| 1 | from dataclasses import dataclass, replace |
| 2 | |
| 3 | import torch |
| 4 | |
| 5 | from ltx_core.guidance.perturbations import BatchedPerturbationConfig, PerturbationType |
| 6 | from ltx_core.model.transformer.adaln import adaln_embedding_coefficient |
| 7 | from ltx_core.model.transformer.attention import Attention, AttentionCallable, AttentionFunction |
| 8 | from ltx_core.model.transformer.feed_forward import FeedForward |
| 9 | from ltx_core.model.transformer.rope import LTXRopeType |
| 10 | from ltx_core.model.transformer.transformer_args import TransformerArgs |
| 11 | from ltx_core.utils import rms_norm |
| 12 | |
| 13 | |
| 14 | @dataclass |
| 15 | class TransformerConfig: |
| 16 | dim: int |
| 17 | heads: int |
| 18 | d_head: int |
| 19 | context_dim: int |
| 20 | apply_gated_attention: bool = False |
| 21 | cross_attention_adaln: bool = False |
| 22 | |
| 23 | |
| 24 | class BasicAVTransformerBlock(torch.nn.Module): |
| 25 | def __init__( |
| 26 | self, |
| 27 | idx: int, |
| 28 | num_layers: int, |
| 29 | video: TransformerConfig | None = None, |
| 30 | audio: TransformerConfig | None = None, |
| 31 | rope_type: LTXRopeType = LTXRopeType.INTERLEAVED, |
| 32 | norm_eps: float = 1e-6, |
| 33 | attention_function: AttentionFunction | AttentionCallable = AttentionFunction.DEFAULT, |
| 34 | ): |
| 35 | super().__init__() |
| 36 | |
| 37 | self.idx = idx |
| 38 | self.num_layers = num_layers |
| 39 | if video is not None: |
| 40 | self.attn1 = Attention( |
| 41 | query_dim=video.dim, |
| 42 | heads=video.heads, |
| 43 | dim_head=video.d_head, |
| 44 | context_dim=None, |
| 45 | rope_type=rope_type, |
| 46 | norm_eps=norm_eps, |
| 47 | attention_function=attention_function, |
| 48 | apply_gated_attention=video.apply_gated_attention, |
| 49 | ) |
| 50 | self.attn2 = Attention( |
| 51 | query_dim=video.dim, |
| 52 | context_dim=video.context_dim, |
| 53 | heads=video.heads, |
| 54 | dim_head=video.d_head, |
| 55 | rope_type=rope_type, |
| 56 | norm_eps=norm_eps, |
| 57 | attention_function=attention_function, |
| 58 | apply_gated_attention=video.apply_gated_attention, |
| 59 | ) |
| 60 | self.ff = FeedForward(video.dim, dim_out=video.dim) |
| 61 | video_sst_size = adaln_embedding_coefficient(video.cross_attention_adaln) |
| 62 | self.scale_shift_table = torch.nn.Parameter(torch.empty(video_sst_size, video.dim)) |
| 63 | |
| 64 | if audio is not None: |
| 65 | self.audio_attn1 = Attention( |
| 66 | query_dim=audio.dim, |
| 67 | heads=audio.heads, |
| 68 | dim_head=audio.d_head, |
| 69 | context_dim=None, |
| 70 | rope_type=rope_type, |
| 71 | norm_eps=norm_eps, |
| 72 | attention_function=attention_function, |
| 73 | apply_gated_attention=audio.apply_gated_attention, |
| 74 | ) |
| 75 | self.audio_attn2 = Attention( |
| 76 | query_dim=audio.dim, |
| 77 | context_dim=audio.context_dim, |
| 78 | heads=audio.heads, |
| 79 | dim_head=audio.d_head, |
| 80 | rope_type=rope_type, |
