2455 lines
110 KiB
Python
2455 lines
110 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 hashlib
|
|
import inspect
|
|
import logging
|
|
import math
|
|
import os
|
|
import pickle
|
|
import random
|
|
import shutil
|
|
import sys
|
|
from typing import (Any, Callable, Dict, List, NamedTuple, Optional, Tuple,
|
|
Union)
|
|
|
|
import accelerate
|
|
import diffusers
|
|
import numpy as np
|
|
import torch
|
|
import torch.nn.functional as F
|
|
import torch.utils.checkpoint
|
|
import transformers
|
|
from accelerate import Accelerator
|
|
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 omegaconf import OmegaConf
|
|
from packaging import version
|
|
from torch.utils.data import BatchSampler, RandomSampler
|
|
from torch.utils.tensorboard import SummaryWriter
|
|
from tqdm.auto import tqdm
|
|
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.bucket_sampler import ASPECT_RATIO_512, RandomSampler
|
|
from videox_fun.data.dataset_image_video import TextDataset
|
|
from videox_fun.dist import set_multi_gpus_devices, shard_model
|
|
from videox_fun.models import (AutoencoderKL, AutoTokenizer,
|
|
CLIPImageProcessor,
|
|
CLIPVisionModelWithProjection,
|
|
Qwen2_5_VLForConditionalGeneration,
|
|
Qwen2Tokenizer, Qwen3ForCausalLM,
|
|
QwenImageTransformer2DModel,
|
|
ZImageTransformer2DModel)
|
|
from videox_fun.pipeline import ZImagePipeline
|
|
from videox_fun.utils.discrete_sampler import DiscreteSampling
|
|
from videox_fun.utils.lora_utils import (convert_peft_lora_to_kohya_lora,
|
|
create_network, merge_lora,
|
|
unmerge_lora)
|
|
from videox_fun.utils.sd3_sde_with_logprob import sde_step_with_logprob
|
|
from videox_fun.utils.tqdm_bar import PauseAwareTqdm
|
|
from videox_fun.utils.utils import 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 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 calculate_shift(
|
|
image_seq_len,
|
|
base_seq_len: int = 256,
|
|
max_seq_len: int = 4096,
|
|
base_shift: float = 0.5,
|
|
max_shift: float = 1.15,
|
|
):
|
|
m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
|
|
b = base_shift - m * base_seq_len
|
|
mu = image_seq_len * m + b
|
|
return mu
|
|
|
|
def encode_prompt(
|
|
prompt: Union[str, List[str]],
|
|
device: Optional[torch.device] = None,
|
|
text_encoder = None,
|
|
tokenizer = None,
|
|
max_sequence_length: int = 512,
|
|
) -> List[torch.FloatTensor]:
|
|
if isinstance(prompt, str):
|
|
prompt = [prompt]
|
|
|
|
for i, prompt_item in enumerate(prompt):
|
|
messages = [
|
|
{"role": "user", "content": prompt_item},
|
|
]
|
|
prompt_item = tokenizer.apply_chat_template(
|
|
messages,
|
|
tokenize=False,
|
|
add_generation_prompt=True,
|
|
enable_thinking=True,
|
|
)
|
|
prompt[i] = prompt_item
|
|
|
|
text_inputs = tokenizer(
|
|
prompt,
|
|
padding="max_length",
|
|
max_length=max_sequence_length,
|
|
truncation=True,
|
|
return_tensors="pt",
|
|
)
|
|
|
|
text_input_ids = text_inputs.input_ids.to(device)
|
|
prompt_masks = text_inputs.attention_mask.to(device).bool()
|
|
|
|
prompt_embeds = text_encoder(
|
|
input_ids=text_input_ids,
|
|
attention_mask=prompt_masks,
|
|
output_hidden_states=True,
|
|
).hidden_states[-2]
|
|
|
|
embeddings_list = []
|
|
|
|
for i in range(len(prompt_embeds)):
|
|
embeddings_list.append(prompt_embeds[i][prompt_masks[i]])
|
|
|
|
return embeddings_list
|
|
|
|
# Fallback implementation if flow_grpo is not available
|
|
class PerPromptStatTracker:
|
|
def __init__(self, global_std=False):
|
|
self.global_std = global_std
|
|
self.stats = {}
|
|
self.history_prompts = set()
|
|
|
|
def update(self, prompts, rewards, type='grpo'):
|
|
prompts = np.array(prompts)
|
|
rewards = np.array(rewards, dtype=np.float64)
|
|
unique = np.unique(prompts)
|
|
advantages = np.empty_like(rewards) * 0.0
|
|
for prompt in unique:
|
|
prompt_rewards = rewards[prompts == prompt]
|
|
if prompt not in self.stats:
|
|
self.stats[prompt] = []
|
|
self.stats[prompt].extend(prompt_rewards)
|
|
self.history_prompts.add(hash(prompt))
|
|
for prompt in unique:
|
|
self.stats[prompt] = np.stack(self.stats[prompt])
|
|
prompt_rewards = rewards[prompts == prompt]
|
|
mean = np.mean(self.stats[prompt], axis=0, keepdims=True)
|
|
if self.global_std:
|
|
std = np.std(rewards, axis=0, keepdims=True) + 1e-4
|
|
else:
|
|
std = np.std(self.stats[prompt], axis=0, keepdims=True) + 1e-4
|
|
if type == 'grpo':
|
|
advantages[prompts == prompt] = (prompt_rewards - mean) / std
|
|
return advantages
|
|
|
|
def get_stats(self):
|
|
avg_group_size = sum(len(v) for v in self.stats.values()) / len(self.stats) if self.stats else 0
|
|
history_prompts = len(self.history_prompts)
|
|
return avg_group_size, history_prompts
|
|
|
|
def clear(self):
|
|
self.stats = {}
|
|
|
|
def calculate_zero_std_ratio(prompts, gathered_rewards):
|
|
"""Calculate the proportion of unique prompts whose reward standard deviation is zero.
|
|
|
|
Args:
|
|
prompts: List of prompts.
|
|
gathered_rewards: Dictionary containing rewards, must include the key 'ori_avg'.
|
|
|
|
Returns:
|
|
zero_std_ratio: Proportion of prompts with zero standard deviation.
|
|
prompt_std_devs: Mean standard deviation across all unique prompts.
|
|
"""
|
|
prompt_array = np.array(prompts)
|
|
unique_prompts, inverse_indices, counts = np.unique(
|
|
prompt_array, return_inverse=True, return_counts=True
|
|
)
|
|
grouped_rewards = gathered_rewards['ori_avg'][np.argsort(inverse_indices)]
|
|
split_indices = np.cumsum(counts)[:-1]
|
|
reward_groups = np.split(grouped_rewards, split_indices)
|
|
prompt_std_devs = np.array([np.std(group) for group in reward_groups])
|
|
zero_std_count = np.count_nonzero(prompt_std_devs == 0)
|
|
zero_std_ratio = zero_std_count / len(prompt_std_devs)
|
|
return zero_std_ratio, prompt_std_devs.mean()
|
|
|
|
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
|
|
def retrieve_timesteps(
|
|
scheduler,
|
|
num_inference_steps: Optional[int] = None,
|
|
device: Optional[Union[str, torch.device]] = None,
|
|
timesteps: Optional[List[int]] = None,
|
|
sigmas: Optional[List[float]] = None,
|
|
**kwargs,
|
|
):
|
|
r"""
|
|
Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
|
|
custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
|
|
|
|
Args:
|
|
scheduler (`SchedulerMixin`):
|
|
The scheduler to get timesteps from.
|
|
num_inference_steps (`int`):
|
|
The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
|
|
must be `None`.
|
|
device (`str` or `torch.device`, *optional*):
|
|
The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
|
|
timesteps (`List[int]`, *optional*):
|
|
Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
|
|
`num_inference_steps` and `sigmas` must be `None`.
|
|
sigmas (`List[float]`, *optional*):
|
|
Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
|
|
`num_inference_steps` and `timesteps` must be `None`.
|
|
|
|
Returns:
|
|
`Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
|
|
second element is the number of inference steps.
|
|
"""
|
|
if timesteps is not None and sigmas is not None:
|
|
raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
|
|
if timesteps is not None:
|
|
accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
|
|
if not accepts_timesteps:
|
|
raise ValueError(
|
|
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
|
|
f" timestep schedules. Please check whether you are using the correct scheduler."
|
|
)
|
|
scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
|
|
timesteps = scheduler.timesteps
|
|
num_inference_steps = len(timesteps)
|
|
elif sigmas is not None:
|
|
accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
|
|
if not accept_sigmas:
|
|
raise ValueError(
|
|
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
|
|
f" sigmas schedules. Please check whether you are using the correct scheduler."
