|
4 | 4 | from flax import nnx |
5 | 5 | from flax.traverse_util import unflatten_dict, flatten_dict |
6 | 6 | from maxdiffusion import max_logging |
7 | | -from ..modeling_flax_pytorch_utils import validate_flax_state_dict |
| 7 | +from ..modeling_flax_pytorch_utils import validate_flax_state_dict, rename_key |
8 | 8 | from .ltx2_utils import load_sharded_checkpoint |
9 | 9 | from .ltx2_utils import ( |
10 | 10 | _tuple_str_to_int, |
11 | 11 | LTX_2_0_VIDEO_VAE_RENAME_DICT, |
| 12 | + rename_for_ltx2_transformer, |
| 13 | + get_key_and_value, |
| 14 | + rename_for_ltx2_audio_vae, |
| 15 | + rename_for_ltx2_vocoder, |
12 | 16 | ) |
| 17 | +def load_ltx2_3_checkpoint(pretrained_model_name_or_path: str, subfolder: str, device: str, filename: str): |
| 18 | + """Loads weights from a single safetensors file for LTX-2.3.""" |
| 19 | + from huggingface_hub import hf_hub_download |
| 20 | + from safetensors import safe_open |
| 21 | + from ..modeling_flax_pytorch_utils import torch2jax |
13 | 22 |
|
| 23 | + ckpt_path = hf_hub_download(pretrained_model_name_or_path, subfolder=subfolder, filename=filename) |
| 24 | + tensors = {} |
| 25 | + with safe_open(ckpt_path, framework="pt") as f: |
| 26 | + for k in f.keys(): |
| 27 | + tensors[k] = torch2jax(f.get_tensor(k)) |
| 28 | + return tensorsdef rename_for_ltx2_3_transformer(key): |
| 29 | + """ |
| 30 | + Renames Diffusers LTX-2.3 keys to MaxDiffusion Flax LTX-2.3 keys. |
| 31 | + """ |
| 32 | + key = key.replace("patchify_proj", "proj_in") |
| 33 | + key = key.replace("audio_patchify_proj", "audio_proj_in") |
| 34 | + key = key.replace("norm_final", "norm_out") |
| 35 | + if "adaLN_modulation_1" in key: |
| 36 | + key = key.replace("adaLN_modulation_1", "scale_shift_table") |
| 37 | + |
| 38 | + if "caption_modulator_1" in key: |
| 39 | + key = key.replace("caption_modulator_1", "video_a2v_cross_attn_scale_shift_table") |
| 40 | + if "audio_caption_modulator_1" in key: |
| 41 | + key = key.replace("audio_caption_modulator_1", "audio_a2v_cross_attn_scale_shift_table") |
| 42 | + if "audio_norm_final" in key: |
| 43 | + key = key.replace("audio_norm_final", "audio_norm_out") |
| 44 | + if ("audio_ff" in key or "ff" in key) and "proj" in key: |
| 45 | + key = key.replace(".proj", "") |
| 46 | + if "to_out_0" in key: |
| 47 | + key = key.replace("to_out_0", "to_out") |
| 48 | + |
| 49 | + # Add missing mappings |
| 50 | + key = key.replace("av_ca_video_scale_shift_adaln_single", "av_cross_attn_video_scale_shift") |
| 51 | + key = key.replace("av_ca_a2v_gate_adaln_single", "av_cross_attn_video_a2v_gate") |
| 52 | + key = key.replace("av_ca_audio_scale_shift_adaln_single", "av_cross_attn_audio_scale_shift") |
| 53 | + key = key.replace("av_ca_v2a_gate_adaln_single", "av_cross_attn_audio_v2a_gate") |
| 54 | + key = key.replace("scale_shift_table_a2v_ca_video", "video_a2v_cross_attn_scale_shift_table") |
| 55 | + key = key.replace("scale_shift_table_a2v_ca_audio", "audio_a2v_cross_attn_scale_shift_table") |
| 56 | + |
| 57 | + # LTX-2.3 specific mappings |