| 81 | norm_eps=norm_eps, |
| 82 | attention_function=attention_function, |
| 83 | apply_gated_attention=audio.apply_gated_attention, |
| 84 | ) |
| 85 | self.audio_ff = FeedForward(audio.dim, dim_out=audio.dim) |
| 86 | audio_sst_size = adaln_embedding_coefficient(audio.cross_attention_adaln) |
| 87 | self.audio_scale_shift_table = torch.nn.Parameter(torch.empty(audio_sst_size, audio.dim)) |
| 88 | |
| 89 | if audio is not None and video is not None: |
| 90 | # Q: Video, K,V: Audio |
| 91 | self.audio_to_video_attn = Attention( |
| 92 | query_dim=video.dim, |
| 93 | context_dim=audio.dim, |
| 94 | heads=audio.heads, |
| 95 | dim_head=audio.d_head, |
| 96 | rope_type=rope_type, |
| 97 | norm_eps=norm_eps, |
| 98 | attention_function=attention_function, |
| 99 | apply_gated_attention=video.apply_gated_attention, |
| 100 | ) |
| 101 | |
| 102 | # Q: Audio, K,V: Video |
| 103 | self.video_to_audio_attn = Attention( |
| 104 | query_dim=audio.dim, |
| 105 | context_dim=video.dim, |
| 106 | heads=audio.heads, |
| 107 | dim_head=audio.d_head, |
| 108 | rope_type=rope_type, |
| 109 | norm_eps=norm_eps, |
| 110 | attention_function=attention_function, |
| 111 | apply_gated_attention=audio.apply_gated_attention, |
| 112 | ) |
| 113 | |
| 114 | self.scale_shift_table_a2v_ca_audio = torch.nn.Parameter(torch.empty(5, audio.dim)) |
| 115 | self.scale_shift_table_a2v_ca_video = torch.nn.Parameter(torch.empty(5, video.dim)) |
| 116 | |
| 117 | self.cross_attention_adaln = (video is not None and video.cross_attention_adaln) or ( |
| 118 | audio is not None and audio.cross_attention_adaln |
| 119 | ) |
| 120 | |
| 121 | if self.cross_attention_adaln and video is not None: |
| 122 | self.prompt_scale_shift_table = torch.nn.Parameter(torch.empty(2, video.dim)) |
| 123 | if self.cross_attention_adaln and audio is not None: |
| 124 | self.audio_prompt_scale_shift_table = torch.nn.Parameter(torch.empty(2, audio.dim)) |
| 125 | |
| 126 | self.norm_eps = norm_eps |
| 127 | |
| 128 | def get_ada_values( |
| 129 | self, scale_shift_table: torch.Tensor, batch_size: int, timestep: torch.Tensor, indices: slice |
| 130 | ) -> tuple[torch.Tensor, ...]: |
| 131 | num_ada_params = scale_shift_table.shape[0] |
| 132 | |
| 133 | ada_values = ( |
| 134 | scale_shift_table[indices].unsqueeze(0).unsqueeze(0).to(device=timestep.device, dtype=timestep.dtype) |
| 135 | + timestep.reshape(batch_size, timestep.shape[1], num_ada_params, -1)[:, :, indices, :] |
| 136 | ).unbind(dim=2) |
| 137 | return ada_values |
| 138 | |
| 139 | def get_av_ca_ada_values( |
| 140 | self, |
| 141 | scale_shift_table: torch.Tensor, |
| 142 | batch_size: int, |
| 143 | scale_shift_timestep: torch.Tensor, |
| 144 | gate_timestep: torch.Tensor, |
| 145 | scale_shift_indices: slice, |
| 146 | num_scale_shift_values: int = 4, |
| 147 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: |
| 148 | scale_shift_ada_values = self.get_ada_values( |
