Spaces:
Running
on
Zero
Running
on
Zero
File size: 13,184 Bytes
ebfc6b3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 |
# ruff: noqa: PLC0415
"""
Model loader for LTX-2 trainer using the new ltx-core package.
This module provides a unified interface for loading LTX-2 model components
for training, using SingleGPUModelBuilder from ltx-core.
Example usage:
# Load individual components
vae_encoder = load_video_vae_encoder("/path/to/checkpoint.safetensors", device="cuda")
vae_decoder = load_video_vae_decoder("/path/to/checkpoint.safetensors", device="cuda")
text_encoder = load_text_encoder("/path/to/checkpoint.safetensors", "/path/to/gemma", device="cuda")
# Load all components at once
components = load_model("/path/to/checkpoint.safetensors", text_encoder_path="/path/to/gemma")
"""
from __future__ import annotations
from dataclasses import dataclass
from pathlib import Path
from typing import TYPE_CHECKING
import torch
from ltx_trainer import logger
# Type alias for device specification
Device = str | torch.device
# Type checking imports (not loaded at runtime)
if TYPE_CHECKING:
from ltx_core.model.audio_vae.audio_vae import Decoder as AudioVAEDecoder
from ltx_core.model.audio_vae.audio_vae import Encoder as AudioVAEEncoder
from ltx_core.model.audio_vae.vocoder import Vocoder
from ltx_core.model.clip.gemma.encoders.av_encoder import AVGemmaTextEncoderModel
from ltx_core.model.transformer.model import LTXModel
from ltx_core.model.video_vae.video_vae import Decoder as VideoVAEDecoder
from ltx_core.model.video_vae.video_vae import Encoder as VideoVAEEncoder
from ltx_core.pipeline.components.schedulers import LTX2Scheduler
def _to_torch_device(device: Device) -> torch.device:
"""Convert device specification to torch.device."""
return torch.device(device) if isinstance(device, str) else device
# =============================================================================
# Individual Component Loaders
# =============================================================================
def load_transformer(
checkpoint_path: str | Path,
device: Device = "cpu",
dtype: torch.dtype = torch.bfloat16,
) -> "LTXModel":
"""Load the LTX transformer model.
Args:
checkpoint_path: Path to the safetensors checkpoint file
device: Device to load model on
dtype: Data type for model weights
Returns:
Loaded LTXModel transformer
"""
from ltx_core.loader.single_gpu_model_builder import SingleGPUModelBuilder
from ltx_core.model.transformer.model_configurator import (
LTXV_MODEL_COMFY_RENAMING_MAP,
LTXModelConfigurator,
)
return SingleGPUModelBuilder(
model_path=str(checkpoint_path),
model_class_configurator=LTXModelConfigurator,
model_sd_ops=LTXV_MODEL_COMFY_RENAMING_MAP,
).build(device=_to_torch_device(device), dtype=dtype)
def load_video_vae_encoder(
checkpoint_path: str | Path,
device: Device = "cpu",
dtype: torch.dtype = torch.bfloat16,
) -> "VideoVAEEncoder":
"""Load the video VAE encoder (for preprocessing).
Args:
checkpoint_path: Path to the safetensors checkpoint file
device: Device to load model on
dtype: Data type for model weights
Returns:
Loaded VideoVAEEncoder
"""
from ltx_core.loader.single_gpu_model_builder import SingleGPUModelBuilder
from ltx_core.model.video_vae.model_configurator import VAE_ENCODER_COMFY_KEYS_FILTER
from ltx_core.model.video_vae.model_configurator import (
VAEEncoderConfigurator as VideoVAEEncoderConfigurator,
)
return SingleGPUModelBuilder(
model_path=str(checkpoint_path),
model_class_configurator=VideoVAEEncoderConfigurator,
model_sd_ops=VAE_ENCODER_COMFY_KEYS_FILTER,
).build(device=_to_torch_device(device), dtype=dtype)
def load_video_vae_decoder(
checkpoint_path: str | Path,
device: Device = "cpu",
dtype: torch.dtype = torch.bfloat16,
) -> "VideoVAEDecoder":
"""Load the video VAE decoder (for inference/validation).
