2565 lines
126 KiB
Python
2565 lines
126 KiB
Python
"""Modified from https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image.py
|
|
"""
|
|
#!/usr/bin/env python
|
|
# coding=utf-8
|
|
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
|
|
import argparse
|
|
import contextlib
|
|
import gc
|
|
import json
|
|
import logging
|
|
import math
|
|
import os
|
|
import pickle
|
|
import shutil
|
|
import sys
|
|
|
|
import accelerate
|
|
import diffusers
|
|
import numpy as np
|
|
import torch
|
|
import torch.distributed as dist
|
|
import torch.nn.functional as F
|
|
import torch.utils.checkpoint
|
|
import torchvision.transforms.functional as TF
|
|
import transformers
|
|
from accelerate import Accelerator, FullyShardedDataParallelPlugin
|
|
from accelerate.logging import get_logger
|
|
from accelerate.state import AcceleratorState
|
|
from accelerate.utils import ProjectConfiguration, set_seed
|
|
from diffusers import DDIMScheduler, FlowMatchEulerDiscreteScheduler
|
|
from diffusers.optimization import get_scheduler
|
|
from diffusers.training_utils import (EMAModel,
|
|
compute_density_for_timestep_sampling,
|
|
compute_loss_weighting_for_sd3)
|
|
from diffusers.utils import check_min_version, deprecate, is_wandb_available
|
|
from diffusers.utils.torch_utils import is_compiled_module
|
|
from einops import rearrange
|
|
from omegaconf import OmegaConf
|
|
from packaging import version
|
|
from PIL import Image
|
|
from torch.distributed.fsdp.fully_sharded_data_parallel import (
|
|
FullOptimStateDictConfig, FullStateDictConfig, ShardedOptimStateDictConfig,
|
|
ShardedStateDictConfig)
|
|
from torch.utils.data import BatchSampler, Dataset, RandomSampler
|
|
from torch.utils.tensorboard import SummaryWriter
|
|
from torchvision import transforms
|
|
from transformers import AutoTokenizer
|
|
from transformers.utils import ContextManagers
|
|
|
|
import datasets
|
|
|
|
current_file_path = os.path.abspath(__file__)
|
|
project_roots = [os.path.dirname(current_file_path), os.path.dirname(os.path.dirname(current_file_path)), os.path.dirname(os.path.dirname(os.path.dirname(current_file_path)))]
|
|
for project_root in project_roots:
|
|
sys.path.insert(0, project_root) if project_root not in sys.path else None
|
|
|
|
from videox_fun.data import (ASPECT_RATIO_512, ASPECT_RATIO_RANDOM_CROP_512,
|
|
ASPECT_RATIO_RANDOM_CROP_PROB,
|
|
AspectRatioBatchImageVideoSampler,
|
|
ImageVideoDataset, ImageVideoSampler,
|
|
RandomSampler, TextDataset, get_closest_ratio,
|
|
get_random_mask)
|
|
from videox_fun.models import (AutoencoderKLWan, AutoencoderKLWan3_8,
|
|
CLIPModel, SparkWan2_2Transformer3DModel,
|
|
Wan2_2Transformer3DModel, WanT5EncoderModel)
|
|
from videox_fun.pipeline import (Wan2_2I2VPipeline, Wan2_2Pipeline,
|
|
Wan2_2TI2VPipeline)
|
|
from videox_fun.utils.discrete_sampler import DiscreteSampling
|
|
from videox_fun.utils.tqdm_bar import PauseAwareTqdm
|
|
from videox_fun.utils.utils import (calculate_dimensions, get_image_latent,
|
|
get_image_to_video_latent,
|
|
save_videos_grid)
|
|
|
|
if is_wandb_available():
|
|
import wandb
|
|
|
|
|
|
def filter_kwargs(cls, kwargs):
|
|
import inspect
|
|
sig = inspect.signature(cls.__init__)
|
|
valid_params = set(sig.parameters.keys()) - {'self', 'cls'}
|
|
filtered_kwargs = {k: v for k, v in kwargs.items() if k in valid_params}
|
|
return filtered_kwargs
|
|
|
|
def get_random_downsample_ratio(sample_size, image_ratio=[],
|
|
all_choices=False, rng=None):
|
|
def _create_special_list(length):
|
|
if length == 1:
|
|
return [1.0]
|
|
if length >= 2:
|
|
first_element = 0.75
|
|
remaining_sum = 1.0 - first_element
|
|
other_elements_value = remaining_sum / (length - 1)
|
|
special_list = [first_element] + [other_elements_value] * (length - 1)
|
|
return special_list
|
|
|
|
if sample_size >= 1536:
|
|
number_list = [1, 1.25, 1.5, 2, 2.5, 3] + image_ratio
|
|
elif sample_size >= 1024:
|
|
number_list = [1, 1.25, 1.5, 2] + image_ratio
|
|
elif sample_size >= 768:
|
|
number_list = [1, 1.25, 1.5] + image_ratio
|
|
elif sample_size >= 512:
|
|
number_list = [1] + image_ratio
|
|
else:
|
|
number_list = [1]
|
|
|
|
if all_choices:
|
|
return number_list
|
|
|
|
number_list_prob = np.array(_create_special_list(len(number_list)))
|
|
if rng is None:
|
|
return np.random.choice(number_list, p = number_list_prob)
|
|
else:
|
|
return rng.choice(number_list, p = number_list_prob)
|
|
|
|
def resize_mask(mask, latent, process_first_frame_only=True):
|
|
latent_size = latent.size()
|
|
batch_size, channels, num_frames, height, width = mask.shape
|
|
|
|
if process_first_frame_only:
|
|
target_size = list(latent_size[2:])
|
|
target_size[0] = 1
|
|
first_frame_resized = F.interpolate(
|
|
mask[:, :, 0:1, :, :],
|
|
size=target_size,
|
|
mode='trilinear',
|
|
align_corners=False
|
|
)
|
|
|
|
target_size = list(latent_size[2:])
|
|
target_size[0] = target_size[0] - 1
|
|
if target_size[0] != 0:
|
|
remaining_frames_resized = F.interpolate(
|
|
mask[:, :, 1:, :, :],
|
|
size=target_size,
|
|
mode='trilinear',
|
|
align_corners=False
|
|
)
|
|
resized_mask = torch.cat([first_frame_resized, remaining_frames_resized], dim=2)
|
|
else:
|
|
resized_mask = first_frame_resized
|
|
else:
|
|
target_size = list(latent_size[2:])
|
|
resized_mask = F.interpolate(
|
|
mask,
|
|
size=target_size,
|
|
mode='trilinear',
|
|
align_corners=False
|
|
)
|
|
return resized_mask
|
|
|
|
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
|
|
check_min_version("0.18.0.dev0")
|
|
|
|
logger = get_logger(__name__, log_level="INFO")
|
|
|
|
def log_validation(vae, text_encoder, tokenizer, transformer3d, args, config, accelerator, weight_dtype, global_step):
|
|
try:
|
|
is_deepspeed = type(transformer3d).__name__ == 'DeepSpeedEngine'
|
|
if is_deepspeed:
|
|
origin_config = transformer3d.config
|
|
transformer3d.config = accelerator.unwrap_model(transformer3d).config
|
|
with torch.no_grad(), torch.cuda.amp.autocast(dtype=weight_dtype), torch.cuda.device(device=accelerator.device):
|
|
# FSDP(CPUOffload)-wrapped modules own their placement: never .to() the sharded
|
|
# text_encoder here or on exit; _te_sharded marks it and the generator below.
|
|
_te_sharded = (
|
|
getattr(accelerator.state, "fsdp_plugin", None) is not None
|
|
or getattr(accelerator.state, "deepspeed_plugin", None) is not None
|
|
)
|
|
logger.info("Running validation... ")
|
|
scheduler = FlowMatchEulerDiscreteScheduler(
|
|
**filter_kwargs(FlowMatchEulerDiscreteScheduler, OmegaConf.to_container(config['scheduler_kwargs']))
|
|
)
|
|
if args.boundary_type == "full":
|
|
transformer3d_1 = accelerator.unwrap_model(transformer3d) if type(transformer3d).__name__ == 'DistributedDataParallel' else transformer3d
|
|
transformer3d_2 = None
|
|
else:
|
|
if args.boundary_type == "low":
|
|
transformer3d_1 = accelerator.unwrap_model(transformer3d) if type(transformer3d).__name__ == 'DistributedDataParallel' else transformer3d
|
|
|
|
sub_path = config['transformer_additional_kwargs'].get('transformer_high_noise_model_subpath', 'transformer')
|
|
transformer3d_2 = (SparkWan2_2Transformer3DModel if args.sparse_student else Wan2_2Transformer3DModel).from_pretrained(
|
|
os.path.join(args.pretrained_model_name_or_path, sub_path),
|
|
transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']),
|
|
low_cpu_mem_usage=True,
|
|
).to(weight_dtype)
|
|
if args.sparse_student:
|
|
transformer3d_2.set_student_topk(args.student_topk)
|
|
else:
|
|
sub_path = config['transformer_additional_kwargs'].get('transformer_low_noise_model_subpath', 'transformer')
|
|
transformer3d_1 = (SparkWan2_2Transformer3DModel if args.sparse_student else Wan2_2Transformer3DModel).from_pretrained(
|
|
os.path.join(args.pretrained_model_name_or_path, sub_path),
|
|
transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']),
|
|
low_cpu_mem_usage=True,
|
|
).to(weight_dtype)
|
|
if args.sparse_student:
|
|
transformer3d_1.set_student_topk(args.student_topk)
|
|
|
|
transformer3d_2 = accelerator.unwrap_model(transformer3d) if type(transformer3d).__name__ == 'DistributedDataParallel' else transformer3d
|
|
|
|
if args.train_mode == "ti2v":
|
|
pipeline = Wan2_2TI2VPipeline(
|
|
vae=vae,
|
|
text_encoder=text_encoder,
|
|
tokenizer=tokenizer,
|
|
transformer=transformer3d_1,
|
|
transformer_2=transformer3d_2,
|
|
scheduler=scheduler,
|
|
)
|
|
elif args.train_mode != "normal":
|
|
pipeline = Wan2_2I2VPipeline(
|
|
vae=vae,
|
|
text_encoder=text_encoder,
|
|
tokenizer=tokenizer,
|
|
transformer=transformer3d_1,
|
|
transformer_2=transformer3d_2,
|
|
scheduler=scheduler,
|
|
)
|
|
else:
|
|
pipeline = Wan2_2Pipeline(
|
|
vae=vae,
|
|
text_encoder=text_encoder,
|
|
tokenizer=tokenizer,
|
|
transformer=transformer3d_1,
|
|
transformer_2=transformer3d_2,
|
|
scheduler=scheduler,
|
|
)
|
|
# Stage only non-FSDP components on GPU; the FSDP-wrapped generator/text_encoder
|
|
# self-stream in place. VAE gets an offload hook and pins the pipeline device.
|
|
from videox_fun.dist import offload_components_cpu
|
|
_fsdp_cls_name = 'FullyShardedDataParallel'
|
|
for _comp in (pipeline.transformer, pipeline.transformer_2):
|
|
if _comp is not None and type(_comp).__name__ != _fsdp_cls_name:
|
|
_comp.to(accelerator.device, dtype=weight_dtype)
|
|
if not _te_sharded:
|
|
text_encoder.to(accelerator.device, dtype=weight_dtype)
|
|
# VAE is never FSDP-wrapped; give it an onload/offload hook and force pipeline device -> GPU.
|
|
offload_components_cpu(pipeline, accelerator.device, ["vae"])
|
|
|
|
if args.seed is None:
|
|
generator = None
|
|
else:
|
|
rank_seed = args.seed + accelerator.process_index
|
|
generator = torch.Generator(device=accelerator.device).manual_seed(rank_seed)
|
|
logger.info(f"Rank {accelerator.process_index} using seed: {rank_seed}")
|
|
|
|
for i in range(len(args.validation_prompts)):
|
|
if args.train_mode != "normal":
|
|
start_image = Image.open(args.validation_paths[i])
|
|
width, height = start_image.width, start_image.height
|
|
width, height = calculate_dimensions(args.image_sample_size * args.image_sample_size, width / height)
|
|
|
|
video_length = int((args.video_sample_n_frames - 1) // vae.config.temporal_compression_ratio * vae.config.temporal_compression_ratio) + 1 if args.video_sample_n_frames != 1 else 1
|
|
input_video, input_video_mask, _ = get_image_to_video_latent(args.validation_paths[i], None, video_length=video_length, sample_size=[height, width])
|
|
sample = pipeline(
|
|
args.validation_prompts[i],
|
|
num_frames = video_length,
|
|
negative_prompt = "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走",
|
|
height = height,
|
|
width = width,
|
|
generator = generator,
|
|
|
|
video = input_video,
|
|
mask_video = input_video_mask,
|
|
num_inference_steps = 4,
|
|
guidance_scale = 1.0,
|
|
boundary = config['transformer_additional_kwargs'].get('boundary', 0.900)
|
|
).videos
|
|
|
|
os.makedirs(os.path.join(args.output_dir, "sample"), exist_ok=True)
|
|
save_videos_grid(
|
|
sample,
|
|
os.path.join(
|
|
args.output_dir,
|
|
f"sample/sample-{global_step}-rank{accelerator.process_index}-image-{i}.mp4"
|
|
)
|
|
)
|
|
else:
|
|
sample = pipeline(
|
|
args.validation_prompts[i],
|
|
num_frames = args.video_sample_n_frames,
|
|
negative_prompt = "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走",
|
|
height = args.video_sample_size,
|
|
width = args.video_sample_size,
|
|
generator = generator,
|
|
num_inference_steps = 4,
|
|
guidance_scale = 1.0,
|
|
boundary = config['transformer_additional_kwargs'].get('boundary', 0.900)
|
|
).videos
|
|
os.makedirs(os.path.join(args.output_dir, "sample"), exist_ok=True)
|
|
save_videos_grid(
|
|
sample,
|
|
os.path.join(
|
|
args.output_dir,
|
|
f"sample/sample-{global_step}-rank{accelerator.process_index}-image-{i}.mp4"
|
|
)
|
|
)
|
|
|
|
# Detach the validation-only VAE offload hook before returning to training so that
|
|
# manual_vae_move keeps sole control of the VAE's device placement.
|
|
for _h in getattr(pipeline, "_cpu_offload_component_hooks", []):
|
|
_h.remove()
|
|
# Drop pipeline AND the local handles to the freshly-loaded complementary expert (a full
|
|
# 14B on GPU, not FSDP-managed). `del pipeline` alone leaves transformer3d_1/_2 pinning it,
|
|
# so empty_cache below reclaims nothing and the 28GB lingers/fragments into the next
|
|
# student backward. The live generator survives via the accelerator's own reference.
|
|
del pipeline, transformer3d_1, transformer3d_2
|
|
gc.collect()
|
|
torch.cuda.empty_cache()
|
|
torch.cuda.ipc_collect()
|
|
vae.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype)
|
|
if not args.enable_text_encoder_in_dataloader and not _te_sharded:
|
|
text_encoder.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype)
|
|
if is_deepspeed:
|
|
transformer3d.config = origin_config
|
|
except Exception as e:
|
|
if 'pipeline' in locals():
|
|
for _h in getattr(pipeline, "_cpu_offload_component_hooks", []):
|
|
_h.remove()
|
|
del pipeline
|
|
# Release the freshly-loaded expert before empty_cache so it is actually reclaimed here.
|
|
if 'transformer3d_1' in locals():
|
|
del transformer3d_1
|
|
if 'transformer3d_2' in locals():
|
|
del transformer3d_2
|
|
gc.collect()
|
|
torch.cuda.empty_cache()
|
|
torch.cuda.ipc_collect()
|
|
print(f"Eval error on rank {accelerator.process_index} with info {e}")
|
|
vae.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype)
|
|
_te_sharded_exc = (
|
|
getattr(accelerator.state, "fsdp_plugin", None) is not None
|
|
or getattr(accelerator.state, "deepspeed_plugin", None) is not None
|
|
)
|
|
if not args.enable_text_encoder_in_dataloader and not _te_sharded_exc:
|
|
text_encoder.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype)
|
|
|
|
|
|
def linear_decay(initial_value, final_value, total_steps, current_step):
|
|
if current_step >= total_steps:
|
|
return final_value
|
|
current_step = max(0, current_step)
|
|
step_size = (final_value - initial_value) / total_steps
|
|
current_value = initial_value + step_size * current_step
|
|
return current_value
|
|
|
|
def generate_timestep_with_lognorm(low, high, shape, device="cpu", generator=None):
|
|
u = torch.normal(mean=0.0, std=1.0, size=shape, device=device, generator=generator)
|
|
t = 1 / (1 + torch.exp(-u)) * (high - low) + low
|
|
return torch.clip(t.to(torch.int32), low, high - 1)
|
|
|
|
def parse_args():
|
|
parser = argparse.ArgumentParser(description="Simple example of a training script.")