| 149 | scale_shift_table[:num_scale_shift_values, :], batch_size, scale_shift_timestep, scale_shift_indices |
| 150 | ) |
| 151 | gate_ada_values = self.get_ada_values( |
| 152 | scale_shift_table[num_scale_shift_values:, :], batch_size, gate_timestep, slice(None, None) |
| 153 | ) |
| 154 | |
| 155 | scale, shift = (t.squeeze(2) for t in scale_shift_ada_values) |
| 156 | (gate,) = (t.squeeze(2) for t in gate_ada_values) |
| 157 | |
| 158 | return scale, shift, gate |
| 159 | |
| 160 | def _apply_text_cross_attention( |
| 161 | self, |
| 162 | x: torch.Tensor, |
| 163 | context: torch.Tensor, |
| 164 | attn: AttentionCallable, |
| 165 | scale_shift_table: torch.Tensor, |
| 166 | prompt_scale_shift_table: torch.Tensor | None, |
| 167 | timestep: torch.Tensor, |
| 168 | prompt_timestep: torch.Tensor | None, |
| 169 | context_mask: torch.Tensor | None, |
| 170 | cross_attention_adaln: bool = False, |
| 171 | ) -> torch.Tensor: |
| 172 | """Apply text cross-attention, with optional AdaLN modulation.""" |
| 173 | if cross_attention_adaln: |
| 174 | shift_q, scale_q, gate = self.get_ada_values(scale_shift_table, x.shape[0], timestep, slice(6, 9)) |
| 175 | return apply_cross_attention_adaln( |
| 176 | x, |
| 177 | context, |
| 178 | attn, |
| 179 | shift_q, |
| 180 | scale_q, |
| 181 | gate, |
| 182 | prompt_scale_shift_table, |
| 183 | prompt_timestep, |
| 184 | context_mask, |
| 185 | self.norm_eps, |
| 186 | ) |
| 187 | return attn(rms_norm(x, eps=self.norm_eps), context=context, mask=context_mask) |
| 188 | |
| 189 | def forward( # noqa: PLR0915 |
| 190 | self, |
| 191 | video: TransformerArgs | None, |
| 192 | audio: TransformerArgs | None, |
| 193 | perturbations: BatchedPerturbationConfig | None = None, |
| 194 | ) -> tuple[TransformerArgs | None, TransformerArgs | None]: |
| 195 | if video is None and audio is None: |
| 196 | raise ValueError("At least one of video or audio must be provided") |
| 197 | |
| 198 | batch_size = (video or audio).x.shape[0] |
| 199 | |
| 200 | if perturbations is None: |
| 201 | perturbations = BatchedPerturbationConfig.empty(batch_size) |
| 202 | |
| 203 | vx = video.x if video is not None else None |
| 204 | ax = audio.x if audio is not None else None |
| 205 | |
| 206 | run_vx = video is not None and video.enabled and vx.numel() > 0 |
| 207 | run_ax = audio is not None and audio.enabled and ax.numel() > 0 |
| 208 | |
| 209 | run_a2v = run_vx and (audio is not None and ax.numel() > 0) |
| 210 | run_v2a = run_ax and (video is not None and vx.numel() > 0) |
| 211 | |
| 212 | if run_vx: |
| 213 | vshift_msa, vscale_msa, vgate_msa = self.get_ada_values( |
| 214 | self.scale_shift_table, vx.shape[0], video.timesteps, slice(0, 3) |
| 215 | ) |
| 216 | norm_vx = rms_norm(vx, eps=self.norm_eps) * (1 + vscale_msa) + vshift_msa |
| 217 | del vshift_msa, vscale_msa |
| 218 | |
| 219 | all_perturbed = perturbations.all_in_batch(PerturbationType.SKIP_VIDEO_SELF_ATTN, self.idx) |