|
|
)
|
|
scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
|
|
timesteps = scheduler.timesteps
|
|
num_inference_steps = len(timesteps)
|
|
else:
|
|
scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
|
|
timesteps = scheduler.timesteps
|
|
return timesteps, num_inference_steps
|
|
|
|
@torch.no_grad()
|
|
def sample_with_cfg(
|
|
model,
|
|
vae,
|
|
noise,
|
|
prompt_embeds,
|
|
neg_prompt_embeds,
|
|
num_steps=25,
|
|
cfg_scale=7.5,
|
|
noise_scheduler=None,
|
|
device='cuda',
|
|
dtype=torch.float32,
|
|
noise_level: float = 0.7,
|
|
sde_window_size: int = 0,
|
|
sde_window_range: tuple[int, int] = (0, 5),
|
|
noise_generator=None,
|
|
diffusion_clip=False,
|
|
diffusion_clip_value=0.45,
|
|
):
|
|
# noise_generator: rank-aware torch.Generator driving the rollout stochasticity, i.e. the
|
|
# SDE window start and the per-step variance noise. When it is None the previous behaviour
|
|
# is kept (draws from the process-global `random` / torch RNGs).
|
|
batch_size = noise.shape[0]
|
|
# The noise/latents here are 5D (B, C, 1, H, W) -- a frame dim is inserted for the
|
|
# Wan-style wrapper -- so the spatial extent must be read from the last two axes.
|
|
# Indexing shape[2]/shape[3] read (1 // 2) * (H // 2) == 0, which pinned `mu` at the
|
|
# base value instead of the resolution-derived one. It is inert with the shipped Z-Image
|
|
# scheduler (`use_dynamic_shifting: false`, static `shift`, which ignores `mu`), but it
|
|
# silently under-shifts the schedule as soon as a dynamic-shifting scheduler is used.
|
|
image_seq_len = (noise.shape[-2] // 2) * (noise.shape[-1] // 2)
|
|
latents = noise.clone().to(torch.float32)
|
|
|
|
mu = calculate_shift(
|
|
image_seq_len,
|
|
noise_scheduler.config.get("base_image_seq_len", 256),
|
|
noise_scheduler.config.get("max_image_seq_len", 4096),
|
|
noise_scheduler.config.get("base_shift", 0.5),
|
|
noise_scheduler.config.get("max_shift", 1.15),
|
|
)
|
|
scheduler_kwargs = {"mu": mu}
|
|
# NOTE: do NOT set `noise_scheduler.sigma_min = 0.0` here, even though
|
|
# `diffusers.ZImagePipeline` / `videox_fun/pipeline/pipeline_z_image.py` do it before
|
|
# `set_timesteps`. It moves the grid from [..., 0.44151, 0.03478, 0.0] to
|
|
# [..., 0.42857, 0.0, 0.0], i.e. the last *executed* step ends up with sigma == 0, and
|
|
# `sde_step_with_logprob` divides by sigma in its `prev_sample_mean` term -> 0/0 = NaN,
|
|
# which poisons the latent that is decoded afterwards (black image). Inference is
|
|
# unaffected because `FlowMatchEulerDiscreteScheduler.step` never divides by sigma (its
|
|
# terminal sigma == 0 step is a plain dt == 0 no-op). The default grid already ends on
|
|
# sigma == 0 as `sigma_prev`, so the last real jump still lands on clean x0.
|
|
timesteps, num_inference_steps = retrieve_timesteps(
|
|
noise_scheduler,
|
|
num_steps,
|
|
device,
|
|
sigmas=None,
|
|
**scheduler_kwargs,
|
|
)
|
|
|
|
if sde_window_size > 0:
|
|
assert sde_window_range[1] - sde_window_size >= sde_window_range[0], (
|
|
f"sde_window_range {sde_window_range} 与 sde_window_size {sde_window_size} 不兼容，"
|
|
f"请保证 range[1] - window_size >= range[0]"
|
|
)
|
|
# Draw the window start from the rank-aware generator instead of the process-global
|
|
# `random` module: with a non device_specific `set_seed(args.seed)` every rank picked
|
|
# the same window at the same training step, so the whole cluster stacked its rollouts
|
|
# on one window simultaneously. The window length stays exactly `sde_window_size` on
|
|
# every rank, so the per-rank backward / gradient-sync counts match.
|
|
if noise_generator is not None:
|
|
start = int(torch.randint(
|
|
sde_window_range[0],
|
|
sde_window_range[1] - sde_window_size + 1,
|
|
(1,),
|
|
generator=noise_generator,
|
|
device=device,
|
|
).item())
|
|
else:
|
|
start = random.randint(
|
|
sde_window_range[0],
|
|
sde_window_range[1] - sde_window_size
|
|
)
|
|
end = start + sde_window_size
|
|
sde_window = (start, end)
|
|
else:
|
|
sde_window = (0, len(timesteps) - 1)
|
|
|
|
all_latents = []
|
|
all_log_probs = []
|
|
all_timesteps = []
|
|
|
|
apply_cfg = cfg_scale > 1.0
|
|
|
|
sampling_bar = tqdm(enumerate(timesteps), total=len(timesteps), desc="Sampling", leave=False, disable=not torch.distributed.is_initialized() or torch.distributed.get_rank() == 0)
|
|
for i, t in sampling_bar:
|
|
if i < sde_window[0]:
|
|
cur_noise_level = 0.0
|
|
elif i == sde_window[0]:
|
|
cur_noise_level = noise_level
|
|
all_latents.append(latents.clone())
|
|
elif sde_window[0] < i < sde_window[1]:
|
|
cur_noise_level = noise_level
|
|
else:
|
|
cur_noise_level = 0.0
|
|
|
|
timestep = t.expand(batch_size)
|
|
timestep_normalized = (1000 - timestep) / 1000.0
|
|
|
|
if apply_cfg:
|
|
latents_typed = latents.to(dtype)
|
|
latent_model_input = latents_typed.repeat(2, 1, 1, 1, 1)
|
|
prompt_embeds_input = prompt_embeds + neg_prompt_embeds
|
|
timestep_input = timestep_normalized.repeat(2)
|
|
else:
|
|
latent_model_input = latents.to(dtype)
|
|
prompt_embeds_input = prompt_embeds
|
|
timestep_input = timestep_normalized
|
|
|
|
latent_model_input_list = list(latent_model_input.unbind(dim=0))
|
|
model_out_list = model(
|
|
latent_model_input_list,
|
|
timestep_input,
|
|
prompt_embeds_input,
|
|
)[0]
|
|
|
|
if apply_cfg:
|
|
pos_out = model_out_list[:batch_size]
|
|
neg_out = model_out_list[batch_size:]
|
|
noise_pred = torch.stack(
|
|
[neg_out[j].float() + cfg_scale * (pos_out[j].float() - neg_out[j].float())
|
|
for j in range(batch_size)],
|
|
dim=0,
|
|
)
|
|
else:
|
|
noise_pred = torch.stack(
|
|
[out.float() for out in model_out_list], dim=0
|
|
)
|
|
|
|
noise_pred = -noise_pred # Sign convention (matches original code)
|
|
|
|
latents, log_prob, prev_latents_mean, std_dev_t = sde_step_with_logprob(
|
|
noise_scheduler,
|
|
noise_pred.float(),
|
|
t.unsqueeze(0).repeat(batch_size),
|
|
latents.float(),
|
|
noise_level=cur_noise_level,
|
|
diffusion_clip=diffusion_clip,
|
|
diffusion_clip_value=diffusion_clip_value,
|
|
# Rank-aware noise for the transition actually taken. The training side later
|
|
# re-evaluates log_prob of the *recorded* transition (prev_sample=next_latents),
|
|
# so no matching draw is needed there and on-policy consistency is preserved.
|
|
generator=noise_generator,
|
|
)
|
|
if sde_window[0] <= i < sde_window[1]:
|
|
all_latents.append(latents.clone())
|
|
all_log_probs.append(log_prob)
|
|
all_timesteps.append(t)
|
|
|
|
|
|
_latents = latents.to(vae.dtype).squeeze(2)
|
|
_latents = (_latents / vae.config.scaling_factor) + vae.config.shift_factor
|
|
images = vae.decode(_latents, return_dict=False)[0].float().unsqueeze(2)
|
|
images = (images / 2 + 0.5).clamp(0, 1)
|
|
|
|
ret = {
|
|
"latents" : latents.double(),
|
|
"all_latents" : all_latents, # List[Tensor], len = window_size+1
|
|
"all_log_probs" : all_log_probs, # List[Tensor], len = window_size
|
|
"all_timesteps" : all_timesteps, # List[Tensor], len = window_size
|
|
"prompt_embeds" : prompt_embeds,
|
|
"negative_prompt_embeds" : neg_prompt_embeds,
|
|
"prompt_embeds_mask" : None,
|
|
"negative_prompt_embeds_mask" : None,
|
|
"images" : images,
|
|
}
|
|
return ret
|
|
|
|
def compute_log_prob(
|
|
model,
|
|
vae,
|
|
sample,
|
|
step_idx,
|
|
noise_scheduler,
|
|
prompt_embeds,
|
|
neg_prompt_embeds,
|
|
cfg_scale=4.5,
|
|
noise_level=0.7,
|
|
dtype=torch.float32,
|
|
ref_model=None,
|
|
diffusion_clip=False,
|
|
diffusion_clip_value=0.45,
|
|
):
|
|
"""
|
|
Compute log probability for GRPO training.