| 58 | + # Handle substrings before they are replaced by shorter patterns below |
| 59 | + key = key.replace("audio_prompt_adaln_single", "audio_prompt_adaln") |
| 60 | + key = key.replace("prompt_adaln_single", "prompt_adaln") |
| 61 | + key = key.replace("audio_prompt_scale_shift_table", "audio_scale_shift_table") |
| 62 | + key = key.replace("prompt_scale_shift_table", "scale_shift_table") |
| 63 | + |
| 64 | + if "prompt_adaln" in key: |
| 65 | + key = key.replace("prompt_adaln", "caption_projection") |
| 66 | + if "audio_prompt_adaln" in key: |
| 67 | + key = key.replace("audio_prompt_adaln", "audio_caption_projection") |
| 68 | + if "video_text_proj_in" in key: |
| 69 | + key = key.replace("video_text_proj_in", "feature_extractor.video_linear") |
| 70 | + if "audio_text_proj_in" in key: |
| 71 | + key = key.replace("audio_text_proj_in", "feature_extractor.audio_linear") |
| 72 | + |
| 73 | + key = key.replace("k_norm", "norm_k") |
| 74 | + key = key.replace("q_norm", "norm_q") |
| 75 | + key = key.replace("adaln_single", "time_embed") |
| 76 | + return keydef rename_for_ltx2_3_vocoder(key): |
| 77 | + """Renames Diffusers LTX-2.3 Vocoder keys to MaxDiffusion Flax keys.""" |
| 78 | + key = key.replace("ups.", "upsamplers.") |
| 79 | + key = key.replace("resblocks.", "resblocks_") |
| 80 | + key = key.replace("conv_post", "conv_out") |
| 81 | + key = key.replace("conv_pre", "conv_in") |
| 82 | + key = key.replace("act_post", "act_out") |
| 83 | + |
| 84 | + # LTX-2.3 specific mappings for Vocoder |
| 85 | + if "downsample" in key and "lowpass" not in key: |
| 86 | + key = key.replace("downsample", "downsample.lowpass") |
| 87 | + |
| 88 | + return key |
14 | 89 |
|
15 | 90 |
|
16 | 91 | LTX_2_3_CONNECTORS_KEYS_RENAME_DICT = { |
|
46 | 121 | "audio_embeddings_connector": "audio_connector", |
47 | 122 | } |
48 | 123 |
|
49 | | -def load_connectors_weights( |
| 124 | +def load_and_segregate_ltx2_3_weights(pretrained_model_name_or_path: str, filename: str = "ltx-2.3-22b-dev.safetensors"): |
| 125 | + """Loads the full LTX-2.3 file once and splits it into component-specific dictionaries.""" |
| 126 | + tensors = load_ltx2_3_checkpoint(pretrained_model_name_or_path, "", "cpu", filename=filename) |
| 127 | + |
| 128 | + segregated = { |
| 129 | + "transformer": {}, |
| 130 | + "vae": {}, |
| 131 | + "audio_vae": {}, |
| 132 | + "connectors": {}, |
| 133 | + "vocoder": {}, |
| 134 | + } |
| 135 | + |
| 136 | + for pt_key, tensor in tensors.items(): |
| 137 | + if pt_key.startswith("model.diffusion_model."): |
| 138 | + segregated["transformer"][pt_key.replace("model.diffusion_model.", "")] = tensor |
| 139 | + elif pt_key.startswith("audio_vae."): |
| 140 | + segregated["audio_vae"][pt_key.replace("audio_vae.", "")] = tensor |
| 141 | + elif pt_key.startswith("vae."): |
| 142 | + segregated["vae"][pt_key] = tensor |
| 143 | + elif pt_key.startswith("vocoder."): |
| 144 | + segregated["vocoder"][pt_key.replace("vocoder.", "")] = tensor |
| 145 | + elif any(x in pt_key for x in ["connectors.", "video_embeddings_connector", "audio_embeddings_connector", "text_embedding_projection"]): |
