 1 |
| 1393 | timestep_query_states = q_fa[:, 0:1, :, :] |
| 1394 | timestep_key_states = k_fa[:, :casual_len, :, :] |
| 1395 | timestep_value_states = v_fa[:, :casual_len, :, :] |
| 1396 | timestep_attn_output = flash_attn_func( |
| 1397 | timestep_query_states, timestep_key_states, timestep_value_states, causal=True) |
| 1398 | image_query_states = q_fa[:, 1:, :, :] |
| 1399 | image_attn_output = flash_attn_func(image_query_states, k_fa, v_fa, causal=False) |
| 1400 | attn_output = torch.cat((timestep_attn_output, image_attn_output), dim=1) |
| 1401 | |
| 1402 | attn_output = attn_output.reshape(bsz, q_len, -1) |
| 1403 | |
| 1404 | attn_output = self.o_proj(attn_output) |
| 1405 | |
| 1406 | return attn_output, None, past_key_value |
| 1407 | |
| 1408 | |
| 1409 | Hunyuan_ATTENTION_CLASSES = { |
| 1410 | "eager": HunyuanImage3SDPAAttention, |
| 1411 | "sdpa": HunyuanImage3SDPAAttention, |
| 1412 | "flash_attention_2": HunyuanImage3FlashAttention2, |
| 1413 | } |
| 1414 | |
| 1415 | |
| 1416 | class HunyuanImage3DecoderLayer(nn.Module): |
| 1417 | def __init__(self, config: HunyuanImage3Config, layer_idx: int): |
| 1418 | super().__init__() |
| 1419 | self.hidden_size = config.hidden_size |
| 1420 | self.layer_idx = layer_idx |
| 1421 | |
| 1422 | attn_impl = config._attn_implementation # noqa |
| 1423 | if attn_impl in Hunyuan_ATTENTION_CLASSES: |
| 1424 | self.self_attn = Hunyuan_ATTENTION_CLASSES[attn_impl](config=config, layer_idx=layer_idx) |
| 1425 | else: |
| 1426 | raise ValueError(f"Unsupported attention implementation: {attn_impl}") |
| 1427 | |
| 1428 | if ((isinstance(config.num_experts, int) and config.num_experts > 1) or ( |
| 1429 | isinstance(config.num_experts, list) and max( |
| 1430 | config.num_experts) > 1)) and layer_idx >= config.moe_layer_num_skipped: |
| 1431 | self.mlp = HunyuanMoE(config, layer_idx=layer_idx) |
| 1432 | else: |
| 1433 | self.mlp = HunyuanMLP(config, layer_idx=layer_idx, is_shared_mlp=False, is_moe=False) |
| 1434 | if config.norm_type == 'hf_rms' or config.norm_type == 'rms': |
| 1435 | self.input_layernorm = HunyuanRMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 1436 | self.post_attention_layernorm = HunyuanRMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 1437 | elif config.norm_type == 'fused' or config.norm_type == 'torch_nn': |
| 1438 | self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 1439 | self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 1440 | else: |
| 1441 | assert False, "other norm_type are not supported" |
| 1442 | |
| 1443 | def forward( |
| 1444 | self, |
| 1445 | hidden_states: torch.Tensor, |
| 1446 | attention_mask: Optional[torch.Tensor] = None, |
| 1447 | position_ids: Optional[torch.LongTensor] = None, |
| 1448 | past_key_value: Optional[Tuple[torch.Tensor]] = None, |
| 1449 | output_attentions: Optional[bool] = False, |
| 1450 | use_cache: Optional[bool] = False, |
| 1451 | custom_pos_emb: Optional[Tuple[torch.FloatTensor]] = None, |
| 1452 | **kwargs, |
| 1453 | ) -> Tuple[torch.FloatTensor | Any]: |
| 1454 | """ |
| 1455 | Args: |
| 1456 | hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` |
| 1457 | attention_mask (`torch.FloatTensor`, *optional*): |
| 1458 | attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1, |
| 1459 | query_sequence_length, key_sequence_length)` if default attention is used. |
| 1460 | position_ids (`torch.LongTensor`, *optional*): |
| 1461 | Indices of positions of each input sequence tokens in the position embeddings. |
| 1462 | output_attentions (`bool`, *optional*): |
| 1463 | Whether or not to return the attentions tensors of all attention layers. See `attentions` under |
| 1464 | returned tensors for more detail. |
| 1465 | use_cache (`bool`, *optional*): |
| 1466 | If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding |
| 1467 | (see `past_key_values`). |
| 1468 | past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states |
| 1469 | custom_pos_emb (`Tuple[torch.FloatTensor]`, *optional*): custom position embedding for rotary |
| 1470 | position embedding |
| 1471 | """ |
| 1472 | if "padding_mask" in kwargs: |
| 1473 | warnings.warn( |
| 1474 | "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use " |
| 1475 | "`attention_mask` instead.`" |
| 1476 | ) |
| 1477 | |
| 1478 | residual = hidden_states |
| 1479 | |
| 1480 | hidden_states = self.input_layernorm(hidden_states) |
| 1481 | |
| 1482 | # Self Attention |
| 1483 | hidden_states, self_attn_weights, present_key_value = self.self_attn( |
| 1484 | hidden_states=hidden_states, |
| 1485 | attention_mask=attention_mask, |
| 1486 | position_ids=position_ids, |
| 1487 | past_key_value=past_key_value, |
| 1488 | output_attentions=output_attentions, |
| 1489 | use_cache=use_cache, |
| 1490 | custom_pos_emb=custom_pos_emb, |
| 1491 | **kwargs, |
| 1492 | ) |
| 1493 | hidden_states = residual + hidden_states |
| 1494 | # Fully Connected |
| 1495 | residual = hidden_states |
| 1496 | hidden_states = self.post_attention_layernorm(hidden_states) |
| 1497 | hidden_states = self.mlp(hidden_states) |
| 1498 | |
| 1499 | hidden_states = residual + hidden_states |
| 1500 | |
| 1501 | outputs = (hidden_states,) |
| 1502 | |
| 1503 | if output_attentions: |
| 1504 | outputs += (self_attn_weights,) |
| 1505 | |
| 1506 | if use_cache: |
| 1507 | outputs += (present_key_value,) |
| 1508 | |
| 1509 | return outputs |
| 1510 | |
| 1511 | |
| 1512 | @add_start_docstrings( |
| 1513 | "The bare Hunyuan Image 3 Model outputting raw hidden-states without any specific head on top.", |
| 1514 | Hunyuan_START_DOCSTRING, |
| 1515 | ) |
| 1516 | class HunyuanImage3PreTrainedModel(PreTrainedModel): |
| 1517 | config_class = HunyuanImage3Config |
| 1518 | base_model_prefix = "" |
| 1519 | supports_gradient_checkpointing = True |
| 1520 | _no_split_modules = ["HunyuanImage3DecoderLayer"] |
| 1521 | _skip_keys_device_placement = "past_key_values" |
| 1522 | _supports_flash_attn_2 = True |
| 1523 | _supports_sdpa = True |
| 1524 | _supports_cache_class = True |
| 1525 | |
| 1526 | def _init_weights(self, module): |
| 1527 | std = self.config.initializer_range |
| 1528 | if isinstance(module, nn.Linear): |
| 1529 | module.weight.data.normal_(mean=0.0, std=std) |
| 1530 | if module.bias is not None: |
| 1531 | module.bias.data.zero_() |
| 1532 | elif isinstance(module, nn.Embedding): |
| 1533 | module.weight.data.normal_(mean=0.0, std=std) |
| 1534 | if module.padding_idx is not None: |
| 1535 | module.weight.data[module.padding_idx].zero_() |
| 1536 | |
| 1537 | |
| 1538 | Hunyuan_INPUTS_DOCSTRING = r""" |
| 1539 | Args: |
| 1540 | input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): |
| 1541 | Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide |
| 1542 | it. |
| 1543 | |
| 1544 | Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and |
| 1545 | [`PreTrainedTokenizer.__call__`] for details. |
| 1546 | |
| 1547 | [What are input IDs?](../glossary#input-ids) |
| 1548 | attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): |
| 1549 | Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: |
| 1550 | |
| 1551 | - 1 for tokens that are **not masked**, |
| 1552 | - 0 for tokens that are **masked**. |
| 1553 | |
| 1554 | [What are attention masks?](../glossary#attention-mask) |
| 1555 | |
| 1556 | Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and |
| 1557 | [`PreTrainedTokenizer.__call__`] for details. |