|
|
parser.add_argument(
|
|
"--input_perturbation", type=float, default=0, help="The scale of input perturbation. Recommended 0.1."
|
|
)
|
|
parser.add_argument(
|
|
"--pretrained_model_name_or_path",
|
|
type=str,
|
|
default=None,
|
|
required=True,
|
|
help="Path to pretrained model or model identifier from huggingface.co/models.",
|
|
)
|
|
parser.add_argument(
|
|
"--revision",
|
|
type=str,
|
|
default=None,
|
|
required=False,
|
|
help="Revision of pretrained model identifier from huggingface.co/models.",
|
|
)
|
|
parser.add_argument(
|
|
"--variant",
|
|
type=str,
|
|
default=None,
|
|
help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
|
|
)
|
|
parser.add_argument(
|
|
"--train_data_dir",
|
|
type=str,
|
|
default=None,
|
|
help=(
|
|
"A folder containing the training data. "
|
|
),
|
|
)
|
|
parser.add_argument(
|
|
"--train_data_meta",
|
|
type=str,
|
|
default=None,
|
|
help=(
|
|
"A csv containing the training data. "
|
|
),
|
|
)
|
|
parser.add_argument(
|
|
"--max_train_samples",
|
|
type=int,
|
|
default=None,
|
|
help=(
|
|
"For debugging purposes or quicker training, truncate the number of training examples to this "
|
|
"value if set."
|
|
),
|
|
)
|
|
parser.add_argument(
|
|
"--validation_prompts",
|
|
type=str,
|
|
default=None,
|
|
nargs="+",
|
|
help=("A set of prompts evaluated every `--validation_epochs` and logged to `--report_to`."),
|
|
)
|
|
parser.add_argument(
|
|
"--validation_paths",
|
|
type=str,
|
|
default=None,
|
|
nargs="+",
|
|
help=("A set of control videos evaluated every `--validation_epochs` and logged to `--report_to`."),
|
|
)
|
|
parser.add_argument(
|
|
"--negative_prompt",
|
|
type=str,
|
|
default="色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走",
|
|
help=("The negative prompt of cfg distill"),
|
|
)
|
|
parser.add_argument(
|
|
"--output_dir",
|
|
type=str,
|
|
default="sd-model-finetuned",
|
|
help="The output directory where the model predictions and checkpoints will be written.",
|
|
)
|
|
parser.add_argument(
|
|
"--cache_dir",
|
|
type=str,
|
|
default=None,
|
|
help="The directory where the downloaded models and datasets will be stored.",
|
|
)
|
|
parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
|
|
parser.add_argument(
|
|
"--random_flip",
|
|
action="store_true",
|
|
help="whether to randomly flip images horizontally",
|
|
)
|
|
parser.add_argument(
|
|
"--use_came",
|
|
action="store_true",
|
|
help="whether to use came",
|
|
)
|
|
parser.add_argument(
|
|
"--multi_stream",
|
|
action="store_true",
|
|
help="whether to use cuda multi-stream",
|
|
)
|
|
parser.add_argument(
|
|
"--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
|
|
)
|
|
parser.add_argument(
|
|
"--vae_mini_batch", type=int, default=32, help="mini batch size for vae."
|
|
)
|
|
parser.add_argument("--num_train_epochs", type=int, default=100)
|
|
parser.add_argument(
|
|
"--max_train_steps",
|
|
type=int,
|
|
default=None,
|
|
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
|
|
)
|
|
parser.add_argument(
|
|
"--gradient_accumulation_steps",
|
|
type=int,
|
|
default=1,
|
|
help="Number of updates steps to accumulate before performing a backward/update pass.",
|
|
)
|
|
parser.add_argument(
|
|
"--gradient_checkpointing",
|
|
action="store_true",
|
|
help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
|
|
)
|
|
parser.add_argument(
|
|
"--learning_rate",
|
|
type=float,
|
|
default=1e-4,
|
|
help="Initial learning rate (after the potential warmup period) to use.",
|
|
)
|
|
parser.add_argument(
|
|
"--learning_rate_critic",
|
|
type=float,
|
|
default=1e-4,
|
|
help="Initial learning rate (after the potential warmup period) to use.",
|
|
)
|
|
parser.add_argument(
|
|
"--scale_lr",
|
|
action="store_true",
|
|
default=False,
|
|
help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
|
|
)
|
|
parser.add_argument(
|
|
"--lr_scheduler",
|
|
type=str,
|
|
default="constant",
|
|
help=(
|
|
'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
|
|
' "constant", "constant_with_warmup"]'
|
|
),
|
|
)
|
|
parser.add_argument(
|
|
"--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
|
|
)
|
|
parser.add_argument(
|
|
"--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
|
|
)
|
|
parser.add_argument(
|
|
"--allow_tf32",
|
|
action="store_true",
|
|
help=(
|
|
"Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
|
|
" https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
|
|
),
|
|
)
|
|
parser.add_argument(
|
|
"--non_ema_revision",
|
|
type=str,
|
|
default=None,
|
|
required=False,
|
|
help=(
|
|
"Revision of pretrained non-ema model identifier. Must be a branch, tag or git identifier of the local or"
|
|
" remote repository specified with --pretrained_model_name_or_path."
|
|
),
|
|
)
|
|
parser.add_argument(
|
|
"--dataloader_num_workers",
|
|
type=int,
|
|
default=0,
|
|
help=(
|
|
"Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
|
|
),
|
|
)
|
|
parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
|
|
parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
|
|
parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
|
|
parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
|
|
parser.add_argument("--max_grad_norm", default=0.0, type=float, help="Max gradient norm. Default 0 disables gradient clipping to match official CrossDistill (grad_clip=False; Table 4 runs unclipped). Set >0 to enable.")
|
|
parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
|
|
parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
|
|
parser.add_argument(
|
|
"--prediction_type",
|
|
type=str,
|
|
default=None,
|
|
help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediciton_type` is chosen.",
|
|
)
|
|
parser.add_argument(
|
|
"--hub_model_id",
|
|
type=str,
|
|
default=None,
|
|
help="The name of the repository to keep in sync with the local `output_dir`.",
|
|
)
|
|
parser.add_argument(
|
|
"--logging_dir",
|
|
type=str,
|
|
default="logs",
|
|
help=(
|
|
"[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
|
|
" *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
|
|
),
|
|
)
|
|
parser.add_argument(
|
|
"--report_model_info", action="store_true", help="Whether or not to report more info about model (such as norm, grad)."
|
|
)
|
|
parser.add_argument(
|
|
"--mixed_precision",
|
|
type=str,
|
|
default=None,
|
|
choices=["no", "fp16", "bf16"],
|
|
help=(
|
|
"Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
|
|
" 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the"
|
|
" flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
|
|
),
|
|
)
|
|
parser.add_argument(
|
|
"--report_to",
|
|
type=str,
|
|
default="tensorboard",
|
|
help=(
|
|
'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
|
|
' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
|
|
),
|
|
)
|
|
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
|
|
parser.add_argument(
|
|
"--checkpointing_steps",
|
|
type=int,
|
|
default=500,
|
|
help=(
|
|
"Save a checkpoint of the training state every X updates. These checkpoints are only suitable for resuming"
|
|
" training using `--resume_from_checkpoint`."
|
|
),
|
|
)
|
|
parser.add_argument(
|
|
"--checkpoints_total_limit",
|
|
type=int,
|
|
default=None,
|
|
help=("Max number of checkpoints to store."),
|
|
)
|
|
parser.add_argument(
|
|
"--resume_from_checkpoint",
|
|
type=str,
|
|
default=None,
|
|
help=(
|
|
"Whether training should be resumed from a previous checkpoint. Use a path saved by"
|
|
' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
|
|
),
|
|
)
|
|
parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.")
|
|
parser.add_argument(
|
|
"--validation_epochs",
|
|
type=int,
|
|
default=5,
|
|
help="Run validation every X epochs.",
|
|
)
|
|
parser.add_argument(
|
|
"--validation_steps",
|
|
type=int,
|
|
default=2000,
|
|
help="Run validation every X steps.",
|
|
)
|
|
parser.add_argument(
|
|
"--tracker_project_name",
|
|
type=str,
|
|
default="text2image-fine-tune",
|
|
help=(
|
|
"The `project_name` argument passed to Accelerator.init_trackers for"
|
|
" more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator"
|
|
),
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--snr_loss", action="store_true", help="Whether or not to use snr_loss."
|
|
)
|
|
parser.add_argument(
|
|
"--uniform_sampling", action="store_true", help="Whether or not to use uniform_sampling."
|
|
)
|
|
parser.add_argument(
|
|
"--enable_text_encoder_in_dataloader", action="store_true", help="Whether or not to use text encoder in dataloader."
|
|
)
|
|
parser.add_argument(
|
|
"--enable_bucket", action="store_true", help="Whether enable bucket sample in datasets."
|
|
)
|
|
parser.add_argument(
|
|
"--random_ratio_crop", action="store_true", help="Whether enable random ratio crop sample in datasets."
|
|
)
|
|
parser.add_argument(
|
|
"--random_frame_crop", action="store_true", help="Whether enable random frame crop sample in datasets."
|
|
)
|
|
parser.add_argument(
|
|
"--random_hw_adapt", action="store_true", help="Whether enable random adapt height and width in datasets."
|
|
)
|
|
parser.add_argument(
|
|
"--training_with_video_token_length", action="store_true", help="The training stage of the model in training.",
|
|
)
|
|
parser.add_argument(
|
|
"--auto_tile_batch_size", action="store_true", help="Whether to auto tile batch size.",
|
|
)
|
|
parser.add_argument(
|
|
"--motion_sub_loss", action="store_true", help="Whether enable motion sub loss."
|
|
)
|
|
parser.add_argument(
|
|
"--motion_sub_loss_ratio", type=float, default=0.25, help="The ratio of motion sub loss."
|
|
)
|
|
parser.add_argument(
|
|
"--train_sampling_steps",
|
|
type=int,
|
|
default=1000,
|
|
help="Run train_sampling_steps.",
|
|
)
|
|
parser.add_argument(
|
|
"--keep_all_node_same_token_length",
|
|
action="store_true",
|
|
help="Reference of the length token.",
|
|
)
|
|
parser.add_argument(
|
|
"--token_sample_size",
|
|
type=int,
|
|
default=512,
|
|
help="Sample size of the token.",
|
|
)
|
|
parser.add_argument(
|
|
"--video_sample_size",
|
|
type=int,
|
|
default=512,
|
|
help="Sample size of the video.",
|
|
)
|
|
parser.add_argument(
|
|
"--image_sample_size",
|
|
type=int,
|
|
default=512,
|
|
help="Sample size of the image.",
|
|
)
|
|
parser.add_argument(
|
|
"--fix_sample_size",
|
|
nargs=2, type=int, default=None,
|
|
help="Fix Sample size [height, width] when using bucket and collate_fn."
|
|
)
|
|
parser.add_argument(
|
|
"--video_sample_stride",
|
|
type=int,
|
|
default=4,
|
|
help="Sample stride of the video.",
|
|
)
|
|
parser.add_argument(
|
|
"--video_sample_n_frames",
|
|
type=int,
|
|
default=17,
|
|
help="Num frame of video.",
|
|
)
|
|
parser.add_argument(
|
|
"--video_repeat",
|
|
type=int,
|
|
default=0,
|
|
help="Num of repeat video.",
|
|
)
|
|
parser.add_argument(
|
|
"--config_path",
|
|
type=str,
|
|
default=None,
|
|
help=(
|
|
"The config of the model in training."
|
|
),
|
|
)
|
|
parser.add_argument(
|
|
"--transformer_path",
|
|
type=str,
|
|
default=None,
|
|
help=("If you want to load the weight from other transformers, input its path."),
|
|
)
|
|
parser.add_argument(
|
|
"--generator_transformer_path",
|
|
type=str,
|
|
default=None,
|
|
help=("Path to a (sparse stage-1) student checkpoint loaded into the generator. When "
|
|
"--sparse_student is off it is also loaded into the dense critic (fake_score); "
|
|
"with --sparse_student the critic stays dense (initialized from the teacher)."),
|
|
)
|
|
parser.add_argument(
|
|
"--fake_score_transformer_path",
|
|
type=str,
|
|
default=None,
|
|
help=("Path to a checkpoint loaded into the dense critic (fake_score) only."),
|
|
)
|
|
parser.add_argument(
|
|
"--vae_path",
|
|
type=str,
|
|
default=None,
|
|
help=("If you want to load the weight from other vaes, input its path."),
|
|
)
|
|
|
|
parser.add_argument(
|
|
'--trainable_modules',
|
|
nargs='+',
|
|
help='Enter a list of trainable modules'
|
|
)
|
|
parser.add_argument(
|
|
'--trainable_modules_low_learning_rate',
|
|
nargs='+',
|
|
default=[],
|
|
help='Enter a list of trainable modules with lower learning rate'
|
|
)
|
|
parser.add_argument(
|
|
'--tokenizer_max_length',
|
|
type=int,
|
|
default=512,
|
|
help='Max length of tokenizer'
|
|
)
|
|
parser.add_argument(
|
|
"--use_deepspeed", action="store_true", help="Whether or not to use deepspeed."
|
|
)
|
|
parser.add_argument(
|
|
"--use_fsdp", action="store_true", help="Whether or not to use fsdp."
|
|
)
|
|
parser.add_argument(
|
|
"--low_vram", action="store_true", help="Whether enable low_vram mode."
|
|
)
|
|
parser.add_argument(
|
|
"--boundary_type",
|
|
type=str,
|
|
default="low",
|
|
help=(
|
|
'The training boundary type for dual-Transformer architecture. '
|
|
'Support `"low"` (train low-noise model), `"high"` (train high-noise model), '
|
|
'and `"full"` (single model training like TI2V-5B).'
|
|
),
|
|
)
|
|
parser.add_argument(
|
|
"--train_mode",
|
|
type=str,
|
|
default="normal",
|
|
help=(
|
|
'The training mode. Support `"normal"` (T2V, default), '
|
|
'`"i2v"` (Image-to-Video), and `"ti2v"` (Text-Image-to-Video).'