| 220 | none_perturbed = not perturbations.any_in_batch(PerturbationType.SKIP_VIDEO_SELF_ATTN, self.idx) |
| 221 | v_mask = ( |
| 222 | perturbations.mask_like(PerturbationType.SKIP_VIDEO_SELF_ATTN, self.idx, vx) |
| 223 | if not all_perturbed and not none_perturbed |
| 224 | else None |
| 225 | ) |
| 226 | vx = ( |
| 227 | vx |
| 228 | + self.attn1( |
| 229 | norm_vx, |
| 230 | pe=video.positional_embeddings, |
| 231 | mask=video.self_attention_mask, |
| 232 | perturbation_mask=v_mask, |
| 233 | all_perturbed=all_perturbed, |
| 234 | ) |
| 235 | * vgate_msa |
| 236 | ) |
| 237 | del vgate_msa, norm_vx, v_mask |
| 238 | vx = vx + self._apply_text_cross_attention( |
| 239 | vx, |
| 240 | video.context, |
| 241 | self.attn2, |
| 242 | self.scale_shift_table, |
| 243 | getattr(self, "prompt_scale_shift_table", None), |
| 244 | video.timesteps, |
| 245 | video.prompt_timestep, |
| 246 | video.context_mask, |
| 247 | cross_attention_adaln=self.cross_attention_adaln, |
| 248 | ) |
| 249 | |
| 250 | if run_ax: |
| 251 | ashift_msa, ascale_msa, agate_msa = self.get_ada_values( |
| 252 | self.audio_scale_shift_table, ax.shape[0], audio.timesteps, slice(0, 3) |
| 253 | ) |
| 254 | |
| 255 | norm_ax = rms_norm(ax, eps=self.norm_eps) * (1 + ascale_msa) + ashift_msa |
| 256 | del ashift_msa, ascale_msa |
| 257 | all_perturbed = perturbations.all_in_batch(PerturbationType.SKIP_AUDIO_SELF_ATTN, self.idx) |
| 258 | none_perturbed = not perturbations.any_in_batch(PerturbationType.SKIP_AUDIO_SELF_ATTN, self.idx) |
| 259 | a_mask = ( |
| 260 | perturbations.mask_like(PerturbationType.SKIP_AUDIO_SELF_ATTN, self.idx, ax) |
| 261 | if not all_perturbed and not none_perturbed |
| 262 | else None |
| 263 | ) |
| 264 | audio_self_attention_mask = audio.self_attention_mask |
| 265 | if self.idx >= int(self.num_layers * 0.7): |
| 266 | audio_self_attention_mask = audio.late_self_attention_mask |
| 267 | ax = ( |
| 268 | ax |
| 269 | + self.audio_attn1( |
| 270 | norm_ax, |
| 271 | pe=audio.positional_embeddings, |
| 272 | mask=audio_self_attention_mask, |
| 273 | perturbation_mask=a_mask, |
| 274 | all_perturbed=all_perturbed, |
| 275 | ) |
| 276 | * agate_msa |
| 277 | ) |
| 278 | del agate_msa, norm_ax, a_mask |
| 279 | ax = ax + self._apply_text_cross_attention( |
| 280 | ax, |
| 281 | audio.context, |
| 282 | self.audio_attn2, |
| 283 | self.audio_scale_shift_table, |
| 284 | getattr(self, "audio_prompt_scale_shift_table", None), |
| 285 | audio.timesteps, |
| 286 | audio.prompt_timestep, |
| 287 | audio.context_mask, |
| 288 | cross_attention_adaln=self.cross_attention_adaln, |
| 289 | ) |
| 290 | |
| 291 | # Audio - Video cross attention. |
| 292 | if run_a2v or run_v2a: |
| 293 | vx_norm3 = rms_norm(vx, eps=self.norm_eps) |
| 294 | ax_norm3 = rms_norm(ax, eps=self.norm_eps) |
| 295 | |
| 296 | if run_a2v and not perturbations.all_in_batch(PerturbationType.SKIP_A2V_CROSS_ATTN, self.idx): |
| 297 | scale_ca_video_a2v, shift_ca_video_a2v, gate_out_a2v = self.get_av_ca_ada_values( |