|
|
|
|
Args:
|
|
model: The transformer model
|
|
vae: VAE for decoding (not used in log_prob computation)
|
|
sample: Dict containing latents, next_latents, timesteps
|
|
step_idx: Index of the timestep to compute log_prob for
|
|
noise_scheduler: The noise scheduler
|
|
prompt_embeds: Prompt embeddings (list of tensors)
|
|
neg_prompt_embeds: Negative prompt embeddings (list of tensors)
|
|
cfg_scale: Classifier-free guidance scale
|
|
noise_level: Noise level for SDE
|
|
dtype: Data type for computation
|
|
ref_model: Reference model for KL computation (optional)
|
|
|
|
Returns:
|
|
log_prob: Log probability of the transition
|
|
prev_sample_mean: Mean of the predicted previous sample
|
|
std_dev_t: Standard deviation at timestep t
|
|
sqrt_dt: sqrt(-dt), so the Gaussian variance is (std_dev_t * sqrt_dt) ** 2
|
|
ref_prev_sample_mean: Mean from reference model (if ref_model provided)
|
|
"""
|
|
batch_size = sample["latents"].shape[0]
|
|
latents = sample["latents"][:, step_idx] # [B, C, 1, H, W]
|
|
next_latents = sample["next_latents"][:, step_idx] # [B, C, 1, H, W]
|
|
timesteps = sample["timesteps"][:, step_idx] # [B]
|
|
|
|
apply_cfg = cfg_scale > 1.0
|
|
|
|
# Prepare timestep
|
|
timestep_normalized = (1000 - timesteps) / 1000.0
|
|
|
|
if apply_cfg:
|
|
latents_typed = latents.to(dtype)
|
|
latent_model_input = latents_typed.repeat(2, 1, 1, 1, 1)
|
|
prompt_embeds_input = prompt_embeds + neg_prompt_embeds
|
|
timestep_input = timestep_normalized.repeat(2)
|
|
else:
|
|
latent_model_input = latents.to(dtype)
|
|
prompt_embeds_input = prompt_embeds
|
|
timestep_input = timestep_normalized
|
|
|
|
# Forward pass through transformer
|
|
latent_model_input_list = list(latent_model_input.unbind(dim=0))
|
|
model_out_list = model(
|
|
latent_model_input_list,
|
|
timestep_input,
|
|
prompt_embeds_input,
|
|
)[0]
|
|
|
|
if apply_cfg:
|
|
pos_out = model_out_list[:batch_size]
|
|
neg_out = model_out_list[batch_size:]
|
|
noise_pred = torch.stack(
|
|
[neg_out[j].float() + cfg_scale * (pos_out[j].float() - neg_out[j].float())
|
|
for j in range(batch_size)],
|
|
dim=0,
|
|
)
|
|
else:
|
|
noise_pred = torch.stack(
|
|
[out.float() for out in model_out_list], dim=0
|
|
)
|
|
|
|
noise_pred = -noise_pred # Sign convention
|
|
|
|
# Compute log prob using SDE step
|
|
_, log_prob, prev_sample_mean, std_dev_t, sqrt_dt = sde_step_with_logprob(
|
|
noise_scheduler,
|
|
noise_pred.float(),
|
|
timesteps,
|
|
latents.float(),
|
|
prev_sample=next_latents.float(),
|
|
noise_level=noise_level,
|
|
return_sqrt_dt=True,
|
|
diffusion_clip=diffusion_clip,
|
|
diffusion_clip_value=diffusion_clip_value,
|
|
)
|
|
|
|
# Compute reference model prediction if provided
|
|
ref_prev_sample_mean = None
|
|
if ref_model is not None:
|
|
with torch.no_grad():
|
|
if apply_cfg:
|
|
ref_model_out_list = ref_model(
|
|
latent_model_input_list,
|
|
timestep_input,
|
|
prompt_embeds_input,
|
|
)[0]
|
|
ref_pos_out = ref_model_out_list[:batch_size]
|
|
ref_neg_out = ref_model_out_list[batch_size:]
|
|
ref_noise_pred = torch.stack(
|
|
[ref_neg_out[j].float() + cfg_scale * (ref_pos_out[j].float() - ref_neg_out[j].float())
|
|
for j in range(batch_size)],
|
|
dim=0,
|
|
)
|
|
else:
|
|
ref_model_out_list = ref_model(
|
|
latent_model_input_list,
|
|
timestep_input,
|
|
prompt_embeds_input,
|
|
)[0]
|
|
ref_noise_pred = torch.stack(
|
|
[out.float() for out in ref_model_out_list], dim=0
|
|
)
|
|
|
|
ref_noise_pred = -ref_noise_pred
|
|
|
|
_, _, ref_prev_sample_mean, _ = sde_step_with_logprob(
|
|
noise_scheduler,
|
|
ref_noise_pred.float(),
|
|
timesteps,
|
|
latents.float(),
|
|
prev_sample=next_latents.float(),
|
|
noise_level=noise_level,
|
|
diffusion_clip=diffusion_clip,
|
|
diffusion_clip_value=diffusion_clip_value,
|
|
)
|
|
|
|
return log_prob, prev_sample_mean, std_dev_t, sqrt_dt, ref_prev_sample_mean
|
|
|
|
# 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, network, args, 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):
|
|
logger.info("Running validation... ")
|
|
scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
|
|
args.pretrained_model_name_or_path,
|
|
subfolder="scheduler"
|
|
)
|
|
pipeline = ZImagePipeline(
|
|
vae=vae,
|
|
text_encoder=text_encoder,
|
|
tokenizer=tokenizer,
|
|
transformer=accelerator.unwrap_model(transformer3d) if type(transformer3d).__name__ == 'DistributedDataParallel' else transformer3d,
|
|
scheduler=scheduler,
|
|
)
|
|
pipeline = pipeline.to(accelerator.device)
|
|
|
|
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}")
|
|
|
|
n_prompts = len(args.validation_prompts)
|
|
if accelerator.num_processes > n_prompts:
|
|
# Fewer prompts than ranks: keep one seed-variant per rank of a single prompt
|
|
# (the previous behaviour), so the multi-seed view is not lost.
|
|
prompts_local = [(accelerator.process_index % n_prompts,
|
|
args.validation_prompts[accelerator.process_index % n_prompts])]
|
|
else:
|
|
# Stride the list by rank: every rank used to loop over the *whole* list with a
|
|
# rank-distinct seed, which produced num_processes near-duplicates of the same
|
|
# prompt and re-rendered longer lists num_processes times over.
|
|
prompts_local = list(enumerate(args.validation_prompts))[accelerator.process_index::accelerator.num_processes]
|
|
for i, prompt in prompts_local:
|
|
sample = pipeline(
|
|
prompt,
|
|
negative_prompt = "bad detailed",
|
|
height = args.image_sample_size,
|
|
width = args.image_sample_size,
|
|
generator = generator,
|
|
guidance_scale = 0 if "Turbo" in args.pretrained_model_name_or_path else 4.5,
|
|
num_inference_steps = 8 if "Turbo" in args.pretrained_model_name_or_path else 25,
|
|
).images
|
|
os.makedirs(os.path.join(args.output_dir, "sample"), exist_ok=True)
|
|
image = sample[0].save(
|
|
os.path.join(
|
|
args.output_dir,
|
|
f"sample/sample-{global_step}-rank{accelerator.process_index}-image-{i}.jpg"
|
|
)
|
|
)
|
|
|
|
del pipeline
|
|
gc.collect()
|
|
torch.cuda.empty_cache()
|
|
torch.cuda.ipc_collect()
|
|
vae.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype)
|
|
transformer3d.to(accelerator.device, dtype=weight_dtype)
|
|
if not args.enable_text_encoder_in_dataloader:
|
|
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:
|
|
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)
|
|
transformer3d.to(accelerator.device, dtype=weight_dtype)
|
|
if not args.enable_text_encoder_in_dataloader:
|
|
text_encoder.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype)
|
|
|
|
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(
|
|
"--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(
|
|
"--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("--use_ema", action="store_true", help="Whether to use EMA model.")
|
|
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=1.0, type=float, help="Max gradient norm.")
|
|
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(
|
|
"--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(
|
|
"--rank",
|
|
type=int,
|
|
default=128,
|
|
help=("The dimension of the LoRA update matrices."),
|
|
)
|
|
parser.add_argument(
|
|
"--network_alpha",
|
|
type=int,
|
|
default=64,
|
|
help=("The dimension of the LoRA update matrices."),
|
|
)
|
|
parser.add_argument(
|
|
"--use_peft_lora", action="store_true", help="Whether or not to use peft lora."