|
|
),
|
|
)
|
|
parser.add_argument(
|
|
"--gen_update_interval",
|
|
type=int,
|
|
default=5,
|
|
help="The ratio to update transformer3d.",
|
|
)
|
|
parser.add_argument(
|
|
"--fake_guidance_scale",
|
|
type=float,
|
|
default=0.0,
|
|
help="The cfg scale for fake iscore.",
|
|
)
|
|
parser.add_argument(
|
|
"--real_guidance_scale",
|
|
type=float,
|
|
default=5.0,
|
|
help="The cfg scale for real score (official teacher_guidance default = 5.0).",
|
|
)
|
|
parser.add_argument(
|
|
'--denoising_step_indices_list',
|
|
nargs='+',
|
|
default=[1000, 750, 500, 250],
|
|
help="The denoising step list.",
|
|
)
|
|
parser.add_argument(
|
|
"--randomize_step_indices",
|
|
action="store_true",
|
|
help="whether to use randomize timesteps indices in training.",
|
|
)
|
|
|
|
parser.add_argument(
|
|
"--distill_loss",
|
|
type=str,
|
|
default="tdm",
|
|
choices=["tdm", "dmd"],
|
|
help="Student distribution-matching loss. TDM and DMD are mutually exclusive; exactly one of --loss_scale_tdm / --loss_scale_dmd must be > 0. 'tdm' = Trajectory Distribution Matching, 'dmd' = Distribution Matching Distillation.",
|
|
)
|
|
parser.add_argument(
|
|
"--loss_scale_tdm", type=float, default=1.0, help="Weight of the TDM student loss (0 disables it)."
|
|
)
|
|
parser.add_argument(
|
|
"--loss_scale_dmd", type=float, default=1.0, help="Weight of the DMD student loss (0 disables it)."
|
|
)
|
|
parser.add_argument(
|
|
"--div_mode",
|
|
type=str,
|
|
default="pcm_random",
|
|
choices=["pcm_random", "pcm_fixed", "ode_match"],
|
|
help="Diversity supervision algorithm (shared by TDM/DMD, keeps the few-step student from mode collapse). 'ode_match' matches the student rf velocity at s=1 to a teacher K-step rollout; 'pcm_*' does PCM-style boundary alignment on phase 0.",
|
|
)
|
|
parser.add_argument(
|
|
"--loss_scale_div", type=float, default=0.0, help="Weight of the diversity loss (0 disables it). Official default is 0; the CrossDistill recipe enables it only for the high-noise expert (see train_distill.sh), keeping the low-noise expert div-free (joint_low_loss_scale_div=0)."
|
|
)
|
|
parser.add_argument(
|
|
"--div_K", type=int, default=8, help="Teacher rollout step count for ode_match diversity."
|
|
)
|
|
parser.add_argument(
|
|
"--div_sigma_max", type=float, default=5000.0, help="sigma_max used to build the ode_match teacher rf grid."
|
|
)
|
|
parser.add_argument(
|
|
"--dcm_total_steps", type=int, default=48, help="Discrete-time CM grid size used by the pcm-div modes."
|
|
)
|
|
parser.add_argument(
|
|
"--dcm_skipping_interval_steps", type=int, default=1, help="Skipping interval on the CM grid used by the pcm-div modes."
|
|
)
|
|
parser.add_argument(
|
|
"--max_simulation_steps_fake", type=int, default=4, help="Max number of student Euler steps during deterministic rf-Euler backward simulation (all times are rf s in [0,1])."
|
|
)
|
|
parser.add_argument(
|
|
"--joint_backward_timesteps",
|
|
nargs='+',
|
|
type=float,
|
|
default=[0.933781, 0.875, 0.608979],
|
|
help="Canonical joint few-step rf grid (rCM, descending); its MIDDLE knot is the MoE boundary. Value = rf time tan(theta)/(1+tan(theta)) at theta=[1.5, boundary, 1.0]. Each expert's --backward_timesteps / --init_rollout_timesteps are auto-split from it at the config boundary.",
|
|
)
|
|
parser.add_argument(
|
|
"--backward_timesteps",
|
|
nargs='+',
|
|
type=float,
|
|
default=[],
|
|
help="rf mid-knots of this expert's backward-simulation schedule (descending). Empty = auto: joint grid split at the boundary (high: knots >= boundary, low: knots < boundary).",
|
|
)
|
|
parser.add_argument(
|
|
"--backward_simulation_start_t",
|
|
type=float,
|
|
default=-1.0,
|
|
help="rf start time of backward simulation; -1 = auto (1.0 = pure noise). For the low expert set to the MoE boundary (e.g. 0.875).",
|
|
)
|
|
parser.add_argument(
|
|
"--backward_simulation_end_t", type=float, default=0.0, help="rf end time of backward simulation; 0 = x0."
|
|
)
|
|
parser.add_argument(
|
|
"--teacher_timestep_shift",
|
|
type=float,
|
|
default=12.0,
|
|
help="Official teacher sampling shift; defines the teacher timestep density (Wan2.2 t2v A14B = 12.0). Real (teacher) guidance is reused from --real_guidance_scale; fake-score guidance from --fake_guidance_scale.",
|
|
)
|
|
parser.add_argument(
|
|
"--rf_t_min",
|
|
type=float,
|
|
default=-1.0,
|
|
help="Lower rf-time bound of this expert's region for partition time sampling (D-time and div drawn from p_D then clamped to [rf_t_min, rf_t_max]). -1 = auto from --boundary_type + config boundary (high: boundary, low/full: 0.0).",
|
|
)
|
|
parser.add_argument(
|
|
"--rf_t_max",
|
|
type=float,
|
|
default=-1.0,
|
|
help="Upper rf-time bound of this expert's region. -1 = auto from --boundary_type + config boundary (low: boundary, high/full: 1.0).",
|
|
)
|
|
parser.add_argument(
|
|
"--p_D",
|
|
type=str,
|
|
default="lognormal",
|
|
choices=["uniform_shift", "lognormal"],
|
|
help="Distribution used to draw the critic/teacher comparison time s_D. Default 'lognormal' matches official p_D=LogNormal(p_mean=0.0, p_std=1.6).",
|
|
)
|
|
parser.add_argument(
|
|
"--p_D_shift", type=float, default=12.0, help="shift for uniform_shift p_D (match teacher_timestep_shift)."
|
|
)
|
|
parser.add_argument(
|
|
"--p_D_mean", type=float, default=0.0, help="mean of lognormal p_D (log-sigma space)."
|
|
)
|
|
parser.add_argument(
|
|
"--p_D_std", type=float, default=1.6, help="std of lognormal p_D (log-sigma space)."
|
|
)
|
|
parser.add_argument(
|
|
"--student_update_freq", type=int, default=5, help="Run the student update once every this many iterations (critic in between)."
|
|
)
|
|
parser.add_argument(
|
|
"--critic_warmup", type=int, default=0, help="Number of initial iterations that always run the student (critic frozen)."
|
|
)
|
|
parser.add_argument(
|
|
"--init_rollout_path",
|
|
type=str,
|
|
default="",
|
|
help="Path to the trained high-noise student safetensors/bin used to roll the frozen trained high student and build the low-noise expert's backward-simulation starting latent online (empty = start from pure noise).",
|
|
)
|
|
parser.add_argument(
|
|
"--sparse_student",
|
|
action="store_true",
|
|
help="Distill a sparse (RoLa/SLA-style) student against the dense teacher: generator and init_rollout use SparkWan2_2Transformer3DModel (sparse + linear self-attention), teacher and critic stay dense.",
|
|
)
|
|
parser.add_argument(
|
|
"--student_topk",
|
|
type=float,
|
|
default=0.03,
|
|
help="Ratio of key blocks kept per query in the sparse student self-attention (sparsity = 1 - topk). Must match the --student_topk used in train_finetune_sparse.py.",
|
|
)
|
|
parser.add_argument(
|
|
"--init_rollout_timesteps",
|
|
nargs='+',
|
|
type=float,
|
|
default=[],
|
|
help="rf mid-knots of the HIGH segment for the init rollout path [1.0, *init_rollout_timesteps, backward_simulation_start_t]. Empty = auto: joint-grid knots above the boundary.",
|
|
)
|
|
|
|
args = parser.parse_args()
|
|
env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
|
|
if env_local_rank != -1 and env_local_rank != args.local_rank:
|
|
args.local_rank = env_local_rank
|
|
|
|
# default to using the same revision for the non-ema model if not specified
|
|
if args.non_ema_revision is None:
|
|
args.non_ema_revision = args.revision
|
|
|
|
return args
|
|
|
|
|
|
def main():
|
|
args = parse_args()
|
|
|
|
# --- SparkDiffusion CrossDistill Argument Validation ---
|
|
if args.train_mode != "normal":
|
|
raise ValueError("SparkDiffusion distillation supports --train_mode normal (T2V) only.")
|
|
if args.distill_loss == "tdm":
|
|
args.loss_scale_dmd = 0.0
|
|
if args.loss_scale_tdm <= 0:
|
|
raise ValueError("--distill_loss tdm requires --loss_scale_tdm > 0.")
|
|
elif args.distill_loss == "dmd":
|
|
args.loss_scale_tdm = 0.0
|
|
if args.loss_scale_dmd <= 0:
|
|
raise ValueError("--distill_loss dmd requires --loss_scale_dmd > 0.")
|
|
assert not (args.loss_scale_tdm > 0 and args.loss_scale_dmd > 0), \
|
|
"loss_scale_tdm and loss_scale_dmd are mutually exclusive; set exactly one > 0."
|
|
if args.loss_scale_div > 0:
|
|
assert args.max_simulation_steps_fake >= 2, (
|
|
"--loss_scale_div > 0 requires --max_simulation_steps_fake >= 2 so that the "
|
|
"first (rf s=1) segment of backward simulation stays no_grad.")
|
|
if args.student_update_freq <= 0:
|
|
raise ValueError("--student_update_freq must be greater than zero.")
|
|
if args.trainable_modules is None:
|
|
args.trainable_modules = [""]
|
|
|
|
if args.report_to == "wandb" and args.hub_token is not None:
|
|
raise ValueError(
|
|
"You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token."
|
|
" Please use `huggingface-cli login` to authenticate with the Hub."
|
|
)
|
|
|
|
if args.non_ema_revision is not None:
|
|
deprecate(
|
|
"non_ema_revision!=None",
|
|
"0.15.0",
|
|
message=(
|
|
"Downloading 'non_ema' weights from revision branches of the Hub is deprecated. Please make sure to"
|
|
" use `--variant=non_ema` instead."
|
|
),
|
|
)
|
|
logging_dir = os.path.join(args.output_dir, args.logging_dir)
|
|
|
|
config = OmegaConf.load(args.config_path)
|
|
|
|
# --- Auto-derive this expert's rf-time region + rollout start from --boundary_type
|
|
# and the model's MoE `boundary` (the same value the pipeline routes on: the high
|
|
# expert handles t >= boundary*1000, i.e. rf s >= boundary). Explicit --rf_t_min /
|
|
# --rf_t_max / --backward_simulation_start_t (>= 0) still override. ---
|
|
_boundary = float(config['transformer_additional_kwargs'].get('boundary', 0.875))
|
|
if args.rf_t_min < 0:
|
|
args.rf_t_min = _boundary if args.boundary_type == "high" else 0.0
|
|
if args.rf_t_max < 0:
|
|
args.rf_t_max = _boundary if args.boundary_type == "low" else 1.0
|
|
if args.backward_simulation_start_t < 0:
|
|
args.backward_simulation_start_t = _boundary if args.boundary_type == "low" else 1.0
|
|
if not args.backward_timesteps:
|
|
if args.boundary_type == "high":
|
|
args.backward_timesteps = [k for k in args.joint_backward_timesteps if k >= _boundary]
|
|
elif args.boundary_type == "low":
|
|
args.backward_timesteps = [k for k in args.joint_backward_timesteps if k < _boundary]
|
|
else: # full
|
|
args.backward_timesteps = [k for k in args.joint_backward_timesteps if 0.0 < k < 1.0]
|
|
if not args.init_rollout_timesteps:
|
|
args.init_rollout_timesteps = [k for k in args.joint_backward_timesteps if k > _boundary]
|
|
if not (0.0 <= args.rf_t_min < args.rf_t_max <= 1.0):
|
|
raise ValueError("require 0 <= --rf_t_min < --rf_t_max <= 1.")
|
|
accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
|
|
|
|
accelerator = Accelerator(
|
|
gradient_accumulation_steps=args.gradient_accumulation_steps,
|
|
mixed_precision=args.mixed_precision,
|
|
log_with=args.report_to,
|
|
project_config=accelerator_project_config,
|
|
)
|
|
accelerator_fake_score_transformer3d = Accelerator(
|
|
gradient_accumulation_steps=args.gradient_accumulation_steps,
|
|
mixed_precision=args.mixed_precision,
|
|
log_with=args.report_to,
|
|
project_config=accelerator_project_config,
|
|
)
|
|
|
|
deepspeed_plugin = accelerator.state.deepspeed_plugin if hasattr(accelerator.state, "deepspeed_plugin") else None
|
|
fsdp_plugin = accelerator.state.fsdp_plugin if hasattr(accelerator.state, "fsdp_plugin") else None
|
|
if deepspeed_plugin is not None:
|
|
zero_stage = int(deepspeed_plugin.zero_stage)
|
|
fsdp_stage = 0
|
|
print(f"Using DeepSpeed Zero stage: {zero_stage}")
|
|
|
|
args.use_deepspeed = True
|
|
if zero_stage == 3:
|
|
print(f"Auto set save_state to True because zero_stage == 3")
|
|
args.save_state = True
|
|
elif fsdp_plugin is not None:
|
|
from torch.distributed.fsdp import ShardingStrategy
|
|
zero_stage = 0
|
|
if fsdp_plugin.sharding_strategy is ShardingStrategy.FULL_SHARD:
|
|
fsdp_stage = 3
|
|
elif fsdp_plugin.sharding_strategy is None: # The fsdp_plugin.sharding_strategy is None in FSDP 2.
|
|
fsdp_stage = 3
|
|
elif fsdp_plugin.sharding_strategy is ShardingStrategy.SHARD_GRAD_OP:
|
|
fsdp_stage = 2
|
|
else:
|
|
fsdp_stage = 0
|
|
print(f"Using FSDP stage: {fsdp_stage}")
|
|
|
|
args.use_fsdp = True
|
|
if fsdp_stage == 3:
|
|
print(f"Auto set save_state to True because fsdp_stage == 3")
|
|
args.save_state = True
|
|
else:
|
|
zero_stage = 0
|
|
fsdp_stage = 0
|
|
print("DeepSpeed is not enabled.")