| 146 | + segregated["connectors"][pt_key] = tensor |
| 147 | + |
| 148 | + return segregated |
| 149 | + |
| 150 | + |
| 151 | +def load_transformer_weights_2_3( |
| 152 | + eval_shapes: dict, |
| 153 | + device: str, |
| 154 | + tensors: dict, |
| 155 | + num_layers: int = 48, |
| 156 | + scan_layers: bool = True, |
| 157 | +): |
| 158 | + device = jax.local_devices(backend=device)[0] |
| 159 | + max_logging.log(f"Load and port LTX-2.3 transformer on {device}") |
| 160 | + |
| 161 | + with jax.default_device(device): |
| 162 | + flax_state_dict = {} |
| 163 | + cpu = jax.local_devices(backend="cpu")[0] |
| 164 | + flattened_dict = flatten_dict(eval_shapes) |
| 165 | + |
| 166 | + random_flax_state_dict = {} |
| 167 | + for key in flattened_dict: |
| 168 | + random_flax_state_dict[tuple(str(item) for item in key)] = flattened_dict[key] |
| 169 | + |
| 170 | + for pt_key, tensor in tensors.items(): |
| 171 | + # Keys are already filtered and stripped of "model.diffusion_model." by load_and_segregate |
| 172 | + if pt_key.startswith("audio_embeddings_connector") or pt_key.startswith("video_embeddings_connector"): |
| 173 | + continue |
| 174 | + |
| 175 | + renamed_pt_key = rename_key(pt_key) |
| 176 | + renamed_pt_key = rename_for_ltx2_3_transformer(renamed_pt_key) |
| 177 | + |
| 178 | + pt_tuple_key = tuple(renamed_pt_key.split(".")) |
| 179 | + |
| 180 | + flax_key, flax_tensor = get_key_and_value( |
| 181 | + pt_tuple_key, tensor, flax_state_dict, random_flax_state_dict, scan_layers, num_layers |
| 182 | + ) |
| 183 | + |
| 184 | + flax_state_dict[flax_key] = jax.device_put(jnp.asarray(flax_tensor), device=cpu) |
| 185 | + |
| 186 | + validate_flax_state_dict(eval_shapes, flax_state_dict) |
| 187 | + flax_state_dict = unflatten_dict(flax_state_dict) |
| 188 | + jax.clear_caches() |
| 189 | + return flax_state_dict |
| 190 | + |
| 191 | + |
| 192 | +def load_audio_vae_weights_2_3( |
| 193 | + eval_shapes: dict, |
| 194 | + device: str, |
| 195 | + tensors: dict, |
| 196 | +): |
| 197 | + flax_state_dict = {} |
| 198 | + cpu = jax.local_devices(backend="cpu")[0] |
| 199 | + |
| 200 | + flattened_eval = flatten_dict(eval_shapes) |
| 201 | + |
| 202 | + for pt_key, tensor in tensors.items(): |
| 203 | + # Keys are already filtered and stripped of "audio_vae." by load_and_segregate |
| 204 | + key = rename_for_ltx2_audio_vae(pt_key) |
| 205 | + |
| 206 | + if key.endswith(".kernel") and tensor.ndim == 4: |
| 207 | + tensor = tensor.transpose(2, 3, 1, 0) |
| 208 | + |
| 209 | + flax_key = _tuple_str_to_int(key.split(".")) |
| 210 | + |
| 211 | + if "up_stages" in flax_key: |
| 212 | + up_stages_idx = flax_key.index("up_stages") |
| 213 | + if up_stages_idx + 1 < len(flax_key) and isinstance(flax_key[up_stages_idx + 1], int): |
| 214 | + flax_key_list = list(flax_key) |
| 215 | + flax_key_list[up_stages_idx + 1] = 2 - flax_key[up_stages_idx + 1] |
| 216 | + flax_key = tuple(flax_key_list) |
| 217 | + |
| 218 | + flax_state_dict[flax_key] = jax.device_put(tensor, device=cpu) |