| 1558 | |
| 1559 | If `past_key_values` is used, optionally only the last `input_ids` have to be input (see |
| 1560 | `past_key_values`). |
| 1561 | |
| 1562 | If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`] |
| 1563 | and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more |
| 1564 | information on the default strategy. |
| 1565 | |
| 1566 | - 1 indicates the head is **not masked**, |
| 1567 | - 0 indicates the head is **masked**. |
| 1568 | position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): |
| 1569 | Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, |
| 1570 | config.n_positions - 1]`. |
| 1571 | |
| 1572 | [What are position IDs?](../glossary#position-ids) |
| 1573 | past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*): |
| 1574 | Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention |
| 1575 | blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values` |
| 1576 | returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`. |
| 1577 | |
| 1578 | Two formats are allowed: |
| 1579 | - a [`~cache_utils.Cache`] instance; |
| 1580 | - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of |
| 1581 | shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy |
| 1582 | cache format. |
| 1583 | |
| 1584 | The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the |
| 1585 | legacy cache format will be returned. |
| 1586 | |
| 1587 | If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't |
| 1588 | have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids` |
| 1589 | of shape `(batch_size, sequence_length)`. |
| 1590 | inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): |
| 1591 | Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This |
| 1592 | is useful if you want more control over how to convert `input_ids` indices into associated vectors than the |
| 1593 | model's internal embedding lookup matrix. |
| 1594 | use_cache (`bool`, *optional*): |
| 1595 | If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see |
| 1596 | `past_key_values`). |
| 1597 | output_attentions (`bool`, *optional*): |
| 1598 | Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned |
| 1599 | tensors for more detail. |
| 1600 | output_hidden_states (`bool`, *optional*): |
| 1601 | Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for |
| 1602 | more detail. |
| 1603 | return_dict (`bool`, *optional*): |
| 1604 | Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. |
| 1605 | """ |
| 1606 | |
| 1607 | |
| 1608 | @add_start_docstrings( |
| 1609 | "The bare Hunyuan Model outputting raw hidden-states without any specific head on top.", |
| 1610 | Hunyuan_START_DOCSTRING, |
| 1611 | ) |
| 1612 | class HunyuanImage3Model(HunyuanImage3PreTrainedModel): |
| 1613 | def __init__(self, config: HunyuanImage3Config): |
| 1614 | super().__init__(config) |
| 1615 | self.padding_idx = config.pad_token_id |
| 1616 | self.vocab_size = config.vocab_size |
| 1617 | self.add_classification_head = config.add_classification_head |
| 1618 | self.wte = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx) |
| 1619 | self.layers = nn.ModuleList( |
| 1620 | [HunyuanImage3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] |
| 1621 | ) |
| 1622 | if not config.add_classification_head: |
| 1623 | self.ln_f = HunyuanRMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 1624 | |
| 1625 | # Initialize weights and apply final processing |
| 1626 | self.post_init() |
| 1627 | |
| 1628 | self.shared_tensor = None |
| 1629 | |
| 1630 | @add_start_docstrings_to_model_forward(Hunyuan_INPUTS_DOCSTRING) |
| 1631 | def forward( |
| 1632 | self, |
| 1633 | input_ids: torch.LongTensor = None, |
| 1634 | attention_mask: Optional[torch.Tensor] = None, |
| 1635 | position_ids: Optional[torch.LongTensor] = None, |
| 1636 | past_key_values: Optional[List[torch.FloatTensor]] = None, |
| 1637 | inputs_embeds: Optional[torch.FloatTensor] = None, |
| 1638 | use_cache: Optional[bool] = None, |
| 1639 | output_attentions: Optional[bool] = None, |
| 1640 | output_hidden_states: Optional[bool] = None, |
| 1641 | return_dict: Optional[bool] = None, |
| 1642 | custom_pos_emb: Optional[Tuple[torch.FloatTensor]] = None, |
| 1643 | mode: str = "gen_text", |
| 1644 | first_step: Optional[bool] = None, |
| 1645 | gen_timestep_scatter_index: Optional[torch.Tensor] = None, |
| 1646 | ) -> Union[Tuple, BaseModelOutputWithPast]: |
| 1647 | |
| 1648 | output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions |
| 1649 | output_hidden_states = ( |
| 1650 | output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states |
| 1651 | ) |
| 1652 | use_cache = use_cache if use_cache is not None else self.config.use_cache |
| 1653 | |
| 1654 | return_dict = return_dict if return_dict is not None else self.config.use_return_dict |
| 1655 | |
| 1656 | if inputs_embeds is None: |
| 1657 | inputs_embeds = self.wte(input_ids) |
| 1658 | |
| 1659 | # embed positions |
| 1660 | hidden_states = inputs_embeds |
| 1661 | |
| 1662 | # decoder layers |
| 1663 | all_hidden_states = () if output_hidden_states else None |
| 1664 | all_self_attns = () if output_attentions else None |
| 1665 | next_decoder_cache = None |
| 1666 | |
| 1667 | for layer_idx, decoder_layer in enumerate(self.layers): |
| 1668 | if output_hidden_states: |
| 1669 | all_hidden_states += (hidden_states,) |
| 1670 | |
| 1671 | layer_outputs = decoder_layer( |
| 1672 | hidden_states, |
| 1673 | attention_mask=attention_mask, |
| 1674 | position_ids=position_ids, |
| 1675 | past_key_value=past_key_values, |
| 1676 | output_attentions=output_attentions, |
| 1677 | use_cache=use_cache, |
| 1678 | custom_pos_emb=custom_pos_emb, |
| 1679 | mode=mode, |
| 1680 | first_step=first_step, |
| 1681 | gen_timestep_scatter_index=gen_timestep_scatter_index, |
| 1682 | ) |
| 1683 | |
| 1684 | hidden_states = layer_outputs[0] |
| 1685 | |
| 1686 | if use_cache: |
| 1687 | next_decoder_cache = layer_outputs[2 if output_attentions else 1] |
| 1688 | |
| 1689 | if output_attentions: |
| 1690 | all_self_attns += (layer_outputs[1],) |
| 1691 | |
| 1692 | if not self.add_classification_head: |
| 1693 | # Do ln_f outside of the model for compatibility with image generation. |
| 1694 | pass |
| 1695 | # hidden_states = self.ln_f(hidden_states) |
| 1696 | |
| 1697 | # add hidden states from the last decoder layer |
| 1698 | if output_hidden_states: |
| 1699 | all_hidden_states += (hidden_states,) |
| 1700 | |
| 1701 | next_cache = None |
| 1702 | if use_cache: |
| 1703 | next_cache = next_decoder_cache |
| 1704 | if not return_dict: |
| 1705 | return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None) |
| 1706 | return BaseModelOutputWithPast( |
| 1707 | last_hidden_state=hidden_states, |
| 1708 | past_key_values=next_cache, |
| 1709 | hidden_states=all_hidden_states, |
| 1710 | attentions=all_self_attns, |
| 1711 | ) |
| 1712 | |
| 1713 | |
| 1714 | class HunyuanImage3ForCausalMM(HunyuanImage3PreTrainedModel, GenerationMixin): |
| 1715 | def __init__(self, config: HunyuanImage3Config): |
| 1716 | super().__init__(config) |
| 1717 | self.config = config |
| 1718 | self._tkwrapper: Optional[TokenizerWrapper] = None |
| 1719 | |
| 1720 | # Initialize image preprocessor (for conditional images) |
| 1721 | self.image_processor = HunyuanImage3ImageProcessor(config) |
| 1722 | |
| 1723 | # vae and gen_image pipeline |