|
|
|
|
if accelerator.is_main_process:
|
|
writer = SummaryWriter(log_dir=logging_dir)
|
|
|
|
# Make one log on every process with the configuration for debugging.
|
|
logging.basicConfig(
|
|
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
|
|
datefmt="%m/%d/%Y %H:%M:%S",
|
|
level=logging.INFO,
|
|
)
|
|
logger.info(accelerator.state, main_process_only=False)
|
|
if accelerator.is_local_main_process:
|
|
datasets.utils.logging.set_verbosity_warning()
|
|
transformers.utils.logging.set_verbosity_warning()
|
|
diffusers.utils.logging.set_verbosity_info()
|
|
else:
|
|
datasets.utils.logging.set_verbosity_error()
|
|
transformers.utils.logging.set_verbosity_error()
|
|
diffusers.utils.logging.set_verbosity_error()
|
|
|
|
# If passed along, set the training seed now.
|
|
if args.seed is not None:
|
|
set_seed(args.seed)
|
|
rng = np.random.default_rng(np.random.PCG64(args.seed + accelerator.process_index))
|
|
torch_rng = torch.Generator(accelerator.device).manual_seed(args.seed + accelerator.process_index)
|
|
else:
|
|
rng = None
|
|
torch_rng = None
|
|
index_rng = np.random.default_rng(np.random.PCG64(43))
|
|
print(f"Init rng with seed {args.seed + accelerator.process_index}. Process_index is {accelerator.process_index}")
|
|
|
|
# Handle the repository creation
|
|
if accelerator.is_main_process:
|
|
if args.output_dir is not None:
|
|
os.makedirs(args.output_dir, exist_ok=True)
|
|
|
|
# For mixed precision training we cast all non-trainable weigths (vae, non-lora text_encoder and non-lora transformer3d) to half-precision
|
|
# as these weights are only used for inference, keeping weights in full precision is not required.
|
|
weight_dtype = torch.float32
|
|
if accelerator.mixed_precision == "fp16":
|
|
weight_dtype = torch.float16
|
|
args.mixed_precision = accelerator.mixed_precision
|
|
elif accelerator.mixed_precision == "bf16":
|
|
weight_dtype = torch.bfloat16
|
|
args.mixed_precision = accelerator.mixed_precision
|
|
|
|
args.denoising_step_indices_list = [int(i) for i in args.denoising_step_indices_list]
|
|
# Load scheduler, tokenizer and models.
|
|
noise_scheduler = FlowMatchEulerDiscreteScheduler(
|
|
**filter_kwargs(FlowMatchEulerDiscreteScheduler, OmegaConf.to_container(config['scheduler_kwargs']))
|
|
)
|
|
|
|
# Get Tokenizer
|
|
tokenizer = AutoTokenizer.from_pretrained(
|
|
os.path.join(args.pretrained_model_name_or_path, config['text_encoder_kwargs'].get('tokenizer_subpath', 'tokenizer')),
|
|
)
|
|
|
|
def deepspeed_zero_init_disabled_context_manager():
|
|
"""
|
|
returns either a context list that includes one that will disable zero.Init or an empty context list
|
|
"""
|
|
deepspeed_plugin = AcceleratorState().deepspeed_plugin if accelerate.state.is_initialized() else None
|
|
if deepspeed_plugin is None:
|
|
return []
|
|
|
|
return [deepspeed_plugin.zero3_init_context_manager(enable=False)]
|
|
|
|
# Currently Accelerate doesn't know how to handle multiple models under Deepspeed ZeRO stage 3.
|
|
# For this to work properly all models must be run through `accelerate.prepare`. But accelerate
|
|
# will try to assign the same optimizer with the same weights to all models during
|
|
# `deepspeed.initialize`, which of course doesn't work.
|
|
#
|
|
# For now the following workaround will partially support Deepspeed ZeRO-3, by excluding the 2
|
|
# frozen models from being partitioned during `zero.Init` which gets called during
|
|
# `from_pretrained` So CLIPTextModel and AutoencoderKL will not enjoy the parameter sharding
|
|
# across multiple gpus and only UNet2DConditionModel will get ZeRO sharded.
|
|
with ContextManagers(deepspeed_zero_init_disabled_context_manager()):
|
|
# Get Text encoder
|
|
text_encoder = WanT5EncoderModel.from_pretrained(
|
|
os.path.join(args.pretrained_model_name_or_path, config['text_encoder_kwargs'].get('text_encoder_subpath', 'text_encoder')),
|
|
additional_kwargs=OmegaConf.to_container(config['text_encoder_kwargs']),
|
|
low_cpu_mem_usage=True,
|
|
torch_dtype=weight_dtype,
|
|
)
|
|
text_encoder = text_encoder.eval()
|
|
# Get Vae
|
|
Chosen_AutoencoderKL = {
|
|
"AutoencoderKLWan": AutoencoderKLWan,
|
|
"AutoencoderKLWan3_8": AutoencoderKLWan3_8
|
|
}[config['vae_kwargs'].get('vae_type', 'AutoencoderKLWan')]
|
|
vae = Chosen_AutoencoderKL.from_pretrained(
|
|
os.path.join(args.pretrained_model_name_or_path, config['vae_kwargs'].get('vae_subpath', 'vae')),
|
|
additional_kwargs=OmegaConf.to_container(config['vae_kwargs']),
|
|
)
|
|
vae.eval()
|
|
|
|
# Get Transformer
|
|
if args.boundary_type == "low" or args.boundary_type == "full":
|
|
sub_path = config['transformer_additional_kwargs'].get('transformer_low_noise_model_subpath', 'transformer')
|
|
else:
|
|
sub_path = config['transformer_additional_kwargs'].get('transformer_high_noise_model_subpath', 'transformer')
|
|
|
|
# Get Transformer
|
|
# Sparse student (SparkDiffusion stage-1 output): block-sparse + linear self-attention
|
|
# (SparkWan2_2Transformer3DModel); the dense teacher and the dense critic are untouched,
|
|
# matching the reference pipeline where fake_score initializes from the teacher weights.
|
|
def build_student_net(weights_dir):
|
|
if args.sparse_student:
|
|
net = SparkWan2_2Transformer3DModel.from_pretrained(
|
|
weights_dir,
|
|
transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']),
|
|
low_cpu_mem_usage=True,
|
|
).to(weight_dtype)
|
|
# Materialize the zero-init sparse branch that the dense base cannot provide.
|
|
net.set_student_topk(args.student_topk)
|
|
return net
|
|
return Wan2_2Transformer3DModel.from_pretrained(
|
|
weights_dir,
|
|
transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']),
|
|
low_cpu_mem_usage=True,
|
|
).to(weight_dtype)
|
|
|
|
generator_transformer3d = build_student_net(os.path.join(args.pretrained_model_name_or_path, sub_path))
|
|
real_score_transformer3d = Wan2_2Transformer3DModel.from_pretrained(
|
|
os.path.join(args.pretrained_model_name_or_path, sub_path),
|
|
transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']),
|
|
low_cpu_mem_usage=True,
|
|
).to(weight_dtype)
|
|
fake_score_transformer3d = Wan2_2Transformer3DModel.from_pretrained(
|
|
os.path.join(args.pretrained_model_name_or_path, sub_path),
|
|
transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']),
|
|
low_cpu_mem_usage=True,
|
|
).to(weight_dtype)
|
|
|
|
# Wan2.2 low-noise expert: a FROZEN copy of the trained high-noise student used to roll
|
|
# the backward-simulation starting latent from rf 1.0 down to --backward_simulation_start_t.
|
|
init_rollout_transformer3d = None
|
|
if args.init_rollout_path and args.init_rollout_path.strip():
|
|
high_sub_path = config['transformer_additional_kwargs'].get('transformer_high_noise_model_subpath', 'transformer')
|
|
init_rollout_transformer3d = build_student_net(os.path.join(args.pretrained_model_name_or_path, high_sub_path))
|
|
_rollout_ckpt = args.init_rollout_path
|
|
if os.path.isdir(_rollout_ckpt):
|
|
_rollout_ckpt = os.path.join(_rollout_ckpt, "transformer", "diffusion_pytorch_model.safetensors") \
|
|
if os.path.exists(os.path.join(_rollout_ckpt, "transformer", "diffusion_pytorch_model.safetensors")) \
|
|
else os.path.join(_rollout_ckpt, "diffusion_pytorch_model.safetensors")
|
|
if _rollout_ckpt.endswith("safetensors"):
|
|
from safetensors.torch import load_file
|
|
_rollout_sd = load_file(_rollout_ckpt)
|
|
else:
|
|
_rollout_sd = torch.load(_rollout_ckpt, map_location="cpu")
|
|
_rollout_sd = _rollout_sd["state_dict"] if "state_dict" in _rollout_sd else _rollout_sd
|
|
_rollout_sd = { (k.split(".", 1)[1] if k.startswith(("generator_transformer3d.", "transformer.", "model.")) else k): v for k, v in _rollout_sd.items() }
|
|
m, u = init_rollout_transformer3d.load_state_dict(_rollout_sd, strict=False)
|
|
print(f"init_rollout missing keys: {len(m)}, unexpected keys: {len(u)}")
|
|
init_rollout_transformer3d.requires_grad_(False)
|
|
init_rollout_transformer3d.eval()
|
|
|
|
# Freeze vae and text_encoder and set generator_transformer3d to trainable
|
|
vae.requires_grad_(False)
|
|
text_encoder.requires_grad_(False)
|
|
generator_transformer3d.requires_grad_(False)
|
|
real_score_transformer3d.requires_grad_(False)
|
|
fake_score_transformer3d.requires_grad_(False)
|
|
|
|
if args.transformer_path is not None:
|
|
print(f"From checkpoint: {args.transformer_path}")
|
|
if args.transformer_path.endswith("safetensors"):
|
|
from safetensors.torch import load_file, safe_open
|
|
state_dict = load_file(args.transformer_path)
|
|
else:
|
|
state_dict = torch.load(args.transformer_path, map_location="cpu")
|
|
state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict
|
|
|
|
m, u = generator_transformer3d.load_state_dict(state_dict, strict=False)
|
|
print(f"generator missing keys: {len(m)}, unexpected keys: {len(u)}")
|
|
assert len(u) == 0
|
|
m, u = fake_score_transformer3d.load_state_dict(state_dict, strict=False)
|
|
print(f"fake_score missing keys: {len(m)}, unexpected keys: {len(u)}")
|
|
assert len(u) == 0
|
|
m, u = real_score_transformer3d.load_state_dict(state_dict, strict=False)
|
|
print(f"real_score missing keys: {len(m)}, unexpected keys: {len(u)}")
|
|
assert len(u) == 0
|
|
|
|
if args.generator_transformer_path is not None:
|
|
print(f"From generator/fake-score checkpoint: {args.generator_transformer_path}")
|
|
if args.generator_transformer_path.endswith("safetensors"):
|
|
from safetensors.torch import load_file, safe_open
|
|
state_dict = load_file(args.generator_transformer_path)
|
|
else:
|
|
state_dict = torch.load(args.generator_transformer_path, map_location="cpu")
|
|
state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict
|
|
|
|
m, u = generator_transformer3d.load_state_dict(state_dict, strict=False)
|
|
print(f"generator missing keys: {len(m)}, unexpected keys: {len(u)}")
|
|
assert len(u) == 0
|
|
if args.sparse_student:
|
|
# Sparse attn_op.* keys have no dense-critic counterpart; keep fake_score dense
|
|
# (teacher-init), as in the reference pipeline where the critic skips student weights.
|
|
print("sparse_student: keeping fake_score dense (skipping sparse checkpoint load)")
|
|
else:
|
|
m, u = fake_score_transformer3d.load_state_dict(state_dict, strict=False)
|
|
print(f"fake_score missing keys: {len(m)}, unexpected keys: {len(u)}")
|
|
assert len(u) == 0
|
|
|
|
if args.fake_score_transformer_path is not None:
|
|
print(f"From fake-score checkpoint: {args.fake_score_transformer_path}")
|
|
if args.fake_score_transformer_path.endswith("safetensors"):
|
|
from safetensors.torch import load_file, safe_open
|
|
state_dict = load_file(args.fake_score_transformer_path)
|
|
else:
|
|
state_dict = torch.load(args.fake_score_transformer_path, map_location="cpu")
|
|
state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict
|
|
|
|
m, u = fake_score_transformer3d.load_state_dict(state_dict, strict=False)
|
|
print(f"fake_score missing keys: {len(m)}, unexpected keys: {len(u)}")
|
|
assert len(u) == 0
|
|
|
|
if args.vae_path is not None:
|
|
print(f"From checkpoint: {args.vae_path}")
|
|
if args.vae_path.endswith("safetensors"):
|
|
from safetensors.torch import load_file, safe_open
|
|
state_dict = load_file(args.vae_path)
|
|
else:
|
|
state_dict = torch.load(args.vae_path, map_location="cpu")
|
|
state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict
|
|
|
|
m, u = vae.load_state_dict(state_dict, strict=False)
|
|
print(f"missing keys: {len(m)}, unexpected keys: {len(u)}")
|
|
assert len(u) == 0
|
|
|
|
# A good trainable modules is showed below now.
|
|
# For 3D Patch: trainable_modules = ['ff.net', 'pos_embed', 'attn2', 'proj_out', 'timepositionalencoding', 'h_position', 'w_position']
|
|
# For 2D Patch: trainable_modules = ['ff.net', 'attn2', 'timepositionalencoding', 'h_position', 'w_position']
|
|
generator_transformer3d.train()
|
|
fake_score_transformer3d.train()
|
|
if accelerator.is_main_process:
|
|
accelerator.print(
|
|
f"Trainable modules '{args.trainable_modules}'."
|
|
)
|
|
for name, param in generator_transformer3d.named_parameters():
|
|
for trainable_module_name in args.trainable_modules + args.trainable_modules_low_learning_rate:
|
|
if trainable_module_name in name:
|
|
param.requires_grad = True
|
|
break
|
|
for name, param in fake_score_transformer3d.named_parameters():
|
|
for trainable_module_name in args.trainable_modules + args.trainable_modules_low_learning_rate:
|
|
if trainable_module_name in name:
|
|
param.requires_grad = True
|
|
break
|
|
|
|
# `accelerate` 0.16.0 will have better support for customized saving
|
|
if version.parse(accelerate.__version__) >= version.parse("0.16.0"):
|
|
# create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
|
|
if fsdp_stage != 0 or zero_stage == 3:
|
|
def save_model_hook(models, weights, output_dir):
|
|
accelerate_state_dict = accelerator.get_state_dict(models[-1], unwrap=True)
|
|
if accelerator.is_main_process:
|
|
from safetensors.torch import save_file
|
|
|
|
safetensor_save_path = os.path.join(output_dir, f"diffusion_pytorch_model.safetensors")
|
|
accelerate_state_dict = {k: v.to(dtype=weight_dtype) for k, v in accelerate_state_dict.items()}
|
|
save_file(accelerate_state_dict, safetensor_save_path, metadata={"format": "pt"})
|
|
|
|
with open(os.path.join(output_dir, "sampler_pos_start.pkl"), 'wb') as file:
|
|
pickle.dump([batch_sampler.sampler._pos_start, first_epoch], file)
|
|
|
|
def load_model_hook(models, input_dir):
|
|
pkl_path = os.path.join(input_dir, "sampler_pos_start.pkl")
|
|
if os.path.exists(pkl_path):
|
|
with open(pkl_path, 'rb') as file:
|
|
loaded_number, _ = pickle.load(file)
|
|
batch_sampler.sampler._pos_start = max(loaded_number - args.dataloader_num_workers * accelerator.num_processes * 2, 0)
|
|
print(f"Load pkl from {pkl_path}. Get loaded_number = {loaded_number}.")