|
|
)
|
|
parser.add_argument(
|
|
"--train_text_encoder",
|
|
action="store_true",
|
|
help="Whether to train the text encoder. If set, the text encoder should be float32 precision.",
|
|
)
|
|
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_hw_adapt", action="store_true", help="Whether enable random adapt height and width in datasets."
|
|
)
|
|
parser.add_argument(
|
|
"--train_sampling_steps",
|
|
type=int,
|
|
default=1000,
|
|
help="Run train_sampling_steps.",
|
|
)
|
|
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(
|
|
"--transformer_path",
|
|
type=str,
|
|
default=None,
|
|
help=("If you want to load the weight from other transformers, input its path."),
|
|
)
|
|
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("--save_state", action="store_true", help="Whether or not to save state.")
|
|
|
|
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(
|
|
"--weighting_scheme",
|
|
type=str,
|
|
default="none",
|
|
choices=["sigma_sqrt", "logit_normal", "mode", "cosmap", "none"],
|
|
help=('We default to the "none" weighting scheme for uniform sampling and uniform loss'),
|
|
)
|
|
parser.add_argument(
|
|
"--logit_mean", type=float, default=0.0, help="mean to use when using the `'logit_normal'` weighting scheme."
|
|
)
|
|
parser.add_argument(
|
|
"--logit_std", type=float, default=1.0, help="std to use when using the `'logit_normal'` weighting scheme."
|
|
)
|
|
parser.add_argument(
|
|
"--mode_scale",
|
|
type=float,
|
|
default=1.29,
|
|
help="Scale of mode weighting scheme. Only effective when using the `'mode'` as the `weighting_scheme`.",
|
|
)
|
|
parser.add_argument(
|
|
"--lora_skip_name",
|
|
type=str,
|
|
default=None,
|
|
help=("The module is not trained in loras. "),
|
|
)
|
|
parser.add_argument(
|
|
"--target_name",
|
|
type=str,
|
|
default=None,
|
|
help=("The module is trained in loras. "),
|
|
)
|
|
parser.add_argument(
|
|
"--training_with_video_token_length",
|
|
action="store_true",
|
|
help="Whether to train with video token length. When set, the text encoder is not trained in lora mode.",
|
|
)
|
|
# GRPO specific arguments
|
|
parser.add_argument(
|
|
"--grpo_num_steps",
|
|
type=int,
|
|
default=20,
|
|
help="Number of inference steps for GRPO sampling.",
|
|
)
|
|
parser.add_argument(
|
|
"--grpo_cfg_scale",
|
|
type=float,
|
|
default=4.5,
|
|
help="Classifier-free guidance scale for GRPO sampling.",
|
|
)
|
|
parser.add_argument(
|
|
"--noise_level",
|
|
type=float,
|
|
default=1.2,
|
|
help="Noise level for SDE sampling in GRPO.",
|
|
)
|
|
parser.add_argument(
|
|
"--diffusion_clip",
|
|
action=argparse.BooleanOptionalAction,
|
|
default=False,
|
|
help="Truncated noise schedule (GenRL/flow_grpo): cap the per-step transition noise "
|
|
"std (std_dev_t*sqrt(-dt)) at --diffusion_clip_value. Tames overshoot at high "
|
|
"noise_level / high-noise steps. 0-effect when disabled.",
|
|
)
|
|
parser.add_argument(
|
|
"--diffusion_clip_value",
|
|
type=float,
|
|
default=0.45,
|
|
help="Max per-step transition noise std when --diffusion_clip is enabled.",
|
|
)
|
|
parser.add_argument(
|
|
"--same_latent",
|
|
action=argparse.BooleanOptionalAction,
|
|
default=False,
|
|
help="GRPO variance reduction: make every rollout of the same prompt start from an "
|
|
"identical initial latent (seeded by epoch+prompt) instead of an independent draw, "
|
|
"mirroring GenRL/flow_grpo same_latent.",
|
|
)
|
|
parser.add_argument(
|
|
"--sde_window_size",
|
|
type=int,
|
|
default=2,
|
|
help="SDE window size for GRPO training. 0 means use all steps.",
|
|
)
|
|
parser.add_argument(
|
|
"--sde_window_range",
|
|
nargs=2,
|
|
type=int,
|
|
default=[0, 5],
|
|
help="SDE window range [start, end] for GRPO training.",
|
|
)
|
|
parser.add_argument(
|
|
"--clip_range",
|
|
type=float,
|
|
default=1e-5,
|
|
help="PPO clip range for GRPO training.",
|
|
)
|
|
parser.add_argument(
|
|
"--adv_clip_max",
|
|
type=float,
|
|
default=5.0,
|
|
help="Maximum value for advantage clipping.",
|
|
)
|
|
parser.add_argument(
|
|
"--grpo_beta",
|
|
type=float,
|
|
default=0.0,
|
|
help="KL divergence coefficient for GRPO. 0 means no KL regularization.",
|
|
)
|
|
parser.add_argument(
|
|
"--per_prompt_stat_tracking",
|
|
action=argparse.BooleanOptionalAction,
|
|
default=True,
|
|
help="Whether to use per-prompt statistics tracking for advantage normalization. Use --no-per_prompt_stat_tracking to disable it.",
|
|
)
|
|
parser.add_argument(
|
|
"--global_std",
|
|
action=argparse.BooleanOptionalAction,
|
|
default=True,
|
|
help="Divide advantages by the std over all rewards of the round instead of the per-prompt group std. Use --no-global_std for strict group-relative normalization.",
|
|
)
|
|
parser.add_argument(
|
|
"--num_image_per_prompt",
|
|
type=int,
|
|
default=16,
|
|
help="Number of images to generate per prompt for GRPO group comparison. "
|
|
"Each prompt is repeated this many times consecutively by the sampler, and the advantage is "
|
|
"normalized within the group, so it is recommended that `num_batches_per_epoch * train_batch_size` "
|
|
"is divisible by this value.",
|
|
)
|
|
parser.add_argument(
|
|
"--num_batches_per_epoch",
|
|
type=int,
|
|
default=16,
|
|
help="Number of sampling batches to collect per epoch before training. "
|
|
"All batches are sampled with the same model, then concatenated for advantage computation and training.",
|
|
)
|
|
parser.add_argument(
|
|
"--reward_fn",
|
|
type=str,
|
|
default="MPSReward",
|
|
help="Reward function to use for GRPO training. For multiple rewards, use comma-separated names like 'HPSReward,MPSReward'.",
|
|
)
|
|
parser.add_argument(
|
|
"--reward_fn_kwargs",
|
|
type=str,
|
|
default=None,
|
|
help="JSON string of kwargs for the reward function. For multiple rewards, use JSON dict like '{\"HPSReward\": {\"version\": \"v2.1\"}, \"MPSReward\": {}}'.",
|
|
)
|
|
parser.add_argument(
|
|
"--multi_reward_weights",
|
|
type=str,
|
|
default=None,
|
|
help="JSON string of weights for combining advantages from multiple rewards, e.g., '{\"HPSReward\": 0.5, \"MPSReward\": 0.5}'. If None, use equal weights.",
|
|
)
|
|
|
|
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()
|
|
|
|
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)
|
|
|
|
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,
|
|
)
|
|
|
|
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)
|
|
# Dedicated rank-aware stream for the rollout stochasticity (SDE window start +
|
|
# per-step variance noise). `set_seed` is not device_specific, so the process-global
|
|
# torch/Python RNGs are seeded identically on every rank; drawing the injected SDE
|
|
# noise from them made all ranks inject bit-identical noise, which (together with the
|
|
# shared window start) collapsed cross-rank sample diversity. The offset keeps this
|
|
# stream disjoint from the initial-latent stream drawn from `torch_rng`.
|
|
sde_rng = torch.Generator(accelerator.device).manual_seed(args.seed + accelerator.process_index + 1_000_003)
|
|
else:
|
|
rng = None
|
|
torch_rng = None
|
|
sde_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
|
|
|
|
# When num_image_per_prompt == 1, per-prompt stat tracking is meaningless (no group to compare)
|
|
if args.num_image_per_prompt == 1:
|
|
args.per_prompt_stat_tracking = False
|
|
logger.info(f"num_image_per_prompt=1, auto-disabling per_prompt_stat_tracking")
|
|
else:
|
|
logger.info(f"num_image_per_prompt={args.num_image_per_prompt})")
|
|
|
|
# Initialize per-prompt stat tracker for advantage normalization
|
|
if args.per_prompt_stat_tracking:
|
|
stat_tracker = PerPromptStatTracker(global_std=args.global_std)
|
|
else:
|
|
stat_tracker = None
|
|
|
|
# Calculate number of train timesteps based on SDE window
|
|
if args.sde_window_size > 0:
|
|
num_train_timesteps = args.sde_window_size
|
|
else:
|
|
num_train_timesteps = args.grpo_num_steps - 1
|
|
|
|
# Load scheduler, tokenizer and models.
|
|
noise_scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
|
|
args.pretrained_model_name_or_path,
|
|
subfolder="scheduler"
|
|
)
|
|
|
|
# Get Tokenizer
|
|
tokenizer = AutoTokenizer.from_pretrained(
|
|
args.pretrained_model_name_or_path, subfolder="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 Qwen3ForCausalLM 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 = Qwen3ForCausalLM.from_pretrained(
|
|
args.pretrained_model_name_or_path, subfolder="text_encoder", torch_dtype=weight_dtype
|
|
)
|
|
text_encoder = text_encoder.eval()
|
|
# Get Vae
|
|
vae = AutoencoderKL.from_pretrained(
|
|
args.pretrained_model_name_or_path,
|
|
subfolder="vae"
|
|
).to(weight_dtype)
|
|
vae.eval()
|
|
|
|
import json
|
|
import videox_fun.reward.reward_fn as reward_fn
|
|
|
|
reward_fn_names = [name.strip() for name in args.reward_fn.split(',')]
|
|
is_multi_reward = len(reward_fn_names) > 1
|
|
|
|
# Parse kwargs for each reward function
|
|
reward_fn_kwargs_all = {}
|
|
if args.reward_fn_kwargs is not None:
|
|
parsed_kwargs = json.loads(args.reward_fn_kwargs)
|
|
if isinstance(parsed_kwargs, dict):
|
|
# Check if it's a nested dict (multi-reward format) or flat dict (single reward)
|
|
if is_multi_reward or any(name in parsed_kwargs for name in reward_fn_names):
|
|
reward_fn_kwargs_all = parsed_kwargs
|
|
else:
|
|
# Single reward with flat kwargs
|
|
reward_fn_kwargs_all = {reward_fn_names[0]: parsed_kwargs}
|
|
else:
|
|
reward_fn_kwargs_all = {reward_fn_names[0]: parsed_kwargs}
|
|
|
|
# Parse weights for multi-reward advantage combination
|
|
multi_reward_weights = {}
|
|
if args.multi_reward_weights is not None:
|
|
multi_reward_weights = json.loads(args.multi_reward_weights)
|
|
# Normalize weights to sum to 1
|
|
if is_multi_reward:
|
|
if not multi_reward_weights:
|
|
# Equal weights by default
|
|
multi_reward_weights = {name: 1.0 / len(reward_fn_names) for name in reward_fn_names}
|
|
else:
|
|
total = sum(multi_reward_weights.values())
|
|
multi_reward_weights = {name: multi_reward_weights.get(name, 0.0) / total for name in reward_fn_names}
|
|
|
|
# Initialize all reward functions
|
|
loss_fns = {}
|
|
if accelerator.is_main_process:
|
|
# Check if the models are downloaded in the main process
|
|
for fn_name in reward_fn_names:
|
|
fn_kwargs = reward_fn_kwargs_all.get(fn_name, {})
|
|
logger.info(f"Loading reward function: {fn_name} with kwargs: {fn_kwargs}")
|
|
loss_fns[fn_name] = getattr(reward_fn, fn_name)(device="cpu", dtype=weight_dtype, **fn_kwargs)
|
|
accelerator.wait_for_everyone()
|
|
|
|
# Re-initialize on correct device
|
|
loss_fns = {}
|
|
for fn_name in reward_fn_names:
|
|
fn_kwargs = reward_fn_kwargs_all.get(fn_name, {})
|
|
loss_fns[fn_name] = getattr(reward_fn, fn_name)(device=accelerator.device, dtype=weight_dtype, **fn_kwargs)
|
|
|
|
if is_multi_reward:
|
|
logger.info(f"Multi-reward enabled with {len(reward_fn_names)} rewards: {reward_fn_names}")
|
|
logger.info(f"Advantage combination weights: {multi_reward_weights}")
|
|
else:
|
|
logger.info(f"Single reward function: {reward_fn_names[0]}")
|
|
|
|
# For backward compatibility, keep loss_fn as the first reward function
|
|
loss_fn = loss_fns[reward_fn_names[0]]
|
|
|
|
# Get Transformer
|
|
transformer3d = ZImageTransformer2DModel.from_pretrained(
|
|
args.pretrained_model_name_or_path,
|
|
subfolder="transformer",
|
|
torch_dtype=weight_dtype,
|
|
low_cpu_mem_usage=True,
|
|
).to(weight_dtype)
|
|
ref_transformer3d = ZImageTransformer2DModel.from_pretrained(
|
|
args.pretrained_model_name_or_path,
|
|
subfolder="transformer",
|
|
torch_dtype=weight_dtype,
|
|
low_cpu_mem_usage=True,
|
|
).to(weight_dtype)
|
|
|
|
# Freeze vae and text_encoder and set transformer3d to trainable
|
|
vae.requires_grad_(False)
|
|
text_encoder.requires_grad_(False)
|
|
transformer3d.requires_grad_(False)
|
|
ref_transformer3d.requires_grad_(False)
|
|
|
|
# Lora will work with this...
|
|
if args.use_peft_lora:
|
|
from peft import (LoraConfig, get_peft_model_state_dict,
|
|
inject_adapter_in_model)
|
|
lora_config = LoraConfig(r=args.rank, lora_alpha=args.network_alpha, target_modules=args.target_name.split(","))
|
|
transformer3d = inject_adapter_in_model(lora_config, transformer3d)
|
|
|
|
network = None
|
|
else:
|
|
network = create_network(
|
|
1.0,
|
|
args.rank,
|
|
args.network_alpha,
|
|
text_encoder,
|
|
transformer3d,
|
|
neuron_dropout=None,
|
|
target_name=args.target_name,
|
|
skip_name=args.lora_skip_name,
|
|
)
|
|
network = network.to(weight_dtype)
|
|
network.apply_to(text_encoder, transformer3d, args.train_text_encoder and not args.training_with_video_token_length, True)
|
|
|
|
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 = transformer3d.load_state_dict(state_dict, strict=False)
|
|
print(f"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
|
|
|
|
# `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"lora_diffusion_pytorch_model.safetensors")
|
|
if args.use_peft_lora:
|
|
network_state_dict = get_peft_model_state_dict(accelerator.unwrap_model(models[-1]), accelerate_state_dict)
|
|
network_state_dict_kohya = convert_peft_lora_to_kohya_lora(network_state_dict)
|
|
safetensor_kohya_format_save_path = os.path.join(output_dir, f"lora_diffusion_pytorch_model_compatible_with_comfyui.safetensors")
|
|
save_model(safetensor_kohya_format_save_path, network_state_dict_kohya)
|
|
else:
|
|
network_state_dict = {}
|
|
for key in accelerate_state_dict:
|
|
if "network" in key:
|
|
network_state_dict[key.replace("network.", "")] = accelerate_state_dict[key].to(weight_dtype)
|
|
save_file(network_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):
|
|
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"lora_diffusion_pytorch_model.safetensors")
|
|
if args.use_peft_lora:
|
|
network_state_dict = get_peft_model_state_dict(accelerator.unwrap_model(models[-1]), accelerate_state_dict)
|
|
network_state_dict_kohya = convert_peft_lora_to_kohya_lora(network_state_dict)
|
|
safetensor_kohya_format_save_path = os.path.join(output_dir, f"lora_diffusion_pytorch_model_compatible_with_comfyui.safetensors")
|
|
save_model(safetensor_kohya_format_save_path, network_state_dict_kohya)
|
|
else:
|
|
network_state_dict = {}
|
|
for key in accelerate_state_dict:
|
|
if "network" in key:
|
|
network_state_dict[key.replace("network.", "")] = accelerate_state_dict[key].to(weight_dtype)
|
|
save_file(network_state_dict, safetensor_save_path, metadata={"format": "pt"})
|
|
|
|
if not args.use_deepspeed:
|
|
for _ in range(len(weights)):
|
|
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):
|
|
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)
|
|
|
|
if args.gradient_checkpointing:
|
|
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
|
|
|
|
if args.use_peft_lora:
|
|
logging.info("Add peft parameters")
|
|
trainable_params = list(filter(lambda p: p.requires_grad, transformer3d.parameters()))
|
|
trainable_params_optim = list(filter(lambda p: p.requires_grad, transformer3d.parameters()))
|
|
else:
|
|
logging.info("Add network parameters")
|
|
trainable_params = list(filter(lambda p: p.requires_grad, network.parameters()))
|
|
trainable_params_optim = network.prepare_optimizer_params(args.learning_rate / 2, args.learning_rate, args.learning_rate)
|
|
|
|
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)
|
|
)
|
|
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,
|
|
)
|
|
|
|
# Get the training dataset
|
|
if args.fix_sample_size is not None and args.enable_bucket:
|
|
args.image_sample_size = max(max(args.fix_sample_size), args.image_sample_size)
|
|
args.random_hw_adapt = False
|
|
|
|
# Get the dataset
|
|
train_dataset = TextDataset(
|
|
args.train_data_meta,
|
|
text_drop_ratio=0.0,
|
|
)
|
|
|
|
def worker_init_fn(_seed):
|
|
_seed = _seed * 256
|
|
def _worker_init_fn(worker_id):
|
|
print(f"worker_init_fn with {_seed + worker_id}")
|
|
np.random.seed(_seed + worker_id)
|
|
random.seed(_seed + worker_id)
|
|
return _worker_init_fn
|
|
|
|
if args.enable_bucket:
|
|
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_embeds = encode_prompt(
|
|
new_examples['text'], device="cpu",
|
|
text_encoder=text_encoder,
|
|
tokenizer=tokenizer,
|
|
)
|
|
new_examples['prompt_embeds'] = prompt_embeds
|
|
|
|
neg_prompt_embeds = encode_prompt(
|
|
["亮度过高，过曝，严重的色彩失真，低分辨率，低画质，肢体畸形，手指畸形，画面过饱和，蜡像感，人脸无细节，过度光滑，画面具有AI感。构图混乱。文字模糊，扭曲。"], device="cpu",
|
|
text_encoder=text_encoder,
|
|
tokenizer=tokenizer,
|
|
)
|
|
new_examples['neg_prompt_embeds'] = 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, k_repeat=args.num_image_per_prompt),
|
|
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,
|
|
worker_init_fn=worker_init_fn(args.seed + accelerator.process_index)
|
|
)
|
|
else:
|
|
# DataLoaders creation:
|
|
batch_sampler_generator = torch.Generator().manual_seed(args.seed)
|
|
batch_sampler = BatchSampler(
|
|
RandomSampler(train_dataset, generator=batch_sampler_generator, k_repeat=args.num_image_per_prompt),
|
|
batch_size=args.train_batch_size, drop_last=True,
|
|
)
|
|
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,
|
|
worker_init_fn=worker_init_fn(args.seed + accelerator.process_index)
|
|
)
|
|
|
|
# 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,
|
|
)