| 1724 | self.vae = AutoencoderKLConv3D.from_config(config.vae) |
| 1725 | self._pipeline = None |
| 1726 | |
| 1727 | # vit |
| 1728 | self.vision_model = Siglip2VisionTransformer(config.vit) |
| 1729 | self.vision_aligner = LightProjector(config.vit_aligner) |
| 1730 | |
| 1731 | # image generation related |
| 1732 | self.timestep_emb = TimestepEmbedder(hidden_size=config.hidden_size) |
| 1733 | if config.img_proj_type == "unet": |
| 1734 | self.patch_embed = UNetDown( |
| 1735 | patch_size=config.patch_size, |
| 1736 | emb_channels=config.hidden_size, |
| 1737 | in_channels=config.vae["latent_channels"], |
| 1738 | hidden_channels=config.patch_embed_hidden_dim, |
| 1739 | out_channels=config.hidden_size, |
| 1740 | ) |
| 1741 | self.time_embed = TimestepEmbedder(hidden_size=config.hidden_size) |
| 1742 | |
| 1743 | self.final_layer = UNetUp( |
| 1744 | patch_size=config.patch_size, |
| 1745 | emb_channels=config.hidden_size, |
| 1746 | in_channels=config.hidden_size, |
| 1747 | hidden_channels=config.patch_embed_hidden_dim, |
| 1748 | out_channels=config.vae["latent_channels"], |
| 1749 | out_norm=True, |
| 1750 | ) |
| 1751 | self.time_embed_2 = TimestepEmbedder(hidden_size=config.hidden_size) |
| 1752 | else: |
| 1753 | raise ValueError(f"Unknown img_proj_type {config.img_proj_type}") |
| 1754 | |
| 1755 | # transformer backbone |
| 1756 | self.model = HunyuanImage3Model(config) |
| 1757 | |
| 1758 | self.pad_id = config.pad_id |
| 1759 | self.vocab_size = config.vocab_size |
| 1760 | |
| 1761 | # linear head |
| 1762 | self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) |
| 1763 | |
| 1764 | # Initialize weights and apply final processing |
| 1765 | self.post_init() |
| 1766 | |
| 1767 | @property |
| 1768 | def tokenizer(self): |
| 1769 | if self._tkwrapper is None: |
| 1770 | raise ValueError("Attribute `tokenizer` has not been initialized yet. Please set it first.") |
| 1771 | return self._tkwrapper |
| 1772 | |
| 1773 | def load_tokenizer(self, tokenizer): |
| 1774 | self._tkwrapper = TokenizerWrapper(tokenizer) |
| 1775 | |
| 1776 | @property |
| 1777 | def pipeline(self): |
| 1778 | if self._pipeline is None: |
| 1779 | self.scheduler = FlowMatchDiscreteScheduler( |
| 1780 | shift=self.generation_config.flow_shift, reverse=True, solver="euler", |
| 1781 | ) |
| 1782 | self._pipeline = HunyuanImage3Text2ImagePipeline( |
| 1783 | model=self, scheduler=self.scheduler, vae=self.vae, |
| 1784 | ) |
| 1785 | return self._pipeline |
| 1786 | |
| 1787 | @staticmethod |
| 1788 | def get_pos_emb(custom_pos_emb, position_ids): |
| 1789 | cos, sin = custom_pos_emb |
| 1790 | cos = real_batched_index_select(cos, dim=1, idx=position_ids) |
| 1791 | sin = real_batched_index_select(sin, dim=1, idx=position_ids) |
| 1792 | return cos, sin |
| 1793 | |
| 1794 | def instantiate_vae_image_tokens( |
| 1795 | self, |
| 1796 | x: torch.Tensor, |
| 1797 | images: BatchRaggedImages, |
| 1798 | ts: BatchRaggedTensor, |
| 1799 | image_mask: torch.Tensor, |
| 1800 | ): |
| 1801 | """ |
| 1802 | Instantiate the VAE image embeddings into the input embedding sequence. |
| 1803 | |
| 1804 | Args: |
| 1805 | x: input sequence, (batch_size, seq_len, n_embd) |
| 1806 | images: BatchRaggedImages |
| 1807 | images can be a 4-D tensor, or a list of 4-D tensors, or a list of lists of 3-D tensors. |
| 1808 | ts: BatchRaggedTensor |
| 1809 | ts can be a 1-D tensor, or a list of 1-D tensors |
| 1810 | image_mask: (batch_size, seq_len) |
| 1811 | """ |
| 1812 | batch_size, seq_len, n_embd = x.shape |
| 1813 | |
| 1814 | if isinstance(images, list): |
| 1815 | index = torch.arange(seq_len, device=x.device).unsqueeze(0).repeat(batch_size, 1) |
| 1816 | t_emb = [] |
| 1817 | for i, (image_i, t_i) in enumerate(zip(images, ts)): |
| 1818 | if isinstance(image_i, torch.Tensor): |
| 1819 | # time_embed needs a 1-D tensor as input |
| 1820 | t_i_emb = self.time_embed(t_i) |
| 1821 | # n_{i} x one_image_seq_len x n_embd |
| 1822 | image_i_seq, _, _ = self.patch_embed(image_i, t_i_emb) |
| 1823 | # 1 x (n_{i} * one_image_seq_len) |
| 1824 | image_i_scatter_index = index[i:i + 1].masked_select(image_mask[i:i + 1].bool()).reshape(1, -1) |
| 1825 | x[i:i + 1].scatter_( |
| 1826 | dim=1, |
| 1827 | index=image_i_scatter_index.unsqueeze(-1).repeat(1, 1, n_embd), |
| 1828 | # 1 x (n_{i} * one_image_seq_len) x n_embd |
| 1829 | src=image_i_seq.reshape(1, -1, n_embd), # 1 x (n_{i} * one_image_seq_len) x n_embd |
| 1830 | ) |
| 1831 | t_emb.append(t_i_emb) |
| 1832 | elif isinstance(image_i, list): |
| 1833 | # time_embed needs a 1-D tensor as input |
| 1834 | t_i_emb = self.time_embed(t_i) # n_{i} x d |
| 1835 | image_i_seq_list = [], [] |
| 1836 | for j in range(len(image_i)): |
| 1837 | image_ij = image_i[j] |
| 1838 | if image_ij.dim() == 4: |
| 1839 | assert image_i[j].shape[0] == 1, "image_i[j] should have a batch dimension of 1" |
| 1840 | elif image_ij.dim() == 3: |
| 1841 | image_ij = image_ij.unsqueeze(0) |
| 1842 | else: |
| 1843 | raise ValueError(f"image_i[j] should have 3 or 4 dimensions, got {image_ij.dim()}") |
| 1844 | # 1 x one_image_seq_len_{j} x n_embd |
| 1845 | image_i_seq_j, _, _ = self.patch_embed(image_ij, t_i_emb[j:j + 1]) |
| 1846 | image_i_seq_list.append(image_i_seq_j) |
| 1847 | # 1 x sum_{j}(one_image_seq_len_{j}) x n_embd |
| 1848 | image_i_seq = torch.cat(image_i_seq_list, dim=1) |
| 1849 | # 1 x sum_{j}(one_image_seq_len_{j}) |
| 1850 | image_i_scatter_index = index[i:i + 1].masked_select(image_mask[i:i + 1].bool()).reshape(1, -1) |
| 1851 | x[i:i + 1].scatter_( |
| 1852 | dim=1, |
| 1853 | index=image_i_scatter_index.unsqueeze(-1).repeat(1, 1, n_embd), |
| 1854 | # 1 x sum_{j}(one_image_seq_len_{j}) x n_embd |
| 1855 | src=image_i_seq.reshape(1, -1, n_embd), # 1 x sum_{j}(one_image_seq_len_{j}) x n_embd |
| 1856 | ) |
| 1857 | t_emb.append(t_i_emb) |
| 1858 | else: |
| 1859 | raise TypeError(f"image_i should be a torch.Tensor or a list, got {type(image_i)}") |
| 1860 | token_h, token_w = None, None |
| 1861 | else: |
| 1862 | # images is a 4-D tensor |
| 1863 | batch_size, seq_len, n_embd = x.shape |
| 1864 | index = torch.arange(seq_len, device=x.device).unsqueeze(0).repeat(batch_size, 1) |
| 1865 | t_emb = self.time_embed(ts) |
| 1866 | image_seq, token_h, token_w = self.patch_embed(images, t_emb) |
| 1867 | image_scatter_index = index.masked_select(image_mask.bool()).reshape(batch_size, -1) |
| 1868 | x.scatter_( |
| 1869 | dim=1, |
| 1870 | index=image_scatter_index.unsqueeze(-1).repeat(1, 1, n_embd), |
| 1871 | src=image_seq, |
| 1872 | ) |
| 1873 | |
| 1874 | return x, token_h, token_w |
| 1875 | |
| 1876 | def instantiate_timestep_tokens( |
| 1877 | self, |
| 1878 | x: torch.Tensor, |
| 1879 | t: BatchRaggedTensor, |
| 1880 | timestep_scatter_index: BatchRaggedTensor, |
| 1881 | ): |
| 1882 | batch_size, seq_len, n_embd = x.shape |
| 1883 | # batch_size x n x n_embd |
| 1884 | timestep_scatter_src = self.timestep_emb(t.reshape(-1)).reshape(batch_size, -1, n_embd) |
| 1885 | x.scatter_( |
| 1886 | dim=1, |
| 1887 | index=timestep_scatter_index.unsqueeze(-1).repeat(1, 1, n_embd), |
| 1888 | src=timestep_scatter_src, |
| 1889 | ) |
| 1890 | |
| 1891 | return x |
| 1892 | |
| 1893 | def instantiate_vit_image_tokens( |
| 1894 | self, |
| 1895 | x: torch.Tensor, |
| 1896 | cond_vit_images: Union[torch.Tensor, List[torch.Tensor]], |
| 1897 | cond_vit_image_mask: torch.Tensor, |
| 1898 | vit_kwargs: Dict[str, Any], |