|
|
else:
|
|
# create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
|
|
def save_model_hook(models, weights, output_dir):
|
|
if accelerator.is_main_process:
|
|
models[0].save_pretrained(os.path.join(output_dir, "transformer"))
|
|
if not args.use_deepspeed:
|
|
weights.pop()
|
|
|
|
with open(os.path.join(output_dir, "sampler_pos_start.pkl"), 'wb') as file:
|
|
pickle.dump([batch_sampler.sampler._pos_start, first_epoch], file)
|
|
|
|
def load_model_hook(models, input_dir):
|
|
for i in range(len(models)):
|
|
# pop models so that they are not loaded again
|
|
model = models.pop()
|
|
|
|
# load diffusers style into model
|
|
load_model = Wan2_2Transformer3DModel.from_pretrained(
|
|
input_dir, subfolder="transformer",
|
|
low_cpu_mem_usage=True,
|
|
)
|
|
model.register_to_config(**load_model.config)
|
|
|
|
model.load_state_dict(load_model.state_dict())
|
|
del load_model
|
|
|
|
pkl_path = os.path.join(input_dir, "sampler_pos_start.pkl")
|
|
if os.path.exists(pkl_path):
|
|
with open(pkl_path, 'rb') as file:
|
|
loaded_number, _ = pickle.load(file)
|
|
batch_sampler.sampler._pos_start = max(loaded_number - args.dataloader_num_workers * accelerator.num_processes * 2, 0)
|
|
print(f"Load pkl from {pkl_path}. Get loaded_number = {loaded_number}.")
|
|
|
|
accelerator.register_save_state_pre_hook(save_model_hook)
|
|
accelerator.register_load_state_pre_hook(load_model_hook)
|
|
accelerator_fake_score_transformer3d.register_save_state_pre_hook(save_model_hook)
|
|
accelerator_fake_score_transformer3d.register_load_state_pre_hook(load_model_hook)
|
|
|
|
if args.gradient_checkpointing:
|
|
generator_transformer3d.enable_gradient_checkpointing()
|
|
fake_score_transformer3d.enable_gradient_checkpointing()
|
|
|
|
# Enable TF32 for faster training on Ampere GPUs,
|
|
# cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
|
|
if args.allow_tf32:
|
|
torch.backends.cuda.matmul.allow_tf32 = True
|
|
|
|
if args.scale_lr:
|
|
args.learning_rate = (
|
|
args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
|
|
)
|
|
|
|
# Initialize the optimizer
|
|
if args.use_8bit_adam:
|
|
try:
|
|
import bitsandbytes as bnb
|
|
except ImportError:
|
|
raise ImportError(
|
|
"Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`"
|
|
)
|
|
|
|
optimizer_cls = bnb.optim.AdamW8bit
|
|
elif args.use_came:
|
|
try:
|
|
from came_pytorch import CAME
|
|
except Exception:
|
|
raise ImportError(
|
|
"Please install came_pytorch to use CAME. You can do so by running `pip install came_pytorch`"
|
|
)
|
|
|
|
optimizer_cls = CAME
|
|
else:
|
|
optimizer_cls = torch.optim.AdamW
|
|
|
|
trainable_params = list(filter(lambda p: p.requires_grad, generator_transformer3d.parameters()))
|
|
trainable_params_optim = [
|
|
{'params': [], 'lr': args.learning_rate},
|
|
{'params': [], 'lr': args.learning_rate / 2},
|
|
]
|
|
in_already = []
|
|
for name, param in generator_transformer3d.named_parameters():
|
|
high_lr_flag = False
|
|
if name in in_already:
|
|
continue
|
|
for trainable_module_name in args.trainable_modules:
|
|
if trainable_module_name in name:
|
|
in_already.append(name)
|
|
high_lr_flag = True
|
|
trainable_params_optim[0]['params'].append(param)
|
|
if accelerator.is_main_process:
|
|
print(f"Set {name} to lr : {args.learning_rate}")
|
|
break
|
|
if high_lr_flag:
|
|
continue
|
|
for trainable_module_name in args.trainable_modules_low_learning_rate:
|
|
if trainable_module_name in name:
|
|
in_already.append(name)
|
|
trainable_params_optim[1]['params'].append(param)
|
|
if accelerator.is_main_process:
|
|
print(f"Set {name} to lr : {args.learning_rate / 2}")
|
|
break
|
|
|
|
fake_trainable_params = list(filter(lambda p: p.requires_grad, fake_score_transformer3d.parameters()))
|
|
fake_trainable_params_optim = [
|
|
{'params': [], 'lr': args.learning_rate_critic},
|
|
{'params': [], 'lr': args.learning_rate_critic / 2},
|
|
]
|
|
in_already = []
|
|
for name, param in fake_score_transformer3d.named_parameters():
|
|
high_lr_flag = False
|
|
if name in in_already:
|
|
continue
|
|
for trainable_module_name in args.trainable_modules:
|
|
if trainable_module_name in name:
|
|
in_already.append(name)
|
|
high_lr_flag = True
|
|
fake_trainable_params_optim[0]['params'].append(param)
|
|
if accelerator.is_main_process:
|
|
print(f"Set {name} to lr : {args.learning_rate_critic}")
|
|
break
|
|
if high_lr_flag:
|
|
continue
|
|
for trainable_module_name in args.trainable_modules_low_learning_rate:
|
|
if trainable_module_name in name:
|
|
in_already.append(name)
|
|
fake_trainable_params_optim[1]['params'].append(param)
|
|
if accelerator.is_main_process:
|
|
print(f"Set {name} to lr : {args.learning_rate_critic / 2}")
|
|
break
|
|
|
|
if args.use_came:
|
|
optimizer = optimizer_cls(
|
|
trainable_params_optim,
|
|
lr=args.learning_rate,
|
|
# weight_decay=args.adam_weight_decay,
|
|
betas=(0.9, 0.999, 0.9999),
|
|
eps=(1e-30, 1e-16)
|
|
)
|
|
critic_optimizer = optimizer_cls(
|
|
fake_trainable_params_optim,
|
|
lr=args.learning_rate_critic,
|
|
# weight_decay=args.adam_weight_decay,
|
|
betas=(0.9, 0.999, 0.9999),
|
|
eps=(1e-30, 1e-16)
|
|
)
|
|
else:
|
|
optimizer = optimizer_cls(
|
|
trainable_params_optim,
|
|
lr=args.learning_rate,
|
|
betas=(args.adam_beta1, args.adam_beta2),
|
|
weight_decay=args.adam_weight_decay,
|
|
eps=args.adam_epsilon,
|
|
)
|
|
critic_optimizer = optimizer_cls(
|
|
fake_trainable_params_optim,
|
|
lr=args.learning_rate_critic,
|
|
betas=(args.adam_beta1, args.adam_beta2),
|
|
weight_decay=args.adam_weight_decay,
|
|
eps=args.adam_epsilon,
|
|
)
|
|
|
|
# Get the training dataset
|
|
sample_n_frames_bucket_interval = vae.config.temporal_compression_ratio
|
|
spatial_compression_ratio = vae.config.spatial_compression_ratio
|
|
|
|
if args.fix_sample_size is not None and args.enable_bucket:
|
|
args.video_sample_size = max(max(args.fix_sample_size), args.video_sample_size)
|
|
args.image_sample_size = max(max(args.fix_sample_size), args.image_sample_size)
|
|
args.training_with_video_token_length = False
|
|
args.random_hw_adapt = False
|
|
|
|
# Get the dataset
|
|
# DFD needs paired videos for the teacher-score input, everything else reuses the DMD data path.
|
|
need_real_video = args.train_mode != "normal"
|
|
if need_real_video:
|
|
train_dataset = ImageVideoDataset(
|
|
args.train_data_meta, args.train_data_dir,
|
|
video_sample_size=args.video_sample_size, video_sample_stride=args.video_sample_stride, video_sample_n_frames=args.video_sample_n_frames,
|
|
video_repeat=args.video_repeat,
|
|
image_sample_size=args.image_sample_size,
|
|
text_drop_ratio=0.0,
|
|
enable_bucket=args.enable_bucket, enable_inpaint=True if args.train_mode != "normal" else False,
|
|
)
|
|
else:
|
|
train_dataset = TextDataset(
|
|
args.train_data_meta
|
|
)
|
|
|
|
def get_length_to_frame_num(token_length):
|
|
if args.image_sample_size > args.video_sample_size:
|
|
sample_sizes = list(range(args.video_sample_size, args.image_sample_size + 1, 128))
|
|
|
|
if sample_sizes[-1] != args.image_sample_size:
|
|
sample_sizes.append(args.image_sample_size)
|
|
else:
|
|
sample_sizes = [args.image_sample_size]
|
|
|
|
length_to_frame_num = {
|
|
sample_size: min(token_length / sample_size / sample_size, args.video_sample_n_frames) // sample_n_frames_bucket_interval * sample_n_frames_bucket_interval + 1 for sample_size in sample_sizes
|
|
}
|
|
|
|
return length_to_frame_num
|
|
|
|
if args.enable_bucket and need_real_video:
|
|
aspect_ratio_sample_size = {key : [x / 512 * args.video_sample_size for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()}
|
|
batch_sampler_generator = torch.Generator().manual_seed(args.seed)
|
|
batch_sampler = AspectRatioBatchImageVideoSampler(
|
|
sampler=RandomSampler(train_dataset, generator=batch_sampler_generator), dataset=train_dataset.dataset,
|
|
batch_size=args.train_batch_size, train_folder = args.train_data_dir, drop_last=True,
|
|
aspect_ratios=aspect_ratio_sample_size,
|
|
)
|
|
|
|
def collate_fn(examples):
|
|
# Get token length
|
|
target_token_length = args.video_sample_n_frames * args.token_sample_size * args.token_sample_size
|
|
length_to_frame_num = get_length_to_frame_num(target_token_length)
|
|
|
|
# Create new output
|
|
new_examples = {}
|
|
new_examples["target_token_length"] = target_token_length
|
|
new_examples["pixel_values"] = []
|
|
new_examples["text"] = []
|
|
# Used in Inpaint mode
|
|
if args.train_mode != "normal":
|
|
new_examples["mask_pixel_values"] = []
|
|
new_examples["mask"] = []
|
|
new_examples["clip_pixel_values"] = []
|
|
|
|
# Get downsample ratio in image and videos
|
|
pixel_value = examples[0]["pixel_values"]
|
|
data_type = examples[0]["data_type"]
|
|
f, h, w, c = np.shape(pixel_value)
|
|
if data_type == 'image':
|
|
random_downsample_ratio = 1 if not args.random_hw_adapt else get_random_downsample_ratio(args.image_sample_size, image_ratio=[args.image_sample_size / args.video_sample_size], rng=rng)
|
|
|
|
aspect_ratio_sample_size = {key : [x / 512 * args.image_sample_size / random_downsample_ratio for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()}
|
|
aspect_ratio_random_crop_sample_size = {key : [x / 512 * args.image_sample_size / random_downsample_ratio for x in ASPECT_RATIO_RANDOM_CROP_512[key]] for key in ASPECT_RATIO_RANDOM_CROP_512.keys()}
|
|
|
|
batch_video_length = args.video_sample_n_frames + sample_n_frames_bucket_interval
|
|
else:
|
|
if args.random_hw_adapt:
|
|
if args.training_with_video_token_length:
|
|
local_min_size = np.min(np.array([np.mean(np.array([np.shape(example["pixel_values"])[1], np.shape(example["pixel_values"])[2]])) for example in examples]))
|
|
# The video will be resized to a lower resolution than its own.
|
|
choice_list = [length for length in list(length_to_frame_num.keys()) if length < local_min_size * 1.25]
|
|
if len(choice_list) == 0:
|
|
choice_list = list(length_to_frame_num.keys())
|
|
if rng is None:
|
|
local_video_sample_size = np.random.choice(choice_list)
|
|
else:
|
|
local_video_sample_size = rng.choice(choice_list)
|
|
batch_video_length = length_to_frame_num[local_video_sample_size]
|
|
random_downsample_ratio = args.video_sample_size / local_video_sample_size
|
|
else:
|
|
random_downsample_ratio = get_random_downsample_ratio(
|
|
args.video_sample_size, rng=rng)
|
|
batch_video_length = args.video_sample_n_frames + sample_n_frames_bucket_interval
|
|
else:
|
|
random_downsample_ratio = 1
|
|
batch_video_length = args.video_sample_n_frames + sample_n_frames_bucket_interval
|
|
|
|
aspect_ratio_sample_size = {key : [x / 512 * args.video_sample_size / random_downsample_ratio for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()}
|
|
aspect_ratio_random_crop_sample_size = {key : [x / 512 * args.video_sample_size / random_downsample_ratio for x in ASPECT_RATIO_RANDOM_CROP_512[key]] for key in ASPECT_RATIO_RANDOM_CROP_512.keys()}
|
|
|
|
if args.fix_sample_size is not None:
|
|
fix_sample_size = [int(x / spatial_compression_ratio / 2) * spatial_compression_ratio * 2 for x in args.fix_sample_size]
|
|
elif args.random_ratio_crop:
|
|
if rng is None:
|
|
random_sample_size = aspect_ratio_random_crop_sample_size[
|
|
np.random.choice(list(aspect_ratio_random_crop_sample_size.keys()), p = ASPECT_RATIO_RANDOM_CROP_PROB)
|
|
]
|
|
else:
|
|
random_sample_size = aspect_ratio_random_crop_sample_size[
|
|
rng.choice(list(aspect_ratio_random_crop_sample_size.keys()), p = ASPECT_RATIO_RANDOM_CROP_PROB)
|
|
]
|
|
random_sample_size = [int(x / spatial_compression_ratio / 2) * spatial_compression_ratio * 2 for x in random_sample_size]
|
|
else:
|
|
closest_size, closest_ratio = get_closest_ratio(h, w, ratios=aspect_ratio_sample_size)
|
|
closest_size = [int(x / spatial_compression_ratio / 2) * spatial_compression_ratio * 2 for x in closest_size]
|
|
|
|
min_example_length = min(
|
|
[example["pixel_values"].shape[0] for example in examples]
|
|
)
|
|
batch_video_length = int(min(batch_video_length, min_example_length))
|
|
|
|
# Magvae needs the number of frames to be 4n + 1.
|
|
batch_video_length = (batch_video_length - 1) // sample_n_frames_bucket_interval * sample_n_frames_bucket_interval + 1
|
|
|
|
if batch_video_length <= 0:
|
|
batch_video_length = 1
|
|
|
|
for example in examples:
|
|
if args.fix_sample_size is not None:
|
|
# To 0~1
|
|
pixel_values = torch.from_numpy(example["pixel_values"]).permute(0, 3, 1, 2).contiguous()
|
|
pixel_values = pixel_values / 255.
|
|
|
|
# Get adapt hw for resize
|
|
fix_sample_size = list(map(lambda x: int(x), fix_sample_size))
|
|
transform = transforms.Compose([
|
|
transforms.Resize(fix_sample_size, interpolation=transforms.InterpolationMode.BILINEAR), # Image.BICUBIC
|
|
transforms.CenterCrop(fix_sample_size),
|
|
transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True),
|
|
])
|
|
elif args.random_ratio_crop:
|
|
# To 0~1
|
|
pixel_values = torch.from_numpy(example["pixel_values"]).permute(0, 3, 1, 2).contiguous()
|
|
pixel_values = pixel_values / 255.
|
|
|
|
# Get adapt hw for resize
|
|
b, c, h, w = pixel_values.size()
|
|
th, tw = random_sample_size
|
|
if th / tw > h / w:
|
|
nh = int(th)
|
|
nw = int(w / h * nh)
|
|
else:
|
|
nw = int(tw)
|
|
nh = int(h / w * nw)
|
|
|
|
transform = transforms.Compose([
|
|
transforms.Resize([nh, nw]),
|
|
transforms.CenterCrop([int(x) for x in random_sample_size]),
|
|
transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True),
|
|
])
|
|
else:
|
|
# To 0~1
|
|
pixel_values = torch.from_numpy(example["pixel_values"]).permute(0, 3, 1, 2).contiguous()
|
|
pixel_values = pixel_values / 255.