| 219 | + |
| 220 | + filtered_eval_shapes = { |
| 221 | + k: v for k, v in flattened_eval.items() if not any("dropout" in str(x) or "rngs" in str(x) for x in k) |
| 222 | + } |
| 223 | + |
| 224 | + validate_flax_state_dict(unflatten_dict(filtered_eval_shapes), flax_state_dict) |
| 225 | + return unflatten_dict(flax_state_dict) |
| 226 | + |
| 227 | + |
| 228 | +def load_vae_weights_2_3( |
| 229 | + eval_shapes: dict, |
| 230 | + device: str, |
| 231 | + tensors: dict, |
| 232 | +): |
| 233 | + flax_state_dict = {} |
| 234 | + cpu = jax.local_devices(backend="cpu")[0] |
| 235 | + flattened_eval = flatten_dict(eval_shapes) |
| 236 | + |
| 237 | + random_flax_state_dict = {} |
| 238 | + for key in flattened_eval: |
| 239 | + random_flax_state_dict[tuple(str(item) for item in key)] = flattened_eval[key] |
| 240 | + |
| 241 | + for pt_key, tensor in tensors.items(): |
| 242 | + # Remove 'vae.' prefix if present in safetensors but not in model |
| 243 | + if pt_key.startswith("vae."): |
| 244 | + pt_key = pt_key[len("vae."):] |
| 245 | + |
| 246 | + renamed_pt_key = pt_key.replace("nin_shortcut", "conv_shortcut") |
| 247 | + |
| 248 | + pt_tuple_key = tuple(renamed_pt_key.split(".")) |
| 249 | + |
| 250 | + pt_list = [] |
| 251 | + resnet_index = None |
| 252 | + |
| 253 | + for i, part in enumerate(pt_tuple_key): |
| 254 | + if "_" in part and part.split("_")[-1].isdigit(): |
| 255 | + name = "_".join(part.split("_")[:-1]) |
| 256 | + idx = int(part.split("_")[-1]) |
| 257 | + |
| 258 | + if name == "resnets" or name == "block": |
| 259 | + pt_list.append("resnets") |
| 260 | + resnet_index = idx |
| 261 | + elif name == "upsamplers": |
| 262 | + pt_list.append("upsampler") |
| 263 | + elif name in ["down_blocks", "up_blocks", "downsamplers"]: |
| 264 | + pt_list.append(name) |
| 265 | + pt_list.append(str(idx)) |
| 266 | + else: |
| 267 | + pt_list.append(part) |
| 268 | + elif part == "upsampler": |
| 269 | + pt_list.append("upsampler") |
| 270 | + elif part in ["conv1", "conv2", "conv", "conv_in", "conv_out", "conv_shortcut"]: |
| 271 | + pt_list.append(part) |
| 272 | + if ( |
| 273 | + part != "conv" |
| 274 | + and (i + 1 == len(pt_tuple_key) or pt_tuple_key[i + 1] != "conv") |
| 275 | + and (len(pt_list) < 2 or pt_list[-2] != "conv") |
| 276 | + ): |
| 277 | + pt_list.append("conv") |
| 278 | + else: |
| 279 | + pt_list.append(part) |
| 280 | + |
| 281 | + pt_tuple_key = tuple(pt_list) |
| 282 | + |
| 283 | + from .ltx2_utils import rename_key_and_reshape_tensor |
| 284 | + flax_key, flax_tensor = rename_key_and_reshape_tensor(pt_tuple_key, tensor, random_flax_state_dict) |
| 285 | + |
| 286 | + flax_key = _tuple_str_to_int(flax_key) |
| 287 | + |
| 288 | + if resnet_index is not None: |
| 289 | + str_flax_key = tuple([str(x) for x in flax_key]) |
| 290 | + if str_flax_key in random_flax_state_dict: |
| 291 | + if flax_key not in flax_state_dict: |
| 292 | + target_shape = random_flax_state_dict[str_flax_key].shape |
| 293 | + flax_state_dict[flax_key] = jnp.zeros(target_shape, dtype=flax_tensor.dtype) |