Args:
checkpoint_path: Path to the safetensors checkpoint file
device: Device to load model on
dtype: Data type for model weights
Returns:
Loaded VideoVAEDecoder
"""
from ltx_core.loader.single_gpu_model_builder import SingleGPUModelBuilder
from ltx_core.model.video_vae.model_configurator import VAE_DECODER_COMFY_KEYS_FILTER
from ltx_core.model.video_vae.model_configurator import (
VAEDecoderConfigurator as VideoVAEDecoderConfigurator,
)
return SingleGPUModelBuilder(
model_path=str(checkpoint_path),
model_class_configurator=VideoVAEDecoderConfigurator,
model_sd_ops=VAE_DECODER_COMFY_KEYS_FILTER,
).build(device=_to_torch_device(device), dtype=dtype)
def load_audio_vae_encoder(
checkpoint_path: str | Path,
device: Device = "cpu",
dtype: torch.dtype = torch.bfloat16,
) -> "AudioVAEEncoder":
"""Load the audio VAE encoder (for preprocessing).
Args:
checkpoint_path: Path to the safetensors checkpoint file
device: Device to load model on
dtype: Data type for model weights (default bfloat16, but float32 recommended for quality)
Returns:
Loaded AudioVAEEncoder
"""
from ltx_core.loader.single_gpu_model_builder import SingleGPUModelBuilder
from ltx_core.model.audio_vae.model_configurator import AUDIO_VAE_ENCODER_COMFY_KEYS_FILTER
from ltx_core.model.audio_vae.model_configurator import (
VAEEncoderConfigurator as AudioVAEEncoderConfigurator,
)
return SingleGPUModelBuilder(
model_path=str(checkpoint_path),
model_class_configurator=AudioVAEEncoderConfigurator,
model_sd_ops=AUDIO_VAE_ENCODER_COMFY_KEYS_FILTER,
).build(device=_to_torch_device(device), dtype=dtype)
def load_audio_vae_decoder(
checkpoint_path: str | Path,
device: Device = "cpu",
dtype: torch.dtype = torch.bfloat16,
) -> "AudioVAEDecoder":
"""Load the audio VAE decoder.
Args:
checkpoint_path: Path to the safetensors checkpoint file
device: Device to load model on
dtype: Data type for model weights
Returns:
Loaded AudioVAEDecoder
"""
from ltx_core.loader.single_gpu_model_builder import SingleGPUModelBuilder
from ltx_core.model.audio_vae.model_configurator import AUDIO_VAE_DECODER_COMFY_KEYS_FILTER
from ltx_core.model.audio_vae.model_configurator import (
VAEDecoderConfigurator as AudioVAEDecoderConfigurator,
)
return SingleGPUModelBuilder(
model_path=str(checkpoint_path),
model_class_configurator=AudioVAEDecoderConfigurator,
model_sd_ops=AUDIO_VAE_DECODER_COMFY_KEYS_FILTER,
).build(device=_to_torch_device(device), dtype=dtype)
def load_vocoder(
checkpoint_path: str | Path,
device: Device = "cpu",
dtype: torch.dtype = torch.bfloat16,
) -> "Vocoder":
"""Load the vocoder (for audio waveform generation).
Args:
checkpoint_path: Path to the safetensors checkpoint file
device: Device to load model on
dtype: Data type for model weights
Returns:
Loaded Vocoder
"""
from ltx_core.loader.single_gpu_model_builder import SingleGPUModelBuilder
from ltx_core.model.audio_vae.model_configurator import VOCODER_COMFY_KEYS_FILTER, VocoderConfigurator
return SingleGPUModelBuilder(
model_path=str(checkpoint_path),
model_class_configurator=VocoderConfigurator,
model_sd_ops=VOCODER_COMFY_KEYS_FILTER,
).build(device=_to_torch_device(device), dtype=dtype)
def load_text_encoder(
checkpoint_path: str | Path,
gemma_model_path: str | Path,
device: Device = "cpu",
dtype: torch.dtype = torch.bfloat16,
) -> "AVGemmaTextEncoderModel":
"""Load the Gemma text encoder.
Args:
checkpoint_path: Path to the LTX-2 safetensors checkpoint file
gemma_model_path: Path to Gemma model directory
device: Device to load model on
dtype: Data type for model weights
Returns:
Loaded AVGemmaTextEncoderModel
"""
from ltx_core.loader.single_gpu_model_builder import SingleGPUModelBuilder
from ltx_core.model.clip.gemma.encoders.av_encoder import (
AV_GEMMA_TEXT_ENCODER_KEY_OPS,
AVGemmaTextEncoderModelConfigurator,
)
from ltx_core.model.clip.gemma.encoders.base_encoder import module_ops_from_gemma_root
if not Path(gemma_model_path).is_dir():
raise ValueError(f"Gemma model path is not a directory: {gemma_model_path}")
torch_device = _to_torch_device(device)
text_encoder = SingleGPUModelBuilder(
model_path=str(checkpoint_path),
model_class_configurator=AVGemmaTextEncoderModelConfigurator,
model_sd_ops=AV_GEMMA_TEXT_ENCODER_KEY_OPS,
module_ops=module_ops_from_gemma_root(str(gemma_model_path)),
).build(device=torch_device, dtype=dtype)
return text_encoder
# =============================================================================
# Combined Component Loader
# =============================================================================
@dataclass
class LtxModelComponents:
"""Container for all LTX-2 model components."""