|
|
|
|
# Get adapt hw for resize
|
|
closest_size = list(map(lambda x: int(x), closest_size))
|
|
if closest_size[0] / h > closest_size[1] / w:
|
|
resize_size = closest_size[0], int(w * closest_size[0] / h)
|
|
else:
|
|
resize_size = int(h * closest_size[1] / w), closest_size[1]
|
|
|
|
transform = transforms.Compose([
|
|
transforms.Resize(resize_size, interpolation=transforms.InterpolationMode.BILINEAR), # Image.BICUBIC
|
|
transforms.CenterCrop(closest_size),
|
|
transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True),
|
|
])
|
|
|
|
new_examples["pixel_values"].append(transform(pixel_values)[:batch_video_length])
|
|
new_examples["text"].append(example["text"])
|
|
|
|
if args.train_mode != "normal":
|
|
mask = get_random_mask(new_examples["pixel_values"][-1].size(), image_start_only=True)
|
|
mask_pixel_values = new_examples["pixel_values"][-1] * (1 - mask)
|
|
# Wan 2.1 use 0 for masked pixels
|
|
# + torch.ones_like(new_examples["pixel_values"][-1]) * -1 * mask
|
|
new_examples["mask_pixel_values"].append(mask_pixel_values)
|
|
new_examples["mask"].append(mask)
|
|
|
|
clip_pixel_values = new_examples["pixel_values"][-1][0].permute(1, 2, 0).contiguous()
|
|
clip_pixel_values = (clip_pixel_values * 0.5 + 0.5) * 255
|
|
new_examples["clip_pixel_values"].append(clip_pixel_values)
|
|
|
|
# Limit the number of frames to the same
|
|
new_examples["pixel_values"] = torch.stack([example for example in new_examples["pixel_values"]])
|
|
if args.train_mode != "normal":
|
|
new_examples["mask_pixel_values"] = torch.stack([example for example in new_examples["mask_pixel_values"]])
|
|
new_examples["mask"] = torch.stack([example for example in new_examples["mask"]])
|
|
new_examples["clip_pixel_values"] = torch.stack([example for example in new_examples["clip_pixel_values"]])
|
|
|
|
# Encode prompts when enable_text_encoder_in_dataloader=True
|
|
if args.enable_text_encoder_in_dataloader:
|
|
prompt_ids = tokenizer(
|
|
new_examples['text'],
|
|
max_length=args.tokenizer_max_length,
|
|
padding="max_length",
|
|
add_special_tokens=True,
|
|
truncation=True,
|
|
return_tensors="pt"
|
|
)
|
|
text_input_ids = prompt_ids.input_ids
|
|
prompt_attention_mask = prompt_ids.attention_mask
|
|
|
|
seq_lens = prompt_attention_mask.gt(0).sum(dim=1).long()
|
|
prompt_embeds = text_encoder(text_input_ids.to("cpu"), attention_mask=prompt_attention_mask.to("cpu"))[0]
|
|
prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)]
|
|
|
|
new_examples['encoder_attention_mask'] = prompt_ids.attention_mask
|
|
new_examples['encoder_hidden_states'] = prompt_embeds
|
|
|
|
neg_txt = [
|
|
"色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" for text in batch['text']
|
|
]
|
|
neg_prompt_ids = tokenizer(
|
|
neg_txt,
|
|
max_length=args.tokenizer_max_length,
|
|
padding="max_length",
|
|
add_special_tokens=True,
|
|
truncation=True,
|
|
return_tensors="pt"
|
|
)
|
|
neg_text_input_ids = neg_prompt_ids.input_ids
|
|
neg_prompt_attention_mask = neg_prompt_ids.attention_mask
|
|
|
|
neg_seq_lens = neg_prompt_attention_mask.gt(0).sum(dim=1).long()
|
|
neg_prompt_embeds = text_encoder(neg_text_input_ids.to("cpu"), attention_mask=neg_prompt_attention_mask.to("cpu"))[0]
|
|
neg_prompt_embeds = [u[:v] for u, v in zip(neg_prompt_embeds, neg_seq_lens)]
|
|
|
|
new_examples['neg_encoder_attention_mask'] = neg_prompt_ids.attention_mask
|
|
new_examples['neg_encoder_hidden_states'] = neg_prompt_embeds
|
|
|
|
return new_examples
|
|
|
|
# DataLoaders creation:
|
|
train_dataloader = torch.utils.data.DataLoader(
|
|
train_dataset,
|
|
batch_sampler=batch_sampler,
|
|
collate_fn=collate_fn,
|
|
persistent_workers=True if args.dataloader_num_workers != 0 else False,
|
|
num_workers=args.dataloader_num_workers,
|
|
)
|
|
elif not need_real_video:
|
|
def collate_fn(examples):
|
|
new_examples = {}
|
|
new_examples["text"] = []
|
|
for example in examples:
|
|
new_examples["text"].append(example["text"])
|
|
|
|
# Encode prompts when enable_text_encoder_in_dataloader=True
|
|
if args.enable_text_encoder_in_dataloader:
|
|
prompt_ids = tokenizer(
|
|
new_examples['text'],
|
|
max_length=args.tokenizer_max_length,
|
|
padding="max_length",
|
|
add_special_tokens=True,
|
|
truncation=True,
|
|
return_tensors="pt"
|
|
)
|
|
text_input_ids = prompt_ids.input_ids
|
|
prompt_attention_mask = prompt_ids.attention_mask
|
|
|
|
seq_lens = prompt_attention_mask.gt(0).sum(dim=1).long()
|
|
prompt_embeds = text_encoder(text_input_ids.to("cpu"), attention_mask=prompt_attention_mask.to("cpu"))[0]
|
|
prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)]
|
|
|
|
new_examples['encoder_attention_mask'] = prompt_ids.attention_mask
|
|
new_examples['encoder_hidden_states'] = prompt_embeds
|
|
|
|
neg_txt = [
|
|
"色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" for text in batch['text']
|
|
]
|
|
neg_prompt_ids = tokenizer(
|
|
neg_txt,
|
|
max_length=args.tokenizer_max_length,
|
|
padding="max_length",
|
|
add_special_tokens=True,
|
|
truncation=True,
|
|
return_tensors="pt"
|
|
)
|
|
neg_text_input_ids = neg_prompt_ids.input_ids
|
|
neg_prompt_attention_mask = neg_prompt_ids.attention_mask
|
|
|
|
neg_seq_lens = neg_prompt_attention_mask.gt(0).sum(dim=1).long()
|
|
neg_prompt_embeds = text_encoder(neg_text_input_ids.to("cpu"), attention_mask=neg_prompt_attention_mask.to("cpu"))[0]
|
|
neg_prompt_embeds = [u[:v] for u, v in zip(neg_prompt_embeds, neg_seq_lens)]
|
|
|
|
new_examples['neg_encoder_attention_mask'] = neg_prompt_ids.attention_mask
|
|
new_examples['neg_encoder_hidden_states'] = neg_prompt_embeds
|
|
|
|
return new_examples
|
|
|
|
batch_sampler_generator = torch.Generator().manual_seed(args.seed)
|
|
batch_sampler = BatchSampler(RandomSampler(train_dataset, generator=batch_sampler_generator), batch_size=args.train_batch_size, drop_last=True)
|
|
|
|
# DataLoaders creation:
|
|
train_dataloader = torch.utils.data.DataLoader(
|
|
train_dataset,
|
|
batch_sampler=batch_sampler,
|
|
collate_fn=collate_fn,
|
|
persistent_workers=True if args.dataloader_num_workers != 0 else False,
|
|
num_workers=args.dataloader_num_workers,
|
|
)
|
|
else:
|
|
# DataLoaders creation:
|
|
batch_sampler_generator = torch.Generator().manual_seed(args.seed)
|
|
batch_sampler = ImageVideoSampler(RandomSampler(train_dataset, generator=batch_sampler_generator), train_dataset, args.train_batch_size)
|
|
train_dataloader = torch.utils.data.DataLoader(
|
|
train_dataset,
|
|
batch_sampler=batch_sampler,
|
|
persistent_workers=True if args.dataloader_num_workers != 0 else False,
|
|
num_workers=args.dataloader_num_workers,
|
|
)
|
|
|
|
# Scheduler and math around the number of training steps.
|
|
overrode_max_train_steps = False
|
|
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
|
|
if args.max_train_steps is None:
|
|
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
|
|
overrode_max_train_steps = True
|
|
|
|
lr_scheduler = get_scheduler(
|
|
args.lr_scheduler,
|
|
optimizer=optimizer,
|
|
num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
|
|
num_training_steps=args.max_train_steps * accelerator.num_processes,
|
|
)
|
|
fake_score_lr_scheduler = get_scheduler(
|
|
args.lr_scheduler,
|
|
optimizer=critic_optimizer,
|
|
num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
|
|
num_training_steps=args.max_train_steps * accelerator.num_processes,
|
|
)
|
|
|
|
# Prepare everything with our `accelerator`.
|
|
generator_transformer3d, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
|
|
generator_transformer3d, optimizer, train_dataloader, lr_scheduler
|
|
)
|
|
fake_score_transformer3d, critic_optimizer, fake_score_lr_scheduler= accelerator_fake_score_transformer3d.prepare(
|
|
fake_score_transformer3d, critic_optimizer, fake_score_lr_scheduler
|
|
)
|
|
if fsdp_stage != 0 or zero_stage != 0:
|
|
from functools import partial
|
|
|
|
from videox_fun.dist import set_multi_gpus_devices, shard_model
|
|
# Under --low_vram use FSDP CPUOffload for CPU<->GPU streaming instead of manual
|
|
# .to(device)/.to('cpu'), which fights the FSDP flat_param.
|
|
shard_fn = partial(
|
|
shard_model,
|
|
device_id=accelerator.device,
|
|
param_dtype=weight_dtype,
|
|
offload_to_cpu=args.low_vram,
|
|
)
|
|
real_score_transformer3d = shard_fn(real_score_transformer3d)
|
|
if init_rollout_transformer3d is not None:
|
|
init_rollout_transformer3d = shard_fn(init_rollout_transformer3d)
|
|
text_encoder = shard_fn(text_encoder)
|
|
|
|
# Move text_encode and vae to gpu and cast to weight_dtype.
|
|
# vae is never sharded and needs explicit placement; sharded modules are owned by FSDP,
|
|
# so skip .to() on them to keep the flat_param attached.
|
|
vae.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype)
|
|
manual_te_move = args.low_vram and fsdp_stage == 0 and zero_stage == 0
|
|
manual_vae_move = args.low_vram
|
|
if fsdp_stage == 0 and zero_stage == 0:
|
|
real_score_transformer3d.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype)
|
|
if init_rollout_transformer3d is not None:
|
|
init_rollout_transformer3d.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype)
|
|
if not args.enable_text_encoder_in_dataloader:
|
|
text_encoder.to(accelerator.device if not args.low_vram else "cpu")
|
|
|
|
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
|
|
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
|
|
if overrode_max_train_steps:
|
|
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
|
|
# Afterwards we recalculate our number of training epochs
|
|
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
|
|
|
|
# We need to initialize the trackers we use, and also store our configuration.
|
|
# The trackers initializes automatically on the main process.
|
|
if accelerator.is_main_process:
|
|
tracker_config = dict(vars(args))
|
|
keys_to_pop = [k for k, v in tracker_config.items() if isinstance(v, list)]
|
|
for k in keys_to_pop:
|
|
tracker_config.pop(k)
|
|
print(f"Removed tracker_config['{k}']")
|
|
accelerator.init_trackers(args.tracker_project_name, tracker_config)
|
|
|
|
# Function for unwrapping if model was compiled with `torch.compile`.
|
|
def unwrap_model(model):
|
|
model = accelerator.unwrap_model(model)
|
|
model = model._orig_mod if is_compiled_module(model) else model
|
|
return model
|
|
|
|
# Train!
|
|
total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
|
|
|
|
logger.info("***** Running training *****")
|
|
logger.info(f" Num examples = {len(train_dataset)}")
|
|
logger.info(f" Num Epochs = {args.num_train_epochs}")
|
|
logger.info(f" Instantaneous batch size per device = {args.train_batch_size}")
|
|
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
|
|
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
|
|
logger.info(f" Total optimization steps = {args.max_train_steps}")
|
|
global_step = 0
|
|
first_epoch = 0
|
|
|
|
# Potentially load in the weights and states from a previous save
|
|
if args.resume_from_checkpoint:
|
|
if args.resume_from_checkpoint != "latest":
|
|
path = os.path.basename(args.resume_from_checkpoint)
|
|
else:
|
|
# Get the most recent checkpoint
|
|
dirs = os.listdir(args.output_dir)
|
|
dirs = [d for d in dirs if d.startswith("checkpoint")]
|
|
dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
|
|
path = dirs[-1] if len(dirs) > 0 else None
|
|
|
|
if path is None:
|
|
accelerator.print(
|
|
f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
|
|
)
|
|
args.resume_from_checkpoint = None
|
|
initial_global_step = 0
|
|
else:
|
|
global_step = int(path.split("-")[1])
|
|
|
|
initial_global_step = global_step
|
|
|
|
pkl_path = os.path.join(os.path.join(args.output_dir, path), "sampler_pos_start.pkl")
|
|
if os.path.exists(pkl_path):
|
|
with open(pkl_path, 'rb') as file:
|
|
_, first_epoch = pickle.load(file)
|
|
else:
|
|
first_epoch = global_step // num_update_steps_per_epoch
|
|
print(f"Load pkl from {pkl_path}. Get first_epoch = {first_epoch}.")
|
|
|
|
accelerator.print(f"Resuming from checkpoint {path}")
|
|
fake_score_path = os.path.join(path, "fake_score")
|
|
accelerator.load_state(os.path.join(args.output_dir, path))
|
|
accelerator_fake_score_transformer3d.load_state(os.path.join(args.output_dir, fake_score_path))
|
|
else:
|
|
initial_global_step = 0
|
|
|
|
progress_bar = PauseAwareTqdm(
|
|
range(0, args.max_train_steps),
|
|
initial=initial_global_step,
|
|
desc="Steps",
|
|
# Only show the progress bar once on each machine.
|
|
disable=not accelerator.is_local_main_process,
|
|
)
|
|
|
|
if args.multi_stream and args.train_mode != "normal":
|
|
# create extra cuda streams to speedup inpaint vae computation
|
|
vae_stream_1 = torch.cuda.Stream()
|
|
vae_stream_2 = torch.cuda.Stream()
|
|
else:
|
|
vae_stream_1 = None
|
|
vae_stream_2 = None
|
|
|
|
# --- SparkDiffusion CrossDistill Training Core ---
|
|
# rf-native: time s in [0,1], latent x_s = (1-s)*x0 + s*eps, net fed t = s*T_SCALE predicts v = eps - x0.
|
|
T_SCALE = 1000.0 # rectified_flow_t_scaling_factor
|
|
|
|
def shift_rf_time(u, shift):
|
|
return shift * u / (1.0 + (shift - 1.0) * u)
|
|
|
|
def inv_shift_rf_time(rf, shift):
|
|
return rf / (shift - (shift - 1.0) * rf) if shift > 0 else rf
|
|
|
|
def draw_rf_time(shape, lower_bound=None, strict_lower=False):
|
|
"""Sample the comparison/critic rf time s_D ~ p_D, clamped to this expert's region."""