| 298 | self.scale_shift_table_a2v_ca_video, |
| 299 | vx.shape[0], |
| 300 | video.cross_scale_shift_timestep, |
| 301 | video.cross_gate_timestep, |
| 302 | slice(0, 2), |
| 303 | ) |
| 304 | vx_scaled = vx_norm3 * (1 + scale_ca_video_a2v) + shift_ca_video_a2v |
| 305 | del scale_ca_video_a2v, shift_ca_video_a2v |
| 306 | |
| 307 | scale_ca_audio_a2v, shift_ca_audio_a2v, _ = self.get_av_ca_ada_values( |
| 308 | self.scale_shift_table_a2v_ca_audio, |
| 309 | ax.shape[0], |
| 310 | audio.cross_scale_shift_timestep, |
| 311 | audio.cross_gate_timestep, |
| 312 | slice(0, 2), |
| 313 | ) |
| 314 | ax_scaled = ax_norm3 * (1 + scale_ca_audio_a2v) + shift_ca_audio_a2v |
| 315 | del scale_ca_audio_a2v, shift_ca_audio_a2v |
| 316 | a2v_mask = perturbations.mask_like(PerturbationType.SKIP_A2V_CROSS_ATTN, self.idx, vx) |
| 317 | cross_attention_mask = video.cross_attention_mask |
| 318 | cross_output_mask = video.cross_output_mask |
| 319 | if self.idx >= int(self.num_layers * 0.7): |
| 320 | if video.late_cross_attention_mask is not None: |
| 321 | cross_attention_mask = video.late_cross_attention_mask |
| 322 | if video.late_cross_output_mask is not None: |
| 323 | cross_output_mask = video.late_cross_output_mask |
| 324 | cross_output_mask = cross_output_mask if cross_output_mask is not None else 1.0 |
| 325 | vx = vx + ( |
| 326 | self.audio_to_video_attn( |
| 327 | vx_scaled, |
| 328 | context=ax_scaled, |
| 329 | mask=cross_attention_mask, |
| 330 | pe=video.cross_positional_embeddings, |
| 331 | k_pe=audio.cross_positional_embeddings, |
| 332 | ) |
| 333 | * gate_out_a2v |
| 334 | * a2v_mask |
| 335 | * cross_output_mask |
| 336 | ) |
| 337 | del gate_out_a2v, a2v_mask, vx_scaled, ax_scaled, cross_output_mask |
| 338 | |
| 339 | if run_v2a and not perturbations.all_in_batch(PerturbationType.SKIP_V2A_CROSS_ATTN, self.idx): |
| 340 | scale_ca_audio_v2a, shift_ca_audio_v2a, gate_out_v2a = self.get_av_ca_ada_values( |
| 341 | self.scale_shift_table_a2v_ca_audio, |
| 342 | ax.shape[0], |
| 343 | audio.cross_scale_shift_timestep, |
| 344 | audio.cross_gate_timestep, |
| 345 | slice(2, 4), |
| 346 | ) |
| 347 | ax_scaled = ax_norm3 * (1 + scale_ca_audio_v2a) + shift_ca_audio_v2a |
| 348 | del scale_ca_audio_v2a, shift_ca_audio_v2a |
| 349 | scale_ca_video_v2a, shift_ca_video_v2a, _ = self.get_av_ca_ada_values( |
| 350 | self.scale_shift_table_a2v_ca_video, |
| 351 | vx.shape[0], |
| 352 | video.cross_scale_shift_timestep, |
| 353 | video.cross_gate_timestep, |
| 354 | slice(2, 4), |
| 355 | ) |
| 356 | vx_scaled = vx_norm3 * (1 + scale_ca_video_v2a) + shift_ca_video_v2a |
| 357 | del scale_ca_video_v2a, shift_ca_video_v2a |
| 358 | v2a_mask = perturbations.mask_like(PerturbationType.SKIP_V2A_CROSS_ATTN, self.idx, ax) |
| 359 | cross_attention_mask = audio.cross_attention_mask |
| 360 | cross_output_mask = audio.cross_output_mask |
| 361 | if self.idx >= int(self.num_layers * 0.7): |