transformer: "LTXModel"
video_vae_encoder: "VideoVAEEncoder | None" = None
video_vae_decoder: "VideoVAEDecoder | None" = None
audio_vae_decoder: "AudioVAEDecoder | None" = None
vocoder: "Vocoder | None" = None
text_encoder: "AVGemmaTextEncoderModel | None" = None
scheduler: "LTX2Scheduler | None" = None
def load_model(
checkpoint_path: str | Path,
text_encoder_path: str | Path | None = None,
device: Device = "cpu",
dtype: torch.dtype = torch.bfloat16,
with_video_vae_encoder: bool = False,
with_video_vae_decoder: bool = True,
with_audio_vae_decoder: bool = True,
with_vocoder: bool = True,
with_text_encoder: bool = True,
) -> LtxModelComponents:
"""
Load LTX-2 model components from a safetensors checkpoint.
This is a convenience function that loads multiple components at once.
For loading individual components, use the dedicated functions:
- load_transformer()
- load_video_vae_encoder()
- load_video_vae_decoder()
- load_audio_vae_decoder()
- load_vocoder()
- load_text_encoder()
Args:
checkpoint_path: Path to the safetensors checkpoint file
text_encoder_path: Path to Gemma model directory (required if with_text_encoder=True)
device: Device to load models on ("cuda", "cpu", etc.)
dtype: Data type for model weights
with_video_vae_encoder: Whether to load the video VAE encoder (for preprocessing)
with_video_vae_decoder: Whether to load the video VAE decoder (for inference/validation)
with_audio_vae_decoder: Whether to load the audio VAE decoder
with_vocoder: Whether to load the vocoder
with_text_encoder: Whether to load the text encoder
Returns:
LtxModelComponents containing all loaded model components
"""
from ltx_core.pipeline.components.schedulers import LTX2Scheduler
checkpoint_path = Path(checkpoint_path)
# Validate checkpoint exists
if not checkpoint_path.exists():
raise FileNotFoundError(f"Checkpoint not found: {checkpoint_path}")
logger.info(f"Loading LTX-2 model from {checkpoint_path}")
torch_device = _to_torch_device(device)
# Load transformer
logger.debug("Loading transformer...")
transformer = load_transformer(checkpoint_path, torch_device, dtype)
# Load video VAE encoder
video_vae_encoder = None
if with_video_vae_encoder:
logger.debug("Loading video VAE encoder...")
video_vae_encoder = load_video_vae_encoder(checkpoint_path, torch_device, dtype)
# Load video VAE decoder
video_vae_decoder = None
if with_video_vae_decoder:
logger.debug("Loading video VAE decoder...")
video_vae_decoder = load_video_vae_decoder(checkpoint_path, torch_device, dtype)
# Load audio VAE decoder
audio_vae_decoder = None
if with_audio_vae_decoder:
logger.debug("Loading audio VAE decoder...")
audio_vae_decoder = load_audio_vae_decoder(checkpoint_path, torch_device, dtype)
# Load vocoder
vocoder = None
if with_vocoder:
logger.debug("Loading vocoder...")
vocoder = load_vocoder(checkpoint_path, torch_device, dtype)
# Load text encoder
text_encoder = None
if with_text_encoder:
if text_encoder_path is None:
raise ValueError("text_encoder_path must be provided when with_text_encoder=True")
logger.debug("Loading Gemma text encoder...")
text_encoder = load_text_encoder(checkpoint_path, text_encoder_path, torch_device, dtype)
# Create scheduler (stateless, no loading needed)
scheduler = LTX2Scheduler()
return LtxModelComponents(
transformer=transformer,
video_vae_encoder=video_vae_encoder,
video_vae_decoder=video_vae_decoder,
audio_vae_decoder=audio_vae_decoder,
vocoder=vocoder,
text_encoder=text_encoder,
scheduler=scheduler,
)
|