|
|
dev, dtp = accelerator.device, torch.float64
|
|
lo, hi = float(args.rf_t_min), float(args.rf_t_max)
|
|
n = int(torch.Size(shape).numel()) if not isinstance(shape, int) else int(shape)
|
|
if args.p_D == "uniform_shift":
|
|
shift = args.p_D_shift
|
|
u_lo, u_hi = inv_shift_rf_time(lo, shift), inv_shift_rf_time(hi, shift)
|
|
u = torch.rand(n, device=dev, dtype=dtp) * (u_hi - u_lo) + u_lo
|
|
rf = shift_rf_time(u, shift)
|
|
else: # lognormal in log-sigma space, mapped to rf via sigma/(1+sigma)
|
|
import torch.distributions as D
|
|
normal = D.Normal(args.p_D_mean, args.p_D_std)
|
|
eps_val = torch.finfo(torch.float64).eps
|
|
lg_lo = torch.tensor(lo / (1.0 - lo + eps_val)).log() if lo > 0 else torch.tensor(-40.0, dtype=dtp)
|
|
lg_hi = torch.tensor(hi / (1.0 - hi + eps_val)).log() if hi < 1 else torch.tensor(40.0, dtype=dtp)
|
|
cdf_lo, cdf_hi = normal.cdf(lg_lo.double()), normal.cdf(lg_hi.double())
|
|
q = torch.rand(n, device=dev, dtype=dtp)
|
|
q = (cdf_lo + q * (cdf_hi - cdf_lo)).clamp(min=eps_val, max=1.0 - eps_val)
|
|
rf = normal.icdf(q).exp()
|
|
rf = rf / (rf + 1.0)
|
|
if not isinstance(shape, int):
|
|
rf = rf.view(shape)
|
|
if lower_bound is not None:
|
|
lb = torch.as_tensor(lower_bound, device=dev, dtype=dtp)
|
|
while lb.ndim < rf.ndim:
|
|
lb = lb.unsqueeze(-1)
|
|
rf = torch.maximum(rf, lb + (1e-4 if strict_lower else 0.0))
|
|
return rf.clamp(min=lo, max=1.0).to(torch.float32)
|
|
|
|
# Broadcast a sampled time/noise from rank 0 so all ranks train on identical randomness.
|
|
def sync_broadcast(t):
|
|
if dist.is_initialized():
|
|
t = t.clone()
|
|
dist.broadcast(t, src=0)
|
|
return t
|
|
|
|
# --- Network Forwards (net takes t in [0,1000], returns rf velocity v = eps - x0) ---
|
|
def net_velocity(net, x_5d, s_B_1, embeds, seq_len):
|
|
t = (s_B_1.reshape(-1) * T_SCALE).to(torch.float32)
|
|
v = net(
|
|
x=x_5d,
|
|
context=embeds,
|
|
t=t,
|
|
seq_len=seq_len,
|
|
y=None
|
|
)
|
|
return v.float()
|
|
|
|
# Invert the rf interpolation x_s = (1-s)*x0 + s*eps => x0 = x_s - s*v.
|
|
def net_x0(net, x_5d, s_B_1, embeds, seq_len):
|
|
v = net_velocity(net, x_5d, s_B_1, embeds, seq_len)
|
|
s5 = s_B_1.reshape(-1, 1, 1, 1, 1).to(torch.float64)
|
|
return x_5d.double() - s5 * v.double(), v
|
|
|
|
# One rf-Euler step: x_{s_next} = x_{s_cur} + (s_next - s_cur)*v.
|
|
def _euler(net, x, s_cur_B_1, s_next_B_1, embeds, seq_len):
|
|
v = net_velocity(net, x, s_cur_B_1, embeds, seq_len)
|
|
sc = s_cur_B_1.reshape(-1, 1, 1, 1, 1).to(x.dtype)
|
|
sn = s_next_B_1.reshape(-1, 1, 1, 1, 1).to(x.dtype)
|
|
return x + (sn - sc) * v.to(x.dtype)
|
|
|
|
def backward_simulation(prompt_embeds, size, n_steps, seq_len, with_grad=False, tdm=False):
|
|
"""Roll the student down an rf-Euler schedule; only the last segment carries grad."""
|
|
start_t = args.backward_simulation_start_t
|
|
if start_t is None or start_t < 0:
|
|
start_t = 1.0
|
|
end_t = args.backward_simulation_end_t
|
|
bt = list(args.backward_timesteps)
|
|
traj = [start_t]
|
|
if tdm:
|
|
for i in range(n_steps):
|
|
traj.append(bt[i] if i < len(bt) else end_t)
|
|
else:
|
|
for i in range(n_steps - 1):
|
|
traj.append(bt[i] if i < len(bt) else end_t)
|
|
traj.append(end_t)
|
|
|
|
# Init latent = pure noise; the low-noise expert first rolls the frozen high student to start_t.
|
|
x = torch.randn(size, device=accelerator.device, generator=torch_rng, dtype=weight_dtype)
|
|
if init_rollout_transformer3d is not None and start_t < 1.0:
|
|
s_traj = [1.0, *list(args.init_rollout_timesteps), start_t]
|
|
with torch.no_grad():
|
|
for s_cur, s_next in zip(s_traj[:-1], s_traj[1:]):
|
|
sc = torch.full((size[0], 1), s_cur, device=accelerator.device, dtype=torch.float32)
|
|
sn = torch.full((size[0], 1), s_next, device=accelerator.device, dtype=torch.float32)
|
|
x = _euler(init_rollout_transformer3d, x, sc, sn, prompt_embeds, seq_len)
|
|
x = x.detach()
|
|
|
|
B = size[0]
|
|
t_ts = [torch.full((B, 1), float(v), device=accelerator.device, dtype=torch.float32) for v in traj]
|
|
n_seg = len(traj) - 1
|
|
for step in range(n_seg):
|
|
ctx_fn = torch.enable_grad() if (with_grad and step == n_seg - 1) else torch.no_grad()
|
|
with ctx_fn:
|
|
x = _euler(generator_transformer3d, x, t_ts[step], t_ts[step + 1], prompt_embeds, seq_len)
|
|
return x, t_ts, traj
|
|
|
|
# Teacher rf-time grid for ode_match div: K+1 descending sigmas from sigma_max down toward 0.
|
|
def _build_div_teacher_schedule(K, sigma_max, shift):
|
|
N_INF = 30
|
|
assert K + 1 <= N_INF, f"div_K + 1 = {K+1} must be <= {N_INF}"
|
|
sigma_max_rf = sigma_max / (sigma_max + 1.0)
|
|
unshifted = inv_shift_rf_time(sigma_max_rf, shift)
|
|
raw = torch.linspace(unshifted, 0.0, N_INF + 1, device=accelerator.device, dtype=torch.float32)[:-1]
|
|
sig = shift_rf_time(raw, shift)
|
|
return sig[:K + 1]
|
|
|
|
def is_student_phase(it):
|
|
if args.loss_scale_tdm <= 0 and args.loss_scale_dmd <= 0:
|
|
return True
|
|
if it < args.critic_warmup:
|
|
return True
|
|
return (it - args.critic_warmup) % args.student_update_freq == 0
|
|
|
|
def num_sim_steps(eff):
|
|
if args.loss_scale_div > 0:
|
|
return eff % (args.max_simulation_steps_fake - 1) + 2
|
|
return eff % args.max_simulation_steps_fake + 1
|
|
|
|
# Patch-size for the student, used to compute seq_len once per batch.
|
|
_patch = accelerator.unwrap_model(generator_transformer3d).config.patch_size[1:]
|
|
neg_txt_default = "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走"
|
|
|
|
for epoch in range(first_epoch, args.num_train_epochs):
|
|
train_student = 0.0
|
|
train_div = 0.0
|
|
train_critic = 0.0
|
|
train_gen_cnt = 0
|
|
batch_sampler.sampler.generator = torch.Generator().manual_seed((args.seed if args.seed is not None else 0) + epoch)
|
|
for step, batch in enumerate(train_dataloader):
|
|
iteration = global_step
|
|
student_phase = is_student_phase(iteration)
|
|
|
|
with torch.cuda.amp.autocast(dtype=weight_dtype), torch.cuda.device(device=accelerator.device):
|
|
# --- Build the Target Latent Shape (T2V bucketing, same as the SFT path) ---
|
|
text = batch['text']
|
|
if args.fix_sample_size is not None:
|
|
local_sample_size = [int(x / spatial_compression_ratio / 2) * spatial_compression_ratio * 2 for x in args.fix_sample_size]
|
|
num_frames = args.video_sample_n_frames
|
|
else:
|
|
if args.random_hw_adapt and args.training_with_video_token_length:
|
|
target_token_length = args.video_sample_n_frames * args.token_sample_size * args.token_sample_size
|
|
length_to_frame_num = get_length_to_frame_num(target_token_length)
|
|
local_length = np.random.choice(list(length_to_frame_num.keys())) if rng is None else rng.choice(list(length_to_frame_num.keys()))
|
|
num_frames = length_to_frame_num[local_length]
|
|
aspect_ratio_sample_size = {key: [x / 512 * local_length for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()}
|
|
aspect_ratio_key = np.random.choice(list(aspect_ratio_sample_size.keys())) if rng is None else rng.choice(list(aspect_ratio_sample_size.keys()))
|
|
local_sample_size = aspect_ratio_sample_size[aspect_ratio_key]
|
|
else:
|
|
num_frames = args.video_sample_n_frames
|
|
aspect_ratio_sample_size = {key: [x / 512 * args.video_sample_size for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()}
|
|
aspect_ratio_key = np.random.choice(list(aspect_ratio_sample_size.keys())) if rng is None else rng.choice(list(aspect_ratio_sample_size.keys()))
|
|
local_sample_size = aspect_ratio_sample_size[aspect_ratio_key]
|
|
local_sample_size = [int(x / spatial_compression_ratio / 2) * spatial_compression_ratio * 2 for x in local_sample_size]
|
|
target_shape = (
|
|
len(text),
|
|
vae.latent_channels,
|
|
int((num_frames - 1) // vae.temporal_compression_ratio + 1),
|
|
int(local_sample_size[0] // vae.spatial_compression_ratio),
|
|
int(local_sample_size[1] // vae.spatial_compression_ratio),
|
|
)
|
|
bsz, channel, nfr, height, width = target_shape
|
|
seq_len = math.ceil((width * height) / (_patch[0] * _patch[1]) * nfr)
|
|
|
|
# --- Optional Real-Latent Anchor (only needed by the pcm diversity modes) ---
|
|
anchor_x0 = None
|
|
if args.loss_scale_div > 0 and args.div_mode in ("pcm_fixed", "pcm_random") \
|
|
and 'pixel_values' in batch and batch.get('pixel_values', None) is not None:
|
|
if manual_vae_move:
|
|
vae.to(accelerator.device)
|
|
pixel_values = rearrange(batch["pixel_values"].to(accelerator.device, dtype=weight_dtype, non_blocking=True), "b f c h w -> b c f h w")
|
|
with torch.no_grad():
|
|
bs = args.vae_mini_batch
|
|
enc = []
|
|
for i in range(0, pixel_values.shape[0], bs):
|
|
enc.append(vae.encode(pixel_values[i:i + bs])[0].sample())
|
|
anchor_x0 = torch.cat(enc, dim=0)
|
|
target_shape = anchor_x0.size()
|
|
bsz, channel, nfr, height, width = target_shape
|
|
seq_len = math.ceil((width * height) / (_patch[0] * _patch[1]) * nfr)
|
|
if manual_vae_move:
|
|
vae.to('cpu')
|
|
torch.cuda.empty_cache()
|
|
|
|
# --- Text Embeddings (cond + uncond) ---
|
|
if args.enable_text_encoder_in_dataloader:
|
|
prompt_embeds = batch['encoder_hidden_states'].to(device=accelerator.device)
|
|
neg_prompt_embeds = batch['neg_encoder_hidden_states'].to(device=accelerator.device)
|
|
else:
|
|
if manual_te_move:
|
|
text_encoder.to(accelerator.device)
|
|
with torch.no_grad():
|
|
prompt_ids = tokenizer(
|
|
text,
|
|
padding="max_length",
|
|
max_length=args.tokenizer_max_length,
|
|
truncation=True,
|
|
add_special_tokens=True,
|
|
return_tensors="pt"
|
|
)
|
|
seq_lens = prompt_ids.attention_mask.gt(0).sum(dim=1).long()
|
|
prompt_embeds = text_encoder(prompt_ids.input_ids.to(accelerator.device), attention_mask=prompt_ids.attention_mask.to(accelerator.device))[0]
|
|
prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)]
|
|
|
|
neg_ids = tokenizer(
|
|
[neg_txt_default for _ in text],
|
|
padding="max_length",
|
|
max_length=args.tokenizer_max_length,
|
|
truncation=True,
|
|
add_special_tokens=True,
|
|
return_tensors="pt"
|
|
)
|
|
neg_seq_lens = neg_ids.attention_mask.gt(0).sum(dim=1).long()
|
|
neg_prompt_embeds = text_encoder(neg_ids.input_ids.to(accelerator.device), attention_mask=neg_ids.attention_mask.to(accelerator.device))[0]
|
|
neg_prompt_embeds = [u[:v] for u, v in zip(neg_prompt_embeds, neg_seq_lens)]
|
|
if manual_te_move:
|
|
text_encoder.to('cpu')
|
|
torch.cuda.empty_cache()
|
|
|
|
if student_phase:
|
|
with accelerator.accumulate(generator_transformer3d), \
|
|
torch.cuda.amp.autocast(dtype=weight_dtype), torch.cuda.device(device=accelerator.device):
|
|
eff_it = iteration if iteration < args.critic_warmup else args.critic_warmup + (iteration - args.critic_warmup) // args.student_update_freq
|
|
n_steps = num_sim_steps(eff_it)
|
|
use_tdm = (args.distill_loss == "tdm")
|
|
|
|
# --- Student Distribution-Matching Loss (TDM or DMD) + optional diversity term ---
|
|
G, t_ts, traj = backward_simulation(
|
|
prompt_embeds,
|
|
target_shape,
|
|
n_steps,
|
|
seq_len,
|
|
with_grad=True,
|
|
tdm=use_tdm
|
|
)
|
|
s_im1 = t_ts[-1] if use_tdm else torch.zeros_like(t_ts[-1]) # x0 (t_{i-1}=0) for DMD
|
|
|
|
# Comparison time s_D ~ p_D, drawn strictly above the last sim step.
|
|
D_time = sync_broadcast(draw_rf_time((target_shape[0], 1), lower_bound=s_im1, strict_lower=use_tdm)).to(accelerator.device)
|
|
eps = torch.randn(target_shape, device=accelerator.device, generator=torch_rng, dtype=weight_dtype)
|
|
sD = D_time.reshape(-1, 1, 1, 1, 1).to(torch.float64)
|
|
base = G.float() # fp32 halves this latent's footprint
|
|
|
|
# Re-noise x0=G up to s_D (DMD: rf interp; TDM: bridge conditioned on t_{i-1}).
|
|
if use_tdm:
|
|
s1 = s_im1.reshape(-1, 1, 1, 1, 1).to(torch.float64)
|
|
om = (1.0 - s1).clamp(min=1e-6)
|
|
sigma_cond = (((sD - s1) * (sD + s1 - 2.0 * sD * s1)).clamp(min=0.0)).sqrt() / om
|
|
D_xt = (((1.0 - sD) / om) * base + sigma_cond * eps.double()).to(weight_dtype)
|
|
else:
|
|
D_xt = ((1.0 - sD) * base + sD * eps.double()).to(weight_dtype)