|
|
# Prepare everything with our `accelerator`.
|
|
if args.use_peft_lora:
|
|
transformer3d, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
|
|
transformer3d, optimizer, train_dataloader, lr_scheduler
|
|
)
|
|
else:
|
|
transformer3d.network = network
|
|
transformer3d = transformer3d.to(dtype=weight_dtype)
|
|
transformer3d, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
|
|
transformer3d, optimizer, train_dataloader, 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
|
|
shard_fn = partial(shard_model, device_id=accelerator.device, param_dtype=weight_dtype, module_to_wrapper=text_encoder.model.layers)
|
|
text_encoder = shard_fn(text_encoder)
|
|
|
|
# Move text_encode and vae to gpu and cast to weight_dtype
|
|
vae.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype)
|
|
transformer3d.to(accelerator.device, dtype=weight_dtype)
|
|
ref_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", dtype=weight_dtype)
|
|
|
|
# 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}")
|
|
accelerator.load_state(os.path.join(args.output_dir, path))
|
|
else:
|
|
initial_global_step = 0
|
|
|
|
# function for saving/removing
|
|
def save_model(ckpt_file, unwrapped_nw):
|
|
os.makedirs(args.output_dir, exist_ok=True)
|
|
accelerator.print(f"\nsaving checkpoint: {ckpt_file}")
|
|
if isinstance(unwrapped_nw, dict):
|
|
from safetensors.torch import save_file
|
|
save_file(unwrapped_nw, ckpt_file, metadata={"format": "pt"})
|
|
return ckpt_file
|
|
unwrapped_nw.save_weights(ckpt_file, weight_dtype, None)
|
|
|
|
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:
|
|
# 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
|
|
|
|
for epoch in range(first_epoch, args.num_train_epochs):
|
|
train_loss = 0.0
|
|
batch_sampler.sampler.generator = torch.Generator().manual_seed(args.seed + epoch)
|
|
transformer3d.eval()
|
|
all_samples = []
|
|
|
|
for step, batch in enumerate(train_dataloader):
|
|
#################### SAMPLING (eval mode, fixed model params) ####################
|
|
# Data batch sanity check
|
|
if epoch == first_epoch and step == 0:
|
|
texts = batch['text']
|
|
for idx, text_item in enumerate(texts):
|
|
print(f"[Sanity Check] Sample {idx}: {text_item[:100]}...")
|
|
|
|
with torch.no_grad():
|
|
text = batch['text']
|
|
if args.fix_sample_size is not None:
|
|
local_sample_size = [int(x / 16) * 16 for x in args.fix_sample_size]
|
|
else:
|
|
if args.random_hw_adapt:
|
|
aspect_ratio_sample_size = {key : [x / 512 * args.image_sample_size for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()}
|
|
if rng is None:
|
|
aspect_ratio_key = np.random.choice(list(aspect_ratio_sample_size.keys()))
|
|
else:
|
|
aspect_ratio_key = rng.choice(list(aspect_ratio_sample_size.keys()))
|
|
local_sample_size = aspect_ratio_sample_size[aspect_ratio_key]
|
|
local_sample_size = [int(x / 16) * 16 for x in local_sample_size]
|
|
else:
|
|
local_sample_size = [args.image_sample_size, args.image_sample_size]
|
|
|
|
vae_scale_factor = (
|
|
2 ** (len(vae.config.block_out_channels) - 1)
|
|
)
|
|
target_shape = (
|
|
len(text),
|
|
vae.latent_channels,
|
|
1,
|
|
int(local_sample_size[0] // vae_scale_factor),
|
|
int(local_sample_size[1] // vae_scale_factor),
|
|
)
|
|
|
|
if args.low_vram:
|
|
vae.to('cpu')
|
|
torch.cuda.empty_cache()
|
|
if not args.enable_text_encoder_in_dataloader:
|
|
text_encoder.to(accelerator.device)
|
|
|
|
if args.enable_text_encoder_in_dataloader:
|
|
prompt_embeds = batch['prompt_embeds'].to(dtype=weight_dtype, device=accelerator.device)
|
|
neg_prompt_embeds = batch['neg_prompt_embeds'].to(dtype=weight_dtype, device=accelerator.device)
|
|
else:
|
|
with torch.no_grad():
|
|
prompt_embeds = encode_prompt(
|
|
text,
|
|
device=accelerator.device,
|
|
text_encoder=text_encoder,
|
|
tokenizer=tokenizer,
|
|
)
|
|
# Generate negative embeddings and repeat for batch size
|
|
neg_prompt_single = encode_prompt(
|
|
["亮度过高，过曝，严重的色彩失真，低分辨率，低画质，肢体畸形，手指畸形，画面过饱和，蜡像感，人脸无细节，过度光滑，画面具有AI感。构图混乱。文字模糊，扭曲。"],
|
|
device=accelerator.device,
|
|
text_encoder=text_encoder,
|
|
tokenizer=tokenizer,
|
|
)
|
|
# Repeat negative embedding for each sample in batch
|
|
neg_prompt_embeds = neg_prompt_single * len(text)
|
|
|
|
if args.low_vram and not args.enable_text_encoder_in_dataloader:
|
|
text_encoder.to('cpu')
|
|
torch.cuda.empty_cache()