| 1899 | ): |
| 1900 | # 1. Forward the vit encoder and vit aligner to get the vit image embeddings and align them to the |
| 1901 | # transformer hidden size |
| 1902 | cond_vit_image_embeds = [] |
| 1903 | for batch_idx, image in enumerate(cond_vit_images): |
| 1904 | cur_kwargs = {k: v[batch_idx] for k, v in vit_kwargs.items()} |
| 1905 | image_embed = self.vision_model(image, **cur_kwargs).last_hidden_state |
| 1906 | image_embed = self.vision_aligner(image_embed) |
| 1907 | n, seq_len, dim = image_embed.shape |
| 1908 | image_embed = image_embed.reshape(n * seq_len, dim) |
| 1909 | cond_vit_image_embeds.append(image_embed) |
| 1910 | |
| 1911 | # 2. Instantiate the vit image embeddings into the input sequence |
| 1912 | batch_size, seq_len, n_embd = x.shape |
| 1913 | index = torch.arange(seq_len, device=x.device).unsqueeze(0).repeat(batch_size, 1) |
| 1914 | |
| 1915 | for i, (image_embed, mask) in enumerate(zip(cond_vit_image_embeds, cond_vit_image_mask)): |
| 1916 | image_scatter_index = index[i:i+1].masked_select(mask.bool()).reshape(1, -1) |
| 1917 | x[i:i+1].scatter_( |
| 1918 | dim=1, |
| 1919 | index=image_scatter_index.unsqueeze(-1).repeat(1, 1, n_embd), |
| 1920 | src=image_embed.reshape(1, -1, n_embd), |
| 1921 | ) |
| 1922 | |
| 1923 | return x |
| 1924 | |
| 1925 | def ragged_final_layer(self, x, image_mask, timestep, token_h, token_w, first_step): |
| 1926 | bsz, seq_len, n_embd = x.shape |
| 1927 | if first_step: |
| 1928 | image_output = x.masked_select(image_mask.unsqueeze(-1).bool()).reshape(bsz, -1, n_embd) |
| 1929 | else: |
| 1930 | image_output = x[:, 1:, :] |
| 1931 | timestep_emb = self.time_embed_2(timestep) |
| 1932 | pred = self.final_layer(image_output, timestep_emb, token_h, token_w) |
| 1933 | return pred |
| 1934 | |
| 1935 | @staticmethod |
| 1936 | def _check_inputs(cond, target, check_list): |
| 1937 | if cond: |
| 1938 | for name, item in check_list: |
| 1939 | assert item is not None, f"`{name}` should be provided when `{target}`." |
| 1940 | |
| 1941 | @add_start_docstrings_to_model_forward(Hunyuan_INPUTS_DOCSTRING) |
| 1942 | def forward( |
| 1943 | self, |
| 1944 | input_ids: torch.LongTensor = None, |
| 1945 | attention_mask: Optional[torch.Tensor] = None, |
| 1946 | position_ids: Optional[torch.LongTensor] = None, |
| 1947 | past_key_values: Optional[List[torch.FloatTensor]] = None, |
| 1948 | use_cache: Optional[bool] = None, |
| 1949 | output_attentions: Optional[bool] = None, |
| 1950 | output_hidden_states: Optional[bool] = None, |
| 1951 | return_dict: Optional[bool] = None, |
| 1952 | custom_pos_emb: Optional[Tuple[torch.FloatTensor]] = None, |
| 1953 | mode: str = "gen_text", |
| 1954 | first_step: Optional[bool] = None, |
| 1955 | # for gen image |
| 1956 | images: Optional[BatchRaggedImages] = None, |
| 1957 | image_mask: Optional[torch.Tensor] = None, |
| 1958 | timestep: Optional[BatchRaggedTensor] = None, |
| 1959 | gen_timestep_scatter_index: Optional[torch.Tensor] = None, |
| 1960 | # for cond image |
| 1961 | cond_vae_images: Optional[BatchRaggedImages] = None, |
| 1962 | cond_timestep: Optional[BatchRaggedTensor] = None, |
| 1963 | cond_vae_image_mask: Optional[torch.Tensor] = None, |
| 1964 | cond_vit_images: Optional[BatchRaggedImages] = None, |
| 1965 | cond_vit_image_mask: Optional[torch.Tensor] = None, |
| 1966 | vit_kwargs: Optional[Dict[str, Any]] = None, |
| 1967 | cond_timestep_scatter_index: Optional[torch.Tensor] = None, |
| 1968 | ) -> Union[Tuple, CausalMMOutputWithPast]: |
| 1969 | return_dict = return_dict if return_dict is not None else self.config.use_return_dict |
| 1970 | # Sanity Check of Inputs |
| 1971 | self._check_inputs(mode == "gen_image", "in `gen_image` mode", [ |
| 1972 | ("images", images), ("timestep", timestep), ("gen_timestep_scatter_index", gen_timestep_scatter_index), |
| 1973 | ]) |
| 1974 | self._check_inputs(mode == "gen_image" and first_step, "in `gen_image` mode at the first step", [ |
| 1975 | ("image_mask", image_mask), |
| 1976 | ]) |
| 1977 | self._check_inputs(cond_vae_images is not None, "`cond_vae_images` is provided", [ |
| 1978 | ("cond_timestep", cond_timestep), ("cond_vae_image_mask", cond_vae_image_mask), |
| 1979 | ("cond_timestep_scatter_index", cond_timestep_scatter_index), |
| 1980 | ]) |
| 1981 | self._check_inputs(cond_vit_images is not None, "`cond_vit_images` is provided", [ |
| 1982 | ("cond_vit_image_mask", cond_vit_image_mask), ("vit_kwargs", vit_kwargs), |
| 1983 | ]) |
| 1984 | |
| 1985 | custom_pos_emb = self.get_pos_emb(custom_pos_emb, position_ids) |
| 1986 | |
| 1987 | inputs_embeds = self.model.wte(input_ids) |
| 1988 | bsz, seq_len, n_embd = inputs_embeds.shape |
| 1989 | |
| 1990 | # Instantiate placeholder tokens: <timestep>, <img> for the gen image |
| 1991 | if mode == "gen_text": |
| 1992 | # For gen_text, make sure gen_timestep_scatter_index is None |
| 1993 | gen_timestep_scatter_index = None |
| 1994 | token_h, token_w = None, None |
| 1995 | else: |
| 1996 | if first_step: |
| 1997 | inputs_embeds, token_h, token_w = self.instantiate_vae_image_tokens( |
| 1998 | inputs_embeds, images, timestep, image_mask) |
| 1999 | inputs_embeds = self.instantiate_timestep_tokens( |
| 2000 | inputs_embeds, timestep, gen_timestep_scatter_index) |
| 2001 | else: |
| 2002 | t_emb = self.time_embed(timestep) |
| 2003 | image_emb, token_h, token_w = self.patch_embed(images, t_emb) |
| 2004 | timestep_emb = self.timestep_emb(timestep).reshape(bsz, -1, n_embd) |
| 2005 | inputs_embeds = torch.cat([timestep_emb, image_emb], dim=1) |
| 2006 | |
| 2007 | # Instantiate placeholder tokens: <timestep>, <img> for cond images |
| 2008 | # Should only run once with kv-cache enabled. |
| 2009 | if cond_vae_images is not None: |
| 2010 | inputs_embeds, _, _ = self.instantiate_vae_image_tokens( |
| 2011 | inputs_embeds, cond_vae_images, cond_timestep, cond_vae_image_mask) |
| 2012 | inputs_embeds = self.instantiate_timestep_tokens( |
| 2013 | inputs_embeds, cond_timestep, cond_timestep_scatter_index) |
| 2014 | if cond_vit_images is not None: |
| 2015 | inputs_embeds = self.instantiate_vit_image_tokens( |
| 2016 | inputs_embeds, cond_vit_images, cond_vit_image_mask, vit_kwargs) |
| 2017 | |
| 2018 | # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) |
| 2019 | outputs = self.model( |
| 2020 | input_ids=input_ids, |
| 2021 | attention_mask=attention_mask, |
| 2022 | position_ids=position_ids, |
| 2023 | past_key_values=past_key_values, |
| 2024 | inputs_embeds=inputs_embeds, |
| 2025 | use_cache=use_cache, |
| 2026 | output_attentions=output_attentions, |
| 2027 | output_hidden_states=output_hidden_states, |
| 2028 | return_dict=return_dict, |
| 2029 | custom_pos_emb=custom_pos_emb, |
| 2030 | mode=mode, |
| 2031 | first_step=first_step, |
| 2032 | gen_timestep_scatter_index=gen_timestep_scatter_index, |
| 2033 | ) |
| 2034 | hidden_states = outputs[0] |
| 2035 | |
| 2036 | if mode == "gen_text": |
| 2037 | hidden_states = self.model.ln_f(hidden_states) |
| 2038 | logits = self.lm_head(hidden_states) |
| 2039 | logits = logits.float() |
| 2040 | diffusion_prediction = None |
| 2041 | else: |
| 2042 | logits = None |
| 2043 | hidden_states = hidden_states.to(input_ids.device) |
| 2044 | diffusion_prediction = self.ragged_final_layer( |
| 2045 | hidden_states, image_mask, timestep, token_h, token_w, first_step) |
| 2046 | |
| 2047 | if not return_dict: |