| 362 | if audio.late_cross_attention_mask is not None: |
| 363 | cross_attention_mask = audio.late_cross_attention_mask |
| 364 | if audio.late_cross_output_mask is not None: |
| 365 | cross_output_mask = audio.late_cross_output_mask |
| 366 | cross_output_mask = cross_output_mask if cross_output_mask is not None else 1.0 |
| 367 | v2a_update = ( |
| 368 | self.video_to_audio_attn( |
| 369 | ax_scaled, |
| 370 | context=vx_scaled, |
| 371 | mask=cross_attention_mask, |
| 372 | pe=audio.cross_positional_embeddings, |
| 373 | k_pe=video.cross_positional_embeddings, |
| 374 | ) |
| 375 | * gate_out_v2a |
| 376 | * v2a_mask |
| 377 | * cross_output_mask |
| 378 | ) |
| 379 | v2a_grad_scale = float(getattr(audio, "v2a_grad_scale", 1.0)) |
| 380 | if v2a_grad_scale != 1.0 and torch.is_grad_enabled(): |
| 381 | v2a_update = v2a_update.detach() + v2a_grad_scale * (v2a_update - v2a_update.detach()) |
| 382 | ax = ax + v2a_update |
| 383 | del gate_out_v2a, v2a_mask, ax_scaled, vx_scaled, cross_output_mask, v2a_update |
| 384 | |
| 385 | del vx_norm3, ax_norm3 |
| 386 | |
| 387 | if run_vx: |
| 388 | vshift_mlp, vscale_mlp, vgate_mlp = self.get_ada_values( |
| 389 | self.scale_shift_table, vx.shape[0], video.timesteps, slice(3, 6) |
| 390 | ) |
| 391 | vx_scaled = rms_norm(vx, eps=self.norm_eps) * (1 + vscale_mlp) + vshift_mlp |
| 392 | vx = vx + self.ff(vx_scaled) * vgate_mlp |
| 393 | |
| 394 | del vshift_mlp, vscale_mlp, vgate_mlp, vx_scaled |
| 395 | |
| 396 | if run_ax: |
| 397 | ashift_mlp, ascale_mlp, agate_mlp = self.get_ada_values( |
| 398 | self.audio_scale_shift_table, ax.shape[0], audio.timesteps, slice(3, 6) |
| 399 | ) |
| 400 | ax_scaled = rms_norm(ax, eps=self.norm_eps) * (1 + ascale_mlp) + ashift_mlp |
| 401 | ax = ax + self.audio_ff(ax_scaled) * agate_mlp |
| 402 | |
| 403 | del ashift_mlp, ascale_mlp, agate_mlp, ax_scaled |
| 404 | |
| 405 | return replace(video, x=vx) if video is not None else None, replace(audio, x=ax) if audio is not None else None |
| 406 | |
| 407 | |
| 408 | def apply_cross_attention_adaln( |
| 409 | x: torch.Tensor, |
| 410 | context: torch.Tensor, |
| 411 | attn: AttentionCallable, |
| 412 | q_shift: torch.Tensor, |
| 413 | q_scale: torch.Tensor, |
| 414 | q_gate: torch.Tensor, |
| 415 | prompt_scale_shift_table: torch.Tensor, |
| 416 | prompt_timestep: torch.Tensor, |
| 417 | context_mask: torch.Tensor | None = None, |
| 418 | norm_eps: float = 1e-6, |
| 419 | ) -> torch.Tensor: |
| 420 | batch_size = x.shape[0] |
| 421 | shift_kv, scale_kv = ( |
| 422 | prompt_scale_shift_table[None, None].to(device=x.device, dtype=x.dtype) |
| 423 | + prompt_timestep.reshape(batch_size, prompt_timestep.shape[1], 2, -1) |
| 424 | ).unbind(dim=2) |
| 425 | attn_input = rms_norm(x, eps=norm_eps) * (1 + q_scale) + q_shift |
| 426 | encoder_hidden_states = context * (1 + scale_kv) + shift_kv |
| 427 | return attn(attn_input, context=encoder_hidden_states, mask=context_mask) * q_gate |
| 428 |