|
|
|
|
# same_latent: seed the initial latent from (epoch, prompt) so every rollout of
|
|
# the same prompt -- on every rank and step -- draws an identical x_T. Within-group
|
|
# advantage then reflects only trajectory differences (policy + per-step SDE noise
|
|
# via sde_rng), not different starting latents. Mirrors GenRL/flow_grpo same_latent.
|
|
if args.same_latent:
|
|
_seed_key = f"{epoch}_{list(text)}".encode("utf-8")
|
|
_prompt_seed = int.from_bytes(hashlib.md5(_seed_key).digest()[:8], "little") & 0x7FFFFFFF
|
|
_latent_gen = torch.Generator(device=accelerator.device).manual_seed(_prompt_seed)
|
|
shared_noise = torch.randn(
|
|
target_shape,
|
|
device=accelerator.device,
|
|
generator=_latent_gen,
|
|
dtype=weight_dtype,
|
|
)
|
|
else:
|
|
shared_noise = torch.randn(
|
|
target_shape,
|
|
device=accelerator.device,
|
|
generator=torch_rng,
|
|
dtype=weight_dtype
|
|
)
|
|
|
|
with torch.no_grad():
|
|
collected_data = sample_with_cfg(
|
|
model=transformer3d,
|
|
vae=vae,
|
|
noise=shared_noise.clone(),
|
|
prompt_embeds=prompt_embeds,
|
|
neg_prompt_embeds=neg_prompt_embeds,
|
|
num_steps=args.grpo_num_steps,
|
|
cfg_scale=args.grpo_cfg_scale,
|
|
noise_scheduler=noise_scheduler,
|
|
device=accelerator.device,
|
|
dtype=weight_dtype,
|
|
noise_level=args.noise_level,
|
|
sde_window_size=args.sde_window_size,
|
|
sde_window_range=tuple(args.sde_window_range),
|
|
noise_generator=sde_rng,
|
|
diffusion_clip=args.diffusion_clip,
|
|
diffusion_clip_value=args.diffusion_clip_value,
|
|
)
|
|
|
|
latents = torch.stack(collected_data["all_latents"], dim=1)
|
|
log_probs = torch.stack(collected_data["all_log_probs"], dim=1)
|
|
timesteps = torch.stack(collected_data["all_timesteps"]).unsqueeze(0).repeat(len(text), 1).to(accelerator.device)
|
|
images = collected_data["images"]
|
|
|
|
# Compute rewards (supports multiple reward functions)
|
|
if is_multi_reward:
|
|
# Compute rewards from all reward functions
|
|
rewards_dict_local = {}
|
|
for fn_name, fn in loss_fns.items():
|
|
rewards_dict_local[fn_name] = fn.get_reward(images, text)
|
|
|
|
individual_rewards = rewards_dict_local
|
|
# Use first reward as placeholder for shape compatibility
|
|
rewards = rewards_dict_local[reward_fn_names[0]]
|
|
else:
|
|
# Single reward function
|
|
rewards = loss_fn.get_reward(images, text)
|
|
individual_rewards = {reward_fn_names[0]: rewards}
|
|
|
|
# Save sample images to output_dir for logging
|
|
if step == 0:
|
|
num_log_samples = min(4, images.shape[0])
|
|
sample_indices = random.sample(range(images.shape[0]), num_log_samples)
|
|
os.makedirs(os.path.join(args.output_dir, "sample"), exist_ok=True)
|
|
for log_idx, img_idx in enumerate(sample_indices):
|
|
log_img = images[img_idx, :, 0].permute(1, 2, 0).cpu().float()
|
|
log_img = log_img.clamp(0, 1).numpy() * 255
|
|
from PIL import Image as PILImage
|
|
PILImage.fromarray(np.uint8(log_img)).save(
|
|
os.path.join(args.output_dir, "sample", f"epoch{epoch}_rank{accelerator.process_index}_{log_idx}.jpg")
|
|
)
|
|
|
|
# Append this batch's samples
|
|
all_samples.append({
|
|
"prompt_embeds": collected_data["prompt_embeds"],
|
|
"negative_prompt_embeds": collected_data["negative_prompt_embeds"],
|
|
"timesteps": timesteps,
|
|
"latents": latents[:, :-1],
|
|
"next_latents": latents[:, 1:],
|
|
"log_probs": log_probs,
|
|
"rewards": rewards,
|
|
"individual_rewards": individual_rewards,
|
|
"prompts": text,
|
|
})
|
|
|
|
# Only proceed to training when we've collected enough batches
|
|
if len(all_samples) < args.num_batches_per_epoch:
|
|
continue
|
|
|
|
#################### COLLATE SAMPLES & COMPUTE ADVANTAGES ####################
|
|
# Concatenate all sample batches into one large batch
|
|
all_prompts_local = [p for s in all_samples for p in s["prompts"]]
|
|
|
|
# Collect individual rewards
|
|
all_individual_rewards = {}
|
|
for fn_name in reward_fn_names:
|
|
if fn_name in all_samples[0].get("individual_rewards", {}):
|
|
all_individual_rewards[fn_name] = torch.cat(
|
|
[s["individual_rewards"][fn_name] for s in all_samples], dim=0
|
|
)
|
|
|
|
world_size = accelerator.num_processes
|
|
|
|
# Gather individual rewards from all GPUs
|
|
gathered_individual_rewards = {}
|
|
for fn_name, rewards_tensor in all_individual_rewards.items():
|
|
gathered_individual_rewards[fn_name] = accelerator.gather(rewards_tensor)
|
|
|
|
# Gather prompts
|
|
gathered_prompts_list = [None] * world_size
|
|
import torch.distributed as dist
|
|
if dist.is_initialized():
|
|
dist.all_gather_object(gathered_prompts_list, all_prompts_local)
|
|
gathered_prompts = [p for sublist in gathered_prompts_list for p in sublist]
|
|
else:
|
|
gathered_prompts = all_prompts_local
|
|
|
|
if accelerator.is_main_process:
|
|
for fn_name, ind_rewards in gathered_individual_rewards.items():
|
|
logger.info(f"Epoch {epoch} Batch index {step} - {fn_name}: mean={ind_rewards.mean():.4f}, std={ind_rewards.std():.4f}")
|
|
|
|
# Compute advantages for each reward independently, then combine
|
|
if is_multi_reward and stat_tracker is not None:
|
|
# For multi-reward: compute advantage per reward, then weighted sum
|
|
advantages_per_reward = {}
|
|
for fn_name, gathered_rewards in gathered_individual_rewards.items():
|
|
gathered_rewards_np = gathered_rewards.cpu().float().numpy()
|
|
# Use a separate stat_tracker instance per reward (or share with clear)
|
|
adv = stat_tracker.update(gathered_prompts, gathered_rewards_np)
|
|
advantages_per_reward[fn_name] = adv
|
|
stat_tracker.clear() # Clear for next reward
|
|
|
|
# Weighted combination of advantages
|
|
advantages_all = None
|
|
for fn_name, adv in advantages_per_reward.items():
|
|
weight = multi_reward_weights.get(fn_name, 1.0 / len(reward_fn_names))
|
|
|
|
weighted_adv = adv * weight
|
|
if advantages_all is None:
|
|
advantages_all = weighted_adv
|
|
else:
|
|
advantages_all = advantages_all + weighted_adv
|
|
|
|
if accelerator.is_main_process:
|
|
group_size, trained_prompt_num = stat_tracker.get_stats()
|
|
# Use first reward for zero_std_ratio calculation
|
|
first_reward_np = list(gathered_individual_rewards.values())[0].cpu().float().numpy()
|
|
zero_std_ratio, reward_std_mean = calculate_zero_std_ratio(
|
|
gathered_prompts,
|
|
{"ori_avg": first_reward_np}
|
|
)
|
|
logger.info(f" Per-prompt stats: group_size={group_size:.2f}, trained_prompts={trained_prompt_num}")
|
|
logger.info(f" Combined advantage: mean={advantages_all.mean():.4f}, std={advantages_all.std():.4f}")
|
|
|
|
# Build log dict
|
|
log_dict = {
|
|
"group_size": group_size,
|
|
"trained_prompt_num": trained_prompt_num,
|
|
"zero_std_ratio": zero_std_ratio,
|
|
"reward_std_mean": reward_std_mean,
|
|
"combined_advantage_mean": float(advantages_all.mean()),
|
|
"combined_advantage_std": float(advantages_all.std()),
|
|
}
|
|
|
|
# Add individual reward/advantage stats
|
|
for fn_name, ind_rewards in gathered_individual_rewards.items():
|
|
log_dict[f"{fn_name}_reward_mean"] = ind_rewards.mean().item()
|
|
log_dict[f"{fn_name}_reward_std"] = ind_rewards.std().item()
|
|
log_dict[f"{fn_name}_advantage_mean"] = float(advantages_per_reward[fn_name].mean())
|
|
log_dict[f"{fn_name}_advantage_std"] = float(advantages_per_reward[fn_name].std())
|
|
log_dict[f"{fn_name}_weight"] = multi_reward_weights.get(fn_name, 1.0 / len(reward_fn_names))
|
|
|
|
accelerator.log(log_dict, step=global_step)
|
|
|
|
advantages_all = torch.as_tensor(advantages_all, device=accelerator.device, dtype=weight_dtype)
|
|
if world_size > 1:
|
|
local_total = len(all_prompts_local)
|
|
advantages_local = advantages_all.reshape(world_size, local_total)[accelerator.process_index]
|
|
else:
|
|
advantages_local = advantages_all
|
|
|
|
elif stat_tracker is not None:
|
|