| 2048 | output = (logits,) + outputs[1:] + (diffusion_prediction,) |
| 2049 | return output |
| 2050 | |
| 2051 | output = CausalMMOutputWithPast( |
| 2052 | logits=logits, |
| 2053 | past_key_values=outputs.past_key_values, |
| 2054 | hidden_states=outputs.hidden_states, |
| 2055 | attentions=outputs.attentions, |
| 2056 | diffusion_prediction=diffusion_prediction, |
| 2057 | ) |
| 2058 | |
| 2059 | return output |
| 2060 | |
| 2061 | @staticmethod |
| 2062 | def check_inputs(prompt=None, message_list=None): |
| 2063 | if prompt is None and message_list is None: |
| 2064 | raise ValueError("Either `prompt` or `message_list` should be provided.") |
| 2065 | if prompt is not None and message_list is not None: |
| 2066 | raise ValueError("Only one of `prompt` or `message_list` should be provided.") |
| 2067 | if prompt is not None: |
| 2068 | assert isinstance(prompt, str) or isinstance(prompt, list), \ |
| 2069 | f"`prompt` should be a string or a list of strings, but got {type(prompt)}." |
| 2070 | if isinstance(prompt, list): |
| 2071 | assert len(prompt) > 0 and all(isinstance(p, str) for p in prompt), \ |
| 2072 | "`prompt` should be a non-empty list of strings." |
| 2073 | if message_list is not None: |
| 2074 | if not isinstance(message_list, list): |
| 2075 | raise ValueError(f"`message_list` should be a list of messages, but got {type(message_list)}.") |
| 2076 | assert len(message_list) > 0, "`message_list` should be a non-empty list." |
| 2077 | for message in message_list: |
| 2078 | assert isinstance(message, list) or isinstance(message, dict), \ |
| 2079 | f"Each message should be a list of dicts or a dict, but got {type(message)}." |
| 2080 | |
| 2081 | @staticmethod |
| 2082 | def prepare_seed(seed, batch_size): |
| 2083 | if isinstance(seed, torch.Tensor): |
| 2084 | seed = seed.tolist() |
| 2085 | if seed is None: |
| 2086 | seeds = [random.randint(0, 10_000_000) for _ in range(batch_size)] |
| 2087 | elif isinstance(seed, int): |
| 2088 | seeds = [seed for _ in range(batch_size)] |
| 2089 | elif isinstance(seed, (list, tuple)): |
| 2090 | if len(seed) == batch_size: |
| 2091 | seeds = [int(seed[i]) for i in range(batch_size)] |
| 2092 | else: |
| 2093 | raise ValueError(f"Length of seed must be equal to the batch_size({batch_size}), got {seed}.") |
| 2094 | else: |
| 2095 | raise ValueError(f"Seed must be an integer, a list of integers, or None, got {seed}.") |
| 2096 | return seeds |
| 2097 | |
| 2098 | @staticmethod |
| 2099 | def build_batch_rope_image_info(output, sections): |
| 2100 | rope_image_info = [] |
| 2101 | for image_slices, sections_i in zip(output.all_image_slices, sections): |
| 2102 | image_shapes = [] |
| 2103 | for section in sections_i: |
| 2104 | if 'image' in section['type']: |
| 2105 | if isinstance(section['token_height'], list): |
| 2106 | assert len(section['token_height']) == len(section['token_height']), \ |
| 2107 | (f"token_height and token_width should have the same length, " |
| 2108 | f"but got {len(section['token_height'])} and {len(section['token_width'])}") |
| 2109 | image_shapes.extend(list(zip(section['token_height'], section['token_width']))) |
| 2110 | else: |
| 2111 | image_shapes.append((section['token_height'], section['token_width'])) |
| 2112 | assert len(image_slices) == len(image_shapes), ( |
| 2113 | f"Size miss matching: Image slices({len(image_slices)}) != image shapes({len(image_shapes)})" |
| 2114 | ) |
| 2115 | rope_image_info.append(list(zip(image_slices, image_shapes))) |
| 2116 | return rope_image_info |
| 2117 | |
| 2118 | def vae_encode(self, image, cfg_factor=1): |
| 2119 | config = self.vae.config |
| 2120 | |
| 2121 | with torch.autocast(device_type="cuda", dtype=torch.float16, enabled=True): |
| 2122 | vae_encode_result = self.vae.encode(image) |
| 2123 | if isinstance(vae_encode_result, torch.Tensor): |
| 2124 | latents = vae_encode_result |
| 2125 | else: |
| 2126 | latents = vae_encode_result.latent_dist.sample() |
| 2127 | if hasattr(config, 'shift_factor') and config.shift_factor: |
| 2128 | latents.sub_(config.shift_factor) |
| 2129 | if hasattr(config, 'scaling_factor') and config.scaling_factor: |
| 2130 | latents.mul_(config.scaling_factor) |
| 2131 | |
| 2132 | if hasattr(self.vae, "ffactor_temporal"): |
| 2133 | assert latents.shape[2] == 1, "latents should have shape [B, C, T, H, W] and T should be 1" |
| 2134 | latents = latents.squeeze(2) |
| 2135 | |
| 2136 | # Here we always use t=0 to declare it is a clean conditional image |
| 2137 | t = torch.zeros((latents.shape[0],)) |
| 2138 | |
| 2139 | if cfg_factor > 1: |
| 2140 | t = t.repeat(cfg_factor) |
| 2141 | latents = latents.repeat(cfg_factor, 1, 1, 1) |
| 2142 | |
| 2143 | return t, latents |
| 2144 | |
| 2145 | def _encode_cond_image( |
| 2146 | self, |
| 2147 | batch_cond_image_info_list: List[List[JointImageInfo]], |
| 2148 | cfg_factor: int = 1, |
| 2149 | ): |
| 2150 | # VAE encode one by one, as we assume cond images have different sizes |
| 2151 | batch_cond_vae_images, batch_cond_t, batch_cond_vit_images = [], [], [] |
| 2152 | for cond_image_info_list in batch_cond_image_info_list: |
| 2153 | cond_vae_image_list, cond_t_list, cond_vit_image_list = [], [], [] |
| 2154 | for image_info in cond_image_info_list: |
| 2155 | cond_t_, cond_vae_image_ = self.vae_encode( |
| 2156 | image_info.vae_image_info.image_tensor.to(self.device), |
| 2157 | ) |
| 2158 | cond_vit_image_list.append(image_info.vision_image_info.image_tensor) |
| 2159 | cond_vae_image_list.append(cond_vae_image_.squeeze(0)) |
| 2160 | cond_t_list.append(cond_t_) |
| 2161 | batch_cond_vae_images.append(cond_vae_image_list) |
| 2162 | batch_cond_t.append(cond_t_list) |
| 2163 | batch_cond_vit_images.append(torch.cat(cond_vit_image_list, dim=0)) |
| 2164 | |
| 2165 | # If only one cond image for each sample and all have the same size, we can batch them together |
| 2166 | # In this case, cond_vae_images is a 4-D tensor. |
| 2167 | if all([len(items) == 1 for items in batch_cond_vae_images]) and all( |
| 2168 | items[0].shape == batch_cond_vae_images[0][0].shape for items in batch_cond_vae_images): |
| 2169 | cond_vae_images = torch.stack([items[0] for items in batch_cond_vae_images], dim=0) |
| 2170 | cond_t = torch.cat([items[0] for items in batch_cond_t], dim=0) |
| 2171 | if cfg_factor > 1: |
| 2172 | cond_t = cond_t.repeat(cfg_factor) |
| 2173 | cond_vae_images = cond_vae_images.repeat(cfg_factor, 1, 1, 1) |
| 2174 | else: |
| 2175 | # In this case, cond_vae_images is a list of 4-D tensors or a list of lists of 3-D tensors. |
| 2176 | cond_t = [torch.cat(item, dim=0) for item in batch_cond_t] |
| 2177 | cond_vae_images = [] |
| 2178 | for items in batch_cond_vae_images: |
| 2179 | if all(items[0].shape == item.shape for item in items): |
| 2180 | cond_vae_images.append(torch.stack(items, dim=0)) |
| 2181 | else: |
| 2182 | cond_vae_images.append(items) |
| 2183 | if cfg_factor > 1: |
| 2184 | cond_t = cond_t * cfg_factor |
| 2185 | cond_vae_images = cond_vae_images * cfg_factor |
| 2186 | |
| 2187 | if cfg_factor > 1: |
| 2188 | batch_cond_vit_images = batch_cond_vit_images * cfg_factor |
| 2189 | |
| 2190 | return cond_vae_images, cond_t, batch_cond_vit_images |
| 2191 | |
| 2192 | def prepare_model_inputs( |
| 2193 | self, |
| 2194 | prompt=None, |
| 2195 | mode="gen_text", |
| 2196 | system_prompt=None, |
| 2197 | cot_text=None, |
| 2198 | image_size="auto", |