|
|
|
|
# Critic (fake_score) and CFG teacher both denoise D_xt to x0.
|
|
with torch.no_grad():
|
|
x0_fake = net_x0(
|
|
fake_score_transformer3d,
|
|
D_xt,
|
|
D_time,
|
|
prompt_embeds,
|
|
seq_len
|
|
)[0]
|
|
x0_teacher = net_x0(
|
|
real_score_transformer3d,
|
|
D_xt,
|
|
D_time,
|
|
prompt_embeds,
|
|
seq_len
|
|
)[0]
|
|
if args.real_guidance_scale and args.real_guidance_scale > 1.0:
|
|
x0_uncond = net_x0(
|
|
real_score_transformer3d,
|
|
D_xt,
|
|
D_time,
|
|
neg_prompt_embeds,
|
|
seq_len
|
|
)[0]
|
|
x0_teacher = x0_uncond + args.real_guidance_scale * (x0_teacher - x0_uncond)
|
|
if use_tdm:
|
|
# TDM matches in x_{t_{i-1}} space: reproject each x0 pred to s_im1 via its implied eps.
|
|
def to_t_im1(x_t, x0_pred):
|
|
x_t = x_t.double()
|
|
eps_pred = (x_t - (1.0 - sD) * x0_pred) / sD.clamp(min=1e-6)
|
|
return (1.0 - s1) * x0_pred + s1 * eps_pred
|
|
tgt_fake = to_t_im1(D_xt, x0_fake)
|
|
tgt_tea = to_t_im1(D_xt, x0_teacher)
|
|
else:
|
|
tgt_fake, tgt_tea = x0_fake, x0_teacher
|
|
# Normalized score difference: (fake - teacher) / per-sample |G - teacher|_1.
|
|
wf = (base - tgt_tea).abs().mean(dim=[1, 2, 3, 4], keepdim=True).clamp(min=1e-5)
|
|
grad = (tgt_fake - tgt_tea) / wf
|
|
|
|
# DMD surrogate loss
|
|
surrogate = (base - (base - grad).detach()) ** 2
|
|
surrogate[torch.isnan(surrogate).flatten(start_dim=1).any(dim=1)] = 0
|
|
loss_dm = surrogate.sum(dim=(1, 2, 3, 4))
|
|
scale = args.loss_scale_tdm if use_tdm else args.loss_scale_dmd
|
|
dm_term = scale * loss_dm
|
|
|
|
# Diversity term (ode_match / pcm); zeros_like (no graph) when div<=0 or n_steps<2.
|
|
div_term = torch.zeros_like(dm_term)
|
|
if args.loss_scale_div > 0 and n_steps >= 2:
|
|
if args.div_mode == "ode_match":
|
|
# ODE-match: student 1-step velocity vs teacher ODE rollout (no_grad, CFG).
|
|
s_grid = _build_div_teacher_schedule(args.div_K, args.div_sigma_max, args.teacher_timestep_shift)
|
|
K = s_grid.shape[0] - 1
|
|
eps = sync_broadcast(torch.randn(target_shape, device=accelerator.device, generator=torch_rng, dtype=weight_dtype))
|
|
x = eps
|
|
with torch.no_grad():
|
|
for k in range(K):
|
|
sc = s_grid[k].view(1, 1).expand(target_shape[0], 1).contiguous()
|
|
sn = s_grid[k + 1].view(1, 1).expand(target_shape[0], 1).contiguous()
|
|
v = net_velocity(
|
|
real_score_transformer3d,
|
|
x,
|
|
sc,
|
|
prompt_embeds,
|
|
seq_len
|
|
).to(weight_dtype)
|
|
if args.real_guidance_scale and args.real_guidance_scale > 1.0:
|
|
vu = net_velocity(
|
|
real_score_transformer3d,
|
|
x,
|
|
sc,
|
|
neg_prompt_embeds,
|
|
seq_len
|
|
).to(weight_dtype)
|
|
v = vu + args.real_guidance_scale * (v - vu)
|
|
dt = (sn - sc).reshape(-1, 1, 1, 1, 1).to(x.dtype)
|
|
x = x + dt * v
|
|
t_tilde = float(s_grid[-1].item())
|
|
d = 1.0 - t_tilde
|
|
assert d > 1e-6, f"1 - t_tilde too small: {d}"
|
|
v_tilde = (eps - x) / d
|
|
s1 = torch.ones(target_shape[0], 1, device=accelerator.device, dtype=torch.float32)
|
|
v_stu = net_velocity(
|
|
generator_transformer3d,
|
|
eps,
|
|
s1,
|
|
prompt_embeds,
|
|
seq_len
|
|
)
|
|
diff = v_stu.double() - v_tilde.double()
|
|
div_loss = (diff ** 2).sum(dim=(1, 2, 3, 4))
|
|
div_loss[torch.isnan(div_loss).flatten(start_dim=1).any(dim=1)] = 0
|
|
div_term = args.loss_scale_div * div_loss
|
|
elif args.div_mode in ("pcm_fixed", "pcm_random"):
|
|
# PCM consistency: grad student step must match the no_grad teacher rollout target.
|
|
B = target_shape[0]
|
|
K = args.dcm_skipping_interval_steps
|
|
dw = 1.0 / args.dcm_total_steps
|
|
shift = args.teacher_timestep_shift
|
|
eps = sync_broadcast(torch.randn(target_shape, device=accelerator.device, generator=torch_rng, dtype=weight_dtype))
|
|
rf_b = float(args.backward_timesteps[0]) if len(args.backward_timesteps) > 0 else float(args.backward_simulation_end_t)
|
|
s_b = inv_shift_rf_time(rf_b, shift)
|
|
w_right = 1.0 - s_b
|
|
# phase-0 offset w in [0, valid_width]; pcm_fixed pins w=0, pcm_random samples it.
|
|
valid_width = w_right - K * dw
|
|
if args.div_mode == "pcm_fixed":
|
|
w = torch.zeros(B, 1, device=accelerator.device, dtype=torch.float32)
|
|
else:
|
|
assert valid_width > 0, (
|
|
f"pcm-div phase 0 valid width <= 0: w_right={w_right:.4f}, K*dw={K*dw:.4f}.")
|
|
w = sync_broadcast(torch.rand(B, 1, device=accelerator.device) * valid_width)
|
|
t_list = [shift_rf_time(1.0 - (w + k * dw), shift) for k in range(K + 1)]
|
|
t0, tK = t_list[0], t_list[-1]
|
|
t_bnd = torch.full((B, 1), rf_b, device=accelerator.device, dtype=torch.float32)
|
|
t0c = t0.reshape(-1, 1, 1, 1, 1).double()
|
|
anchor = anchor_x0.double() if anchor_x0 is not None else G.detach().double()
|
|
xt = ((1.0 - t0c) * anchor + t0c * eps.double()).to(weight_dtype)
|
|
# Run the no_grad teacher rollout + target student BEFORE the grad forward (FSDP safety).
|
|
with torch.no_grad():
|
|
xk = xt
|
|
for k in range(K):
|
|
tk, tk1 = t_list[k], t_list[k + 1]
|
|
dt = (tk - tk1).reshape(-1, 1, 1, 1, 1).to(xk.dtype)
|
|
F_tea = net_velocity(
|
|
real_score_transformer3d,
|
|
xk,
|
|
tk,
|
|
prompt_embeds,
|
|
seq_len
|
|
).to(xk.dtype)
|
|
if args.real_guidance_scale and args.real_guidance_scale > 1.0:
|
|
F_unc = net_velocity(
|
|
real_score_transformer3d,
|
|
xk,
|
|
tk,
|
|
neg_prompt_embeds,
|
|
seq_len
|
|
).to(xk.dtype)
|
|
F_tea = F_unc + args.real_guidance_scale * (F_tea - F_unc)
|
|
xk = xk - dt * F_tea
|
|
x_target = _euler(
|
|
generator_transformer3d,
|
|
xk,
|
|
tK,
|
|
t_bnd,
|
|
prompt_embeds,
|
|
seq_len
|
|
).detach()
|
|
x_pred = _euler(
|
|
generator_transformer3d,
|
|
xt,
|
|
t0,
|
|
t_bnd,
|
|
prompt_embeds,
|
|
seq_len
|
|
)
|
|
div_loss = (x_pred.double() - x_target.double()) ** 2
|
|
div_loss[torch.isnan(div_loss).flatten(start_dim=1).any(dim=1)] = 0
|
|
div_term = args.loss_scale_div * div_loss.sum(dim=(1, 2, 3, 4))
|
|
|
|
avg_dm = accelerator.gather(dm_term.detach().mean().repeat(args.train_batch_size)).mean()
|
|
avg_div = accelerator.gather(div_term.detach().mean().repeat(args.train_batch_size)).mean()
|
|
train_student += avg_dm.item() / args.gradient_accumulation_steps
|
|
train_div += avg_div.item() / args.gradient_accumulation_steps
|
|
train_gen_cnt += 1
|
|
|
|
# Single combined backward: a second accelerator.backward() on the same FSDP module
|
|
# trips the "flat_param on cpu" assert under CPUOffload, so dm+div sum into one call.
|
|
student_loss = dm_term.mean()
|
|
if div_term.requires_grad: # div_term has no graph when div<=0 or n_steps<2
|
|
student_loss = student_loss + div_term.mean()
|
|
accelerator.backward(student_loss)
|
|
del dm_term, div_term, student_loss
|
|
if accelerator.sync_gradients and args.max_grad_norm and args.max_grad_norm > 0:
|
|
accelerator.clip_grad_norm_(trainable_params, args.max_grad_norm)
|
|
optimizer.step()
|
|
lr_scheduler.step()
|
|
optimizer.zero_grad()
|
|
sync_now = accelerator.sync_gradients
|
|
active_lr = lr_scheduler
|
|
else:
|
|
with accelerator_fake_score_transformer3d.accumulate(fake_score_transformer3d), \
|
|
torch.cuda.amp.autocast(dtype=weight_dtype), torch.cuda.device(device=accelerator.device):
|
|
eff_it = iteration if iteration < args.critic_warmup else args.critic_warmup + (iteration - args.critic_warmup) // args.student_update_freq
|
|
n_steps = num_sim_steps(iteration - eff_it - 1)
|
|
|
|
# --- Critic (fake_score) Denoising Loss on student x0, x0-space weighted ---
|
|
with torch.no_grad():
|
|
G, _, _ = backward_simulation(
|
|
prompt_embeds,
|
|
target_shape,
|
|
n_steps,
|
|
seq_len,
|
|
with_grad=False,
|
|
tdm=False
|
|
)
|
|
D_time = sync_broadcast(draw_rf_time((target_shape[0], 1))).to(accelerator.device)
|
|
eps = torch.randn(target_shape, device=accelerator.device, generator=torch_rng, dtype=weight_dtype)
|
|
sD = D_time.reshape(-1, 1, 1, 1, 1).to(torch.float64)
|
|
D_xt = ((1.0 - sD) * G.double() + sD * eps.double()).to(weight_dtype)
|
|
x0_fake = net_x0(
|
|
fake_score_transformer3d,
|
|
D_xt,
|
|
D_time,
|
|
prompt_embeds,
|
|
seq_len
|
|
)[0]
|
|
|
|
# x0-space weight (s_D^2+(1-s_D)^2)/s_D^2 keeps the MSE comparable across s_D.
|
|
weight = (sD ** 2 + (1.0 - sD) ** 2) / sD.clamp(min=1e-6) ** 2
|
|
crit = (((G.double() - x0_fake) ** 2 * weight).sum(dim=(1, 2, 3, 4))).mean()
|
|
|
|
avg_c = accelerator.gather(crit.detach().repeat(args.train_batch_size)).mean()
|
|
train_critic += avg_c.item() / args.gradient_accumulation_steps
|
|
accelerator_fake_score_transformer3d.backward(crit)
|
|
if accelerator_fake_score_transformer3d.sync_gradients and args.max_grad_norm and args.max_grad_norm > 0:
|
|
accelerator_fake_score_transformer3d.clip_grad_norm_(fake_trainable_params, args.max_grad_norm)
|
|
critic_optimizer.step()
|
|
fake_score_lr_scheduler.step()
|
|
critic_optimizer.zero_grad()
|
|
sync_now = accelerator_fake_score_transformer3d.sync_gradients
|
|
active_lr = fake_score_lr_scheduler
|
|
|
|
if sync_now:
|
|
progress_bar.update(1)
|
|
global_step += 1
|
|
tracker_logs = {
|
|
"train_student_loss": train_student / max(train_gen_cnt, 1),
|
|
"train_div_loss": train_div / max(train_gen_cnt, 1),
|
|
"train_critic_loss": train_critic,
|
|
}
|
|
accelerator.log(tracker_logs, step=global_step)
|
|
train_student = 0.0
|
|
train_div = 0.0
|
|
train_critic = 0.0
|
|
train_gen_cnt = 0
|
|
|
|
if global_step % args.checkpointing_steps == 0:
|
|
if args.use_deepspeed or args.use_fsdp or accelerator.is_main_process:
|
|
if args.checkpoints_total_limit is not None:
|
|
checkpoints = os.listdir(args.output_dir)
|
|
checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
|
|
checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
|
|
if len(checkpoints) >= args.checkpoints_total_limit:
|
|
num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
|
|
for removing_checkpoint in checkpoints[0:num_to_remove]:
|
|
shutil.rmtree(os.path.join(args.output_dir, removing_checkpoint))
|
|
gc.collect()
|
|
torch.cuda.empty_cache()
|
|
torch.cuda.ipc_collect()
|
|
save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
|
|
with progress_bar.paused():
|
|
fake_score_save_path = os.path.join(save_path, "fake_score")
|
|
accelerator.save_state(save_path)
|
|
accelerator_fake_score_transformer3d.save_state(fake_score_save_path)
|
|
gc.collect()
|
|
torch.cuda.empty_cache()
|
|
torch.cuda.ipc_collect()
|
|
logger.info(f"Saved state to {save_path}")
|
|
|
|
if args.validation_prompts is not None and global_step % args.validation_steps == 0:
|
|
with progress_bar.paused():
|
|
log_validation(
|
|
vae, text_encoder, tokenizer, generator_transformer3d,
|
|
args, config, accelerator, weight_dtype, global_step,
|
|
)
|
|
|
|
logs = {"train_critic_loss": train_critic, "lr": active_lr.get_last_lr()[0], "phase": "student" if student_phase else "critic"}
|
|
if student_phase:
|
|
logs["train_student_loss"] = train_student
|
|
progress_bar.set_postfix(**logs)
|
|
|
|
if global_step >= args.max_train_steps:
|
|
break
|
|
|
|
if args.validation_prompts is not None and epoch % args.validation_epochs == 0:
|
|
with progress_bar.paused():
|
|
log_validation(
|
|
vae, text_encoder, tokenizer, generator_transformer3d,
|
|
args, config, accelerator, weight_dtype, global_step,
|
|
)
|
|
|
|
# Close the bar before the end-of-run checkpoint.
|
|
progress_bar.close()
|
|
|
|
# Create the pipeline using the trained modules and save it.
|
|
accelerator.wait_for_everyone()
|
|
if accelerator.is_main_process:
|
|
generator_transformer3d = unwrap_model(generator_transformer3d)
|
|
|
|
if args.use_deepspeed or args.use_fsdp or accelerator.is_main_process:
|
|
gc.collect()
|
|
torch.cuda.empty_cache()
|
|
torch.cuda.ipc_collect()
|
|
save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
|
|
fake_score_save_path = os.path.join(save_path, "fake_score")
|
|
accelerator.save_state(save_path)
|
|
accelerator_fake_score_transformer3d.save_state(fake_score_save_path)
|
|
logger.info(f"Saved state to {save_path}")
|
|
|
|
accelerator.end_training()
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|