| 294 | + flax_state_dict[flax_key] = flax_state_dict[flax_key].at[resnet_index].set(flax_tensor) |
| 295 | + else: |
| 296 | + flax_state_dict[flax_key] = flax_tensor |
| 297 | + else: |
| 298 | + flax_state_dict[flax_key] = jax.device_put(jnp.asarray(flax_tensor), device=cpu) |
| 299 | + |
| 300 | + filtered_eval_shapes = { |
| 301 | + k: v for k, v in flattened_eval.items() if not any("dropout" in str(x) or "rngs" in str(x) for x in k) |
| 302 | + } |
| 303 | + |
| 304 | + validate_flax_state_dict(unflatten_dict(filtered_eval_shapes), flax_state_dict) |
| 305 | + return unflatten_dict(flax_state_dict) |
| 306 | + |
| 307 | + |
| 308 | +def load_vocoder_weights_2_3( |
| 309 | + eval_shapes: dict, |
| 310 | + device: str, |
| 311 | + tensors: dict, |
| 312 | +): |
| 313 | + flax_state_dict = {} |
| 314 | + cpu = jax.local_devices(backend="cpu")[0] |
| 315 | + |
| 316 | + for pt_key, tensor in tensors.items(): |
| 317 | + # Keys are already filtered and stripped of "vocoder." by load_and_segregate |
| 318 | + key = rename_for_ltx2_3_vocoder(pt_key) |
| 319 | + |
| 320 | + # Always apply LTX-2.3 specific replacement |
| 321 | + key = key.replace("resblocks_", "resnets.") |
| 322 | + |
| 323 | + parts = key.split(".") |
| 324 | + |
| 325 | + if parts[-1] == "weight": |
| 326 | + parts[-1] = "kernel" |
| 327 | + |
| 328 | + flax_key = _tuple_str_to_int(parts) |
| 329 | + |
| 330 | + # Skip filter keys as they are derived in NNX model |
| 331 | + if "filter" in flax_key: |
| 332 | + continue |
| 333 | + |
| 334 | + if flax_key[-1] == "kernel": |
| 335 | + if "upsamplers" in flax_key: |
| 336 | + tensor = tensor.transpose(2, 0, 1)[::-1, :, :] |
| 337 | + else: |
| 338 | + tensor = tensor.transpose(2, 1, 0) |
| 339 | + |
| 340 | + if "mel_stft" in flax_key and ("forward_basis" in flax_key or "inverse_basis" in flax_key): |
| 341 | + tensor = tensor.transpose(2, 1, 0) |
| 342 | + |
| 343 | + flax_state_dict[flax_key] = jax.device_put(tensor, device=cpu) |
| 344 | + |
| 345 | + validate_flax_state_dict(eval_shapes, flax_state_dict) |
| 346 | + return unflatten_dict(flax_state_dict) |
| 347 | + |
| 348 | + |
| 349 | +def load_connectors_weights_2_3( |
50 | 350 | pretrained_model_name_or_path: str, |
51 | 351 | eval_shapes: dict, |
52 | 352 | device: str, |
53 | 353 | hf_download: bool = True, |
54 | 354 | subfolder: str = "", |
55 | 355 | filename: str = None, |
56 | 356 | is_ltx2_3: bool = False, |
| 357 | + tensors: dict = None, |
57 | 358 | ): |
58 | 359 | device = jax.local_devices(backend=device)[0] |
59 | 360 |
|
60 | 361 | with jax.default_device(device): |
61 | | - tensors = load_sharded_checkpoint(pretrained_model_name_or_path, subfolder, device, filename=filename) |
| 362 | + if tensors is None: |
| 363 | + tensors = load_sharded_checkpoint(pretrained_model_name_or_path, subfolder, device, filename=filename) |
62 | 364 | flax_state_dict = {} |
63 | 365 | cpu = jax.local_devices(backend="cpu")[0] |
64 | 366 | flattened_eval = flatten_dict(eval_shapes) |
|
0 commit comments