| 2199 | message_list=None, |
| 2200 | device=None, |
| 2201 | max_new_tokens=None, |
| 2202 | **kwargs, |
| 2203 | ): |
| 2204 | # 1. Sanity check |
| 2205 | self.check_inputs(prompt, message_list) |
| 2206 | device = default(device, self.device) |
| 2207 | |
| 2208 | # 2. Format inputs |
| 2209 | batch_message_list = message_list |
| 2210 | batch_prompt = prompt |
| 2211 | batch_cot_text = cot_text |
| 2212 | batch_system_prompt = system_prompt |
| 2213 | batch_gen_image_info = None |
| 2214 | # TODO: construct with user input images |
| 2215 | batch_cond_image_info = None |
| 2216 | |
| 2217 | # -- 2.1 message_list |
| 2218 | if batch_message_list is not None: |
| 2219 | if isinstance(batch_message_list[0], dict): |
| 2220 | batch_message_list = [batch_message_list] |
| 2221 | batch_size = len(batch_message_list) |
| 2222 | |
| 2223 | batch_gen_image_info = [ |
| 2224 | [message['content'] for message in message_list_ if message['type'] == 'gen_image'] |
| 2225 | for message_list_ in batch_message_list |
| 2226 | ] |
| 2227 | # At most one gen_image is allowed for each message_list |
| 2228 | batch_gen_image_info = [info[-1] if len(info) > 0 else None for info in batch_gen_image_info] |
| 2229 | # Multiple cond images are allowed. |
| 2230 | batch_cond_image_info = [ |
| 2231 | [message['content'] for message in message_list_ if message['type'] == 'joint_image'] |
| 2232 | for message_list_ in batch_message_list |
| 2233 | ] |
| 2234 | |
| 2235 | # -- 2.2 Prompt, cot text, system prompt |
| 2236 | else: |
| 2237 | if isinstance(batch_prompt, str): |
| 2238 | batch_prompt = [batch_prompt] |
| 2239 | batch_size = len(batch_prompt) |
| 2240 | |
| 2241 | if batch_cot_text is not None: |
| 2242 | if isinstance(batch_cot_text, str): |
| 2243 | batch_cot_text = [batch_cot_text] |
| 2244 | else: |
| 2245 | assert isinstance(batch_cot_text, list) and len(batch_cot_text) == batch_size, \ |
| 2246 | "`cot_text` should be a string or a list of strings with the same length as `prompt`." |
| 2247 | |
| 2248 | if batch_system_prompt is not None: |
| 2249 | if isinstance(batch_system_prompt, str): |
| 2250 | batch_system_prompt = [batch_system_prompt] |
| 2251 | else: |
| 2252 | assert isinstance(batch_system_prompt, list) and len(batch_system_prompt) == batch_size, \ |
| 2253 | "`system_prompts` should be a string or a list of strings with the same length as `prompt`." |
| 2254 | |
| 2255 | if mode == "gen_image": |
| 2256 | batch_gen_image_info = [self.image_processor.build_image_info(image_size) for _ in range(batch_size)] |
| 2257 | |
| 2258 | # -- 2.3 seed |
| 2259 | seeds = self.prepare_seed(seed=kwargs.get('seed'), batch_size=batch_size) |
| 2260 | generator = [torch.Generator(self.device).manual_seed(seed) for seed in seeds] |
| 2261 | |
| 2262 | # 3. apply chat template |
| 2263 | cfg_factor = {"gen_text": 1, "gen_image": 2} |
| 2264 | bot_task = kwargs.pop("bot_task", "auto") |
| 2265 | # If `drop_think` enabled, always drop <think> parts in the context. |
| 2266 | drop_think = kwargs.get('drop_think', self.generation_config.drop_think) |
| 2267 | # Apply batched prompt or batched message_list to build input sequence with associated info. |
| 2268 | out = self._tkwrapper.apply_chat_template( |
| 2269 | batch_prompt=batch_prompt, |
| 2270 | batch_message_list=batch_message_list, |
| 2271 | mode=mode, |
| 2272 | batch_gen_image_info=batch_gen_image_info, |
| 2273 | batch_cond_image_info=batch_cond_image_info, |
| 2274 | batch_system_prompt=batch_system_prompt, |
| 2275 | batch_cot_text=batch_cot_text, |
| 2276 | max_length=kwargs.get('max_length'), |
| 2277 | bot_task=bot_task, |
| 2278 | image_base_size=self.config.image_base_size, |
| 2279 | sequence_template=self.generation_config.sequence_template, |
| 2280 | cfg_factor=cfg_factor[mode], |
| 2281 | drop_think=drop_think, |
| 2282 | ) |
| 2283 | output, sections = out['output'], out['sections'] |
| 2284 | |
| 2285 | # 4. Encode conditional images |
| 2286 | if batch_cond_image_info is not None and len(batch_cond_image_info[0]) > 0: |
| 2287 | cond_vae_images, cond_timestep, cond_vit_images = self._encode_cond_image( |
| 2288 | batch_cond_image_info, cfg_factor[mode] |
| 2289 | ) |
| 2290 | # Pack vit kwargs. Siglip2-so requires spatial_shapes and attention_mask for inference. |
| 2291 | vit_kwargs = {"spatial_shapes": [], "attention_mask": []} |
| 2292 | for cond_image_info in batch_cond_image_info: |
| 2293 | vit_kwargs["spatial_shapes"].append( |
| 2294 | torch.stack([item.vision_encoder_kwargs["spatial_shapes"] for item in cond_image_info])) |
| 2295 | vit_kwargs["attention_mask"].append( |
| 2296 | torch.stack([item.vision_encoder_kwargs["pixel_attention_mask"] for item in cond_image_info])) |
| 2297 | if cfg_factor[mode] > 1: |
| 2298 | vit_kwargs["spatial_shapes"] = vit_kwargs["spatial_shapes"] * cfg_factor[mode] |
| 2299 | vit_kwargs["attention_mask"] = vit_kwargs["attention_mask"] * cfg_factor[mode] |
| 2300 | else: |
| 2301 | cond_vae_images, cond_timestep, cond_vit_images = None, None, None |
| 2302 | vit_kwargs = None |
| 2303 | |
| 2304 | # 5. Build position embeddings |
| 2305 | rope_image_info = self.build_batch_rope_image_info(output, sections) |
| 2306 | if mode == "gen_text": |
| 2307 | seq_len = self.generation_config.max_length |
| 2308 | else: |
| 2309 | seq_len = output.tokens.shape[1] |
| 2310 | cos, sin = build_batch_2d_rope( |
| 2311 | image_infos=rope_image_info, |
| 2312 | seq_len=seq_len, |
| 2313 | n_elem=self.config.attention_head_dim, |
| 2314 | device=device, |
| 2315 | base=self.config.rope_theta, |
| 2316 | ) |
| 2317 | |
| 2318 | # 6. Build kv cache |
| 2319 | if bot_task == "img_ratio": |
| 2320 | max_new_tokens = 1 |
| 2321 | if mode == "gen_image": |
| 2322 | # Image generation will not extend sequence length, using token length as max_cache_len is enough. |
| 2323 | max_cache_len = output.tokens.shape[1] |
| 2324 | else: |
| 2325 | max_cache_len = output.tokens.shape[1] + default(max_new_tokens, self.generation_config.max_length) |
| 2326 | cache = HunyuanStaticCache( |
| 2327 | config=self.config, |
| 2328 | batch_size=batch_size * cfg_factor[mode], |
| 2329 | max_cache_len=max_cache_len, |
| 2330 | dtype=torch.bfloat16, |
| 2331 | dynamic=mode == "gen_text", |
| 2332 | ) |
| 2333 | |
| 2334 | # 7. Build position ids |
| 2335 | batch_input_pos = torch.arange( |
| 2336 | 0, output.tokens.shape[1], dtype=torch.long, device=device)[None].expand( |
| 2337 | batch_size * cfg_factor[mode], -1) # use expand to share indices to save memory |
| 2338 | |
| 2339 | # 8. Build model input kwargs |
| 2340 | tkw = self._tkwrapper |
| 2341 | if image_size == "auto": |
| 2342 | extra_auto_stops = [tkw.special_token_map[f"<img_ratio_{i}>"] for i in range(33)] |
| 2343 | else: |
| 2344 | extra_auto_stops = [tkw.boi_token_id] |
| 2345 | stop_token_id = dict( |
| 2346 | auto=[tkw.eos_token_id] + extra_auto_stops, |
| 2347 | image=[tkw.eos_token_id], |
| 2348 | recaption=[tkw.end_recaption_token_id, tkw.end_answer_token_id, tkw.eos_token_id], |
| 2349 | think=[tkw.end_recaption_token_id, tkw.end_answer_token_id, tkw.eos_token_id], |
| 2350 | img_ratio=extra_auto_stops, |
| 2351 | ) |
| 2352 | model_input_kwargs = dict( |
| 2353 | input_ids=output.tokens.to(device), |
| 2354 | position_ids=batch_input_pos, |
| 2355 | past_key_values=cache, |
| 2356 | custom_pos_emb=(cos, sin), |
| 2357 | mode=mode, |
| 2358 | image_mask=to_device(output.gen_image_mask, device), |
| 2359 | gen_timestep_scatter_index=to_device(output.gen_timestep_scatter_index, device), |
| 2360 | cond_vae_images=to_device(cond_vae_images, device), |
| 2361 | cond_timestep=to_device(cond_timestep, device), |
| 2362 | cond_vae_image_mask=to_device(output.cond_vae_image_mask, device), |
| 2363 | cond_vit_images=to_device(cond_vit_images, device), |
| 2364 | cond_vit_image_mask=to_device(output.cond_vit_image_mask, device), |
| 2365 | vit_kwargs={ |
| 2366 | k: to_device(v, self.device) for k, v in vit_kwargs.items() |
| 2367 | } if vit_kwargs is not None else None, |
| 2368 | cond_timestep_scatter_index=to_device(output.cond_timestep_scatter_index, device), |
| 2369 | # for inner usage |
| 2370 | tokenizer_output=output, |
| 2371 | batch_gen_image_info=batch_gen_image_info, |
| 2372 | generator=generator, |
| 2373 | # generation config |
| 2374 | eos_token_id=stop_token_id[bot_task], |
| 2375 | max_new_tokens=max_new_tokens, |
| 2376 | ) |
| 2377 | |
| 2378 | return model_input_kwargs |
| 2379 | |
| 2380 | def _prepare_attention_mask_for_generation( |
| 2381 | self, |
| 2382 | inputs_tensor: torch.Tensor, |
| 2383 | generation_config: GenerationConfig, |
| 2384 | model_kwargs: Dict[str, Any], |
| 2385 | ) -> torch.Tensor: |
| 2386 | # create `4d` bool attention mask (b, 1, seqlen, seqlen) using this implementation to bypass the 2d requirement |
| 2387 | # in the `transformers.generation_utils.GenerationMixin.generate`. |
| 2388 | # This implementation can handle sequences with text and image modalities, where text tokens use causal |
| 2389 | # attention and image tokens use full attention. |
| 2390 | bsz, seq_len = inputs_tensor.shape |
| 2391 | tokenizer_output = model_kwargs["tokenizer_output"] |
| 2392 | batch_image_slices = [ |
| 2393 | tokenizer_output.joint_image_slices[i] + tokenizer_output.gen_image_slices[i] |
| 2394 | for i in range(bsz) |
| 2395 | ] |
| 2396 | attention_mask = torch.ones(seq_len, seq_len, dtype=torch.bool).tril(diagonal=0).repeat(bsz, 1, 1) |
| 2397 | for i in range(bsz): |
| 2398 | for j, image_slice in enumerate(batch_image_slices[i]): |
| 2399 | attention_mask[i, image_slice, image_slice] = True |
| 2400 | attention_mask = attention_mask.unsqueeze(1) |
| 2401 | return attention_mask |
| 2402 | |
| 2403 | def prepare_inputs_for_generation( |
| 2404 | self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, |
| 2405 | tokenizer_output=None, batch_gen_image_info=None, generator=None, **kwargs |
| 2406 | ): |
| 2407 | # if `inputs_embeds` are passed, we only want to use them in the 1st generation step |
| 2408 | if inputs_embeds is not None and past_key_values is None: |
| 2409 | model_inputs = {"inputs_embeds": inputs_embeds} |
| 2410 | else: |
| 2411 | if input_ids.shape[1] != kwargs["position_ids"].shape[1]: # in decode steps |
| 2412 | input_ids = torch.gather(input_ids, dim=1, index=kwargs["position_ids"]) |
| 2413 | model_inputs = {"input_ids": input_ids} |
| 2414 | |
| 2415 | model_inputs.update( |
| 2416 | { |
| 2417 | "attention_mask": attention_mask, |
| 2418 | "position_ids": kwargs["position_ids"], |
| 2419 | "past_key_values": past_key_values, |
| 2420 | "use_cache": kwargs.get("use_cache"), |
| 2421 | "custom_pos_emb": kwargs["custom_pos_emb"], |
| 2422 | "mode": kwargs["mode"], |
| 2423 | "images": kwargs.get("images"), |
| 2424 | "image_mask": kwargs.get("image_mask"), |
| 2425 | "timestep": kwargs.get("timestep"), |
| 2426 | "gen_timestep_scatter_index": kwargs.get("gen_timestep_scatter_index"), |
| 2427 | "cond_vae_images": kwargs.get("cond_vae_images"), |
| 2428 | "cond_timestep": kwargs.get("cond_timestep"), |
| 2429 | "cond_vae_image_mask": kwargs.get("cond_vae_image_mask"), |
| 2430 | "cond_vit_images": kwargs.get("cond_vit_images"), |
| 2431 | "cond_vit_image_mask": kwargs.get("cond_vit_image_mask"), |
| 2432 | "vit_kwargs": kwargs.get("vit_kwargs"), |
| 2433 | "cond_timestep_scatter_index": kwargs.get("cond_timestep_scatter_index"), |
| 2434 | } |
| 2435 | ) |
| 2436 | return model_inputs |
| 2437 | |
| 2438 | def _update_model_kwargs_for_generation( |
| 2439 | self, |
| 2440 | outputs: ModelOutput, |
| 2441 | model_kwargs: Dict[str, Any], |
| 2442 | is_encoder_decoder: bool = False, |
| 2443 | num_new_tokens: int = 1, |
| 2444 | ) -> Dict[str, Any]: |
| 2445 | mode = model_kwargs["mode"] |
| 2446 | |
| 2447 | updated_model_kwargs = { |
| 2448 | "mode": mode, |
| 2449 | "custom_pos_emb": model_kwargs["custom_pos_emb"], |
| 2450 | } |
| 2451 | |
| 2452 | # update past_key_values keeping its naming used in model code |
| 2453 | for possible_cache_name in ALL_CACHE_NAMES: |
| 2454 | if possible_cache_name in outputs: |
| 2455 | # TODO (joao): remove output/input mismatch when these old models (xlnet, reformer) are deprecated |
| 2456 | if possible_cache_name in ("past_buckets_states", "mems"): |
| 2457 | cache_name = "past_key_values" |
| 2458 | else: |
| 2459 | cache_name = possible_cache_name |
| 2460 | updated_model_kwargs[cache_name] = getattr(outputs, possible_cache_name) |
| 2461 | break |
| 2462 | |
| 2463 | if "tokenizer_output" in model_kwargs: |
| 2464 | if mode == "gen_text": |
| 2465 | # When enable batching, we use right padding, which requires a real_pos to index the valid |
| 2466 | # end position of the sequence. If tokenizer_output in model_kwargs, it means we are in the |
| 2467 | # prefill step of generation. |
| 2468 | real_pos = to_device(model_kwargs["tokenizer_output"].real_pos, self.device) |
| 2469 | updated_model_kwargs["position_ids"] = real_pos |
| 2470 | else: |
| 2471 | # position ids |
| 2472 | image_mask = model_kwargs["image_mask"] |
| 2473 | bsz, seq_len = image_mask.shape |
| 2474 | index = torch.arange(seq_len, device=image_mask.device).unsqueeze(0).repeat(bsz, 1) |
| 2475 | position_ids = index.masked_select(image_mask.bool()).reshape(bsz, -1) |
| 2476 | timestep_position_ids = \ |
| 2477 | index[torch.arange(bsz), model_kwargs["gen_timestep_scatter_index"][:, -1]].unsqueeze(-1) |
| 2478 | updated_model_kwargs["position_ids"] = torch.cat([timestep_position_ids, position_ids], dim=1) |
| 2479 | |
| 2480 | # attention mask |
| 2481 | mask_list = [] |
| 2482 | for attention_mask_i, position_ids_i in zip( |
| 2483 | model_kwargs["attention_mask"], updated_model_kwargs["position_ids"]): |
| 2484 | mask_list.append(torch.index_select(attention_mask_i, dim=1, index=position_ids_i.reshape(-1))) |
| 2485 | attention_mask = torch.stack(mask_list, dim=0) |
| 2486 | updated_model_kwargs["attention_mask"] = attention_mask |
| 2487 | updated_model_kwargs["gen_timestep_scatter_index"] = model_kwargs["gen_timestep_scatter_index"] |
| 2488 | |
| 2489 | else: |
| 2490 | if mode == "gen_text": |
| 2491 | # Now we are in the decode steps. |
| 2492 | updated_model_kwargs["position_ids"] = model_kwargs["position_ids"] + 1 |
| 2493 | else: |
| 2494 | updated_model_kwargs["position_ids"] = model_kwargs["position_ids"] |
| 2495 | updated_model_kwargs["attention_mask"] = model_kwargs["attention_mask"] |
| 2496 | updated_model_kwargs["gen_timestep_scatter_index"] = model_kwargs["gen_timestep_scatter_index"] |
| 2497 | |
| 2498 | return updated_model_kwargs |
| 2499 | |
| 2500 | def _generate( |
| 2501 | self, |
| 2502 | inputs: Optional[torch.Tensor] = None, |
| 2503 | generation_config: Optional[GenerationConfig] = None, |
| 2504 | logits_processor: Optional[LogitsProcessorList] = None, |
| 2505 | stopping_criteria: Optional[StoppingCriteriaList] = None, |
| 2506 | prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None, |
| 2507 | synced_gpus: Optional[bool] = None, |
| 2508 | assistant_model: Optional["PreTrainedModel"] = None, |
| 2509 | streamer: Optional["BaseStreamer"] = None, |
| 2510 | negative_prompt_ids: Optional[torch.Tensor] = None, |
| 2511 | negative_prompt_attention_mask: Optional[torch.Tensor] = None, |
| 2512 | use_model_defaults: Optional[bool] = None, |
| 2513 | generator: Optional[List[torch.Generator]] = None, |
| 2514 | verbose: int = 0, |
| 2515 | **kwargs, |
| 2516 | ): |
| 2517 | mode = kwargs.get("mode", "gen_text") |
| 2518 | |
| 2519 | # Log info |
| 2520 | if verbose >= 1: |
| 2521 | output = kwargs["tokenizer_output"] |
| 2522 | context = self._tkwrapper.tokenizer.decode(output.tokens[0], skip_special_tokens=False) |
| 2523 | # Replace <img><img>...<img> with [<img>]{number} |
| 2524 | context = re.sub(r"(<img>)+", lambda m: f"[<img>]{{{len(m.group(0)) // 5}}}", context) |
| 2525 | info_list = [ |
| 2526 | ("token shape", output.tokens.shape), |
| 2527 | ("context[0]", context), |
| 2528 | ] |
| 2529 | gen_config = default(generation_config, self.generation_config) |
| 2530 | if mode == "gen_image": |
| 2531 | if generator is not None: |
| 2532 | info_list.extend([ |
| 2533 | ("seed", [g.initial_seed() for g in generator]), |
| 2534 | ]) |
| 2535 | info_list.extend([ |
| 2536 | ("image_size", [f"{info.image_height}x{info.image_width}" for info in kwargs["batch_gen_image_info"]]), |
| 2537 | ("infer_steps", kwargs.get("diff_infer_steps", gen_config.diff_infer_steps)), |
| 2538 | ("guidance_scale", kwargs.get("diff_guidance_scale", gen_config.diff_guidance_scale)), |
| 2539 | ("flow_shift", kwargs.get("flow_shift", gen_config.flow_shift)), |
| 2540 | ]) |
| 2541 | else: |
| 2542 | info_list.extend([ |
| 2543 | ("do_sample", kwargs.get("do_sample", gen_config.do_sample)), |
| 2544 | ("max_new_tokens", kwargs.get("max_new_tokens", gen_config.max_new_tokens)), |
| 2545 | ("top_k", kwargs.get("top_k", gen_config.top_k)), |
| 2546 | ("top_p", kwargs.get("top_p", gen_config.top_p)), |
| 2547 | ("temperature", kwargs.get("temperature", gen_config.temperature)), |
| 2548 | ("repetition_penalty", kwargs.get("repetition_penalty", gen_config.repetition_penalty)), |
| 2549 | ]) |
| 2550 | max_key_len = max(len(k) for k, _ in info_list) |
| 2551 | info_str = "=" * 50 + \ |
| 2552 | "\nModel input info:\n" + \ |
| 2553 | "\n".join([f" {k.rjust(max_key_len)}: {v}" for k, v in info_list]) + \ |
| 2554 | "\n--------------------------------------------------" |
| 2555 | print(info_str) |
| 2556 | |
| 2557 | if mode == "gen_text": |
| 2558 | with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): |
| 2559 | return super().generate( |
| 2560 | inputs, |
| 2561 | generation_config, |
| 2562 | logits_processor, |
| 2563 | stopping_criteria, |
| 2564 | prefix_allowed_tokens_fn, |
| 2565 | synced_gpus, |
| 2566 | assistant_model, |
| 2567 | streamer, |
| 2568 | negative_prompt_ids, |
| 2569 | negative_prompt_attention_mask, |
| 2570 | use_model_defaults, |
| 2571 | **kwargs, |
| 2572 | ) |
| 2573 | |
| 2574 | elif mode == "gen_image": |
| 2575 | batch_gen_image_info: List[ImageInfo] = kwargs.get("batch_gen_image_info") |
| 2576 | if batch_gen_image_info is None: |
| 2577 | raise ValueError("`batch_gen_image_info` should be provided when `mode` is `gen_image`.") |
| 2578 | |
| 2579 | results = self.pipeline( |
| 2580 | batch_size=len(batch_gen_image_info), |
| 2581 | image_size=[batch_gen_image_info[0].image_height, batch_gen_image_info[0].image_width], |
| 2582 | num_inference_steps=kwargs.get("diff_infer_steps", self.generation_config.diff_infer_steps), |
| 2583 | guidance_scale=kwargs.get("diff_guidance_scale", self.generation_config.diff_guidance_scale), |
| 2584 | generator=generator, |
| 2585 | model_kwargs=kwargs, |
| 2586 | ) |
| 2587 | samples = results[0] |
| 2588 | return samples |
| 2589 | |
| 2590 | else: |
| 2591 | raise ValueError(f"Unknown mode {mode}, only `gen_text` and `gen_image` are supported.") |
| 2592 | |
| 2593 | def get_cot_text(self, output: torch.Tensor): |
| 2594 | if output.ndim == 2: |
| 2595 | return [self.get_cot_text(output_i) for output_i in output] |
| 2596 | elif output.ndim == 1: |
| 2597 | if output[-1] == self._tkwrapper.eos_token_id: |
| 2598 | output = output[:-1] |
| 2599 | cot_text = self._tkwrapper.decode(output).split("Assistant: ")[1] |
| 2600 | return cot_text |
| 2601 | else: |
| 2602 | raise ValueError(f"output should be 1D or 2D tensor, but got {output.ndim}D tensor.") |
| 2603 | |
| 2604 | def generate_image( |
| 2605 | self, |
| 2606 | prompt, |
| 2607 | seed=None, |
| 2608 | image_size="auto", |
| 2609 | use_system_prompt=None, |
| 2610 | system_prompt=None, |
| 2611 | bot_task=None, |
| 2612 | stream=False, |
| 2613 | **kwargs, |
| 2614 | ): |
| 2615 | max_new_tokens = kwargs.pop("max_new_tokens", 8192) |
| 2616 | verbose = kwargs.pop("verbose", 0) |
| 2617 | |
| 2618 | if stream: |
| 2619 | from transformers import TextStreamer |
| 2620 | streamer = TextStreamer(self._tkwrapper.tokenizer, skip_prompt=True, skip_special_tokens=False) |
| 2621 | kwargs["streamer"] = streamer |
| 2622 | |
| 2623 | use_system_prompt = default(use_system_prompt, self.generation_config.use_system_prompt) |
| 2624 | bot_task = default(bot_task, self.generation_config.bot_task) |
| 2625 | system_prompt = get_system_prompt(use_system_prompt, bot_task, system_prompt) |
| 2626 | |
| 2627 | if bot_task in ["think", "recaption"]: |
| 2628 | # Cot |
| 2629 | model_inputs = self.prepare_model_inputs( |
| 2630 | prompt=prompt, bot_task=bot_task, system_prompt=system_prompt, max_new_tokens=max_new_tokens) |
| 2631 | print(f"<{bot_task}>", end="", flush=True) |
| 2632 | outputs = self._generate(**model_inputs, **kwargs, verbose=verbose) |
| 2633 | cot_text = self.get_cot_text(outputs[0]) |
| 2634 | # Switch system_prompt to `en_recaption` if drop_think is enabled. |
| 2635 | if self.generation_config.drop_think and system_prompt: |
| 2636 | system_prompt = t2i_system_prompts["en_recaption"][0] |
| 2637 | else: |
| 2638 | cot_text = None |
| 2639 | |
| 2640 | # Image ratio |
| 2641 | if image_size == "auto": |
| 2642 | model_inputs = self.prepare_model_inputs( |
| 2643 | prompt=prompt, cot_text=cot_text, bot_task="img_ratio", system_prompt=system_prompt, seed=seed) |
| 2644 | outputs = self._generate(**model_inputs, **kwargs, verbose=verbose) |
| 2645 | ratio_index = outputs[0, -1].item() - self._tkwrapper.ratio_token_offset |
| 2646 | # In some cases, the generated ratio_index is out of range. A valid ratio_index should be in [0, 32]. |
| 2647 | # If ratio_index is out of range, we set it to 16 (i.e., 1:1). |
| 2648 | if ratio_index < 0 or ratio_index >= len(self.image_processor.reso_group): |
| 2649 | ratio_index = 16 |
| 2650 | reso = self.image_processor.reso_group[ratio_index] |
| 2651 | image_size = reso.height, reso.width |
| 2652 | |
| 2653 | # Generate image |
| 2654 | model_inputs = self.prepare_model_inputs( |
| 2655 | prompt=prompt, cot_text=cot_text, system_prompt=system_prompt, mode="gen_image", seed=seed, |
| 2656 | image_size=image_size, |
| 2657 | ) |
| 2658 | outputs = self._generate(**model_inputs, **kwargs, verbose=verbose) |
| 2659 | return outputs[0] |
| 2660 | |