# Single reward path
|
|
all_rewards = torch.cat([s["rewards"] for s in all_samples], dim=0)
|
|
gathered_rewards = accelerator.gather(all_rewards)
|
|
gathered_rewards_np = gathered_rewards.cpu().float().numpy()
|
|
|
|
advantages_all = stat_tracker.update(gathered_prompts, gathered_rewards_np)
|
|
|
|
if accelerator.is_main_process:
|
|
group_size, trained_prompt_num = stat_tracker.get_stats()
|
|
zero_std_ratio, reward_std_mean = calculate_zero_std_ratio(
|
|
gathered_prompts,
|
|
{"ori_avg": gathered_rewards_np}
|
|
)
|
|
logger.info(f"Epoch {epoch} Step {step}: gathered rewards mean={gathered_rewards.mean():.4f}, std={gathered_rewards.std():.4f}")
|
|
logger.info(f" Per-prompt stats: group_size={group_size:.2f}, trained_prompts={trained_prompt_num}")
|
|
|
|
log_dict = {
|
|
"reward_mean": gathered_rewards.mean().item(),
|
|
"reward_std": gathered_rewards.std().item(),
|
|
"group_size": group_size,
|
|
"trained_prompt_num": trained_prompt_num,
|
|
"zero_std_ratio": zero_std_ratio,
|
|
"reward_std_mean": reward_std_mean,
|
|
}
|
|
accelerator.log(log_dict, step=global_step)
|
|
|
|
stat_tracker.clear()
|
|
|
|
advantages_all = torch.as_tensor(advantages_all, device=accelerator.device, dtype=weight_dtype)
|
|
if world_size > 1:
|
|
local_total = len(all_prompts_local)
|
|
advantages_local = advantages_all.reshape(world_size, local_total)[accelerator.process_index]
|
|
else:
|
|
advantages_local = advantages_all
|
|
else:
|
|
# No stat_tracker - use global normalization
|
|
all_rewards = torch.cat([s["rewards"] for s in all_samples], dim=0)
|
|
gathered_rewards = accelerator.gather(all_rewards)
|
|
advantages_all_global = (gathered_rewards - gathered_rewards.mean()) / (gathered_rewards.std() + 1e-4)
|
|
if world_size > 1:
|
|
local_total = len(all_prompts_local)
|
|
advantages_local = advantages_all_global.reshape(world_size, local_total)[accelerator.process_index]
|
|
else:
|
|
advantages_local = advantages_all_global
|
|
|
|
# Expand advantages to timestep dimension and assign back to each sample batch
|
|
advantages_local = advantages_local.unsqueeze(1).repeat(1, num_train_timesteps)
|
|
offset = 0
|
|
for s in all_samples:
|
|
bs = s["rewards"].shape[0]
|
|
s["advantages"] = advantages_local[offset:offset+bs]
|
|
offset += bs
|
|
del s["rewards"]
|
|
del s["prompts"]
|
|
if "individual_rewards" in s:
|
|
del s["individual_rewards"]
|
|
|
|
# Concatenate all samples into one big dict (tensor fields only)
|
|
tensor_keys = [k for k in all_samples[0].keys() if isinstance(all_samples[0][k], torch.Tensor)]
|
|
list_keys = [k for k in all_samples[0].keys() if not isinstance(all_samples[0][k], torch.Tensor)]
|
|
samples_concat = {}
|
|
for k in tensor_keys:
|
|
samples_concat[k] = torch.cat([s[k] for s in all_samples], dim=0)
|
|
for k in list_keys:
|
|
samples_concat[k] = [item for s in all_samples for item in s[k]]
|
|
|
|
# Also collect prompt_embeds and neg_prompt_embeds for training
|
|
all_prompt_embeds_flat = []
|
|
all_neg_prompt_embeds_flat = []
|
|
for s in all_samples:
|
|
pe = s["prompt_embeds"]
|
|
ne = s["negative_prompt_embeds"]
|
|
if isinstance(pe, list):
|
|
all_prompt_embeds_flat.extend(pe)
|
|
all_neg_prompt_embeds_flat.extend(ne)
|
|
else:
|
|
all_prompt_embeds_flat.append(pe)
|
|
all_neg_prompt_embeds_flat.append(ne)
|
|
|
|
# Drop samples whose within-group reward std is 0 (advantage exactly 0)
|
|
# so they do not dilute the effective gradient.
|
|
# Mirrors flow_grpo/scripts/train_sd3_fast.py:
|
|
# mask = (samples["advantages"].abs().sum(dim=1) != 0)
|
|
# plus a divisibility fixup so the kept count stays a multiple of
|
|
# num_batches_per_epoch (the rebatch below uses
|
|
# per_batch_size = total // num_batches_per_epoch, so the batch count
|
|
# is always num_batches_per_epoch; hence the backward count and the
|
|
# gradient_accumulation_steps_total normalization stay accurate).
|
|
adv_mask = (samples_concat["advantages"].abs().sum(dim=1) != 0)
|
|
num_batches_ep = args.num_batches_per_epoch
|
|
true_count = int(adv_mask.sum().item())
|
|
remainder = true_count % num_batches_ep
|
|
if true_count == 0 or remainder != 0:
|
|
false_indices = torch.where(~adv_mask)[0]
|
|
num_to_change = num_batches_ep - remainder
|
|
if len(false_indices) >= num_to_change:
|
|
perm = torch.randperm(len(false_indices), device=adv_mask.device)[:num_to_change]
|
|
adv_mask[false_indices[perm]] = True
|
|
else:
|
|
num_to_change -= len(false_indices)
|
|
true_indices = torch.where(adv_mask)[0]
|
|
perm = torch.randperm(len(true_indices), device=adv_mask.device)[:num_to_change]
|
|
adv_mask[true_indices[perm]] = False
|
|
num_dropped = int((~adv_mask).sum().item())
|
|
if num_dropped > 0:
|
|
adv_mask_list = adv_mask.tolist()
|
|
for k in tensor_keys:
|
|
samples_concat[k] = samples_concat[k][adv_mask]
|
|
# Note: list_keys (e.g. samples_concat["prompt_embeds"]) are used
|
|
# neither by the rebatch below (tensor_keys only) nor by the training
|
|
# forward (which uses all_prompt_embeds_flat), so we skip filtering
|
|
# them to avoid an IndexError when their length mismatches the mask.
|
|
all_prompt_embeds_flat = [e for i, e in enumerate(all_prompt_embeds_flat) if adv_mask_list[i]]
|
|
all_neg_prompt_embeds_flat = [e for i, e in enumerate(all_neg_prompt_embeds_flat) if adv_mask_list[i]]
|
|
if accelerator.is_main_process:
|
|
logger.info(
|
|
f"Epoch {epoch}: advantage==0 mask dropped {num_dropped}/{len(adv_mask_list)} samples, "
|
|
f"kept {int(adv_mask.sum().item())}"
|
|
)
|
|
|
|
total_batch_size_collected = samples_concat["timesteps"].shape[0]
|
|
|
|
#################### TRAINING ####################
|
|
# Move ref model to GPU for training phase in low_vram mode
|
|
if args.low_vram and args.grpo_beta > 0:
|
|
ref_transformer3d.to(accelerator.device)
|
|
|
|
gradient_accumulation_steps_total = (args.gradient_accumulation_steps * num_train_timesteps * args.num_batches_per_epoch) // 2
|
|
|
|
# Rebatch: split the big concatenated samples back into smaller batches for training
|
|
per_batch_size = total_batch_size_collected // args.num_batches_per_epoch
|
|
samples_batched = []
|
|
for bi in range(args.num_batches_per_epoch):
|
|
batch_slice = slice(bi * per_batch_size, (bi + 1) * per_batch_size)
|
|
one_batch = {k: samples_concat[k][batch_slice] for k in tensor_keys}
|
|
samples_batched.append(one_batch)
|
|
|
|
transformer3d.train()
|
|
info = {"approx_kl": [], "clipfrac": [], "policy_loss": [], "kl_loss": [], "loss": []}
|
|
for batch_idx, sample in enumerate(samples_batched):
|
|
# Get corresponding prompt_embeds for this sub-batch
|
|
sub_prompt_embeds = all_prompt_embeds_flat[batch_idx * per_batch_size:(batch_idx + 1) * per_batch_size]
|
|
sub_neg_prompt_embeds = all_neg_prompt_embeds_flat[batch_idx * per_batch_size:(batch_idx + 1) * per_batch_size]
|
|
|
|
for j in range(num_train_timesteps):
|
|
# Manual gradient accumulation: sync every gradient_accumulation_steps_total steps
|
|
flat_step = batch_idx * num_train_timesteps + j
|
|
should_sync = (flat_step + 1) % gradient_accumulation_steps_total == 0
|
|
|
|
# Disable gradient sync for accumulation steps, enable for sync steps
|
|
context = contextlib.nullcontext if should_sync else accelerator.no_sync
|
|
with context(transformer3d) if not should_sync else contextlib.nullcontext():
|
|
log_prob, prev_sample_mean, std_dev_t, sqrt_dt, ref_prev_sample_mean = compute_log_prob(
|
|
model=transformer3d,
|
|
vae=vae,
|
|
sample=sample,
|
|
step_idx=j,
|
|
noise_scheduler=noise_scheduler,
|
|
prompt_embeds=sub_prompt_embeds,
|
|
neg_prompt_embeds=sub_neg_prompt_embeds,
|
|
cfg_scale=args.grpo_cfg_scale,
|
|
noise_level=args.noise_level,
|
|
dtype=weight_dtype,
|
|
ref_model=ref_transformer3d if args.grpo_beta > 0 else None,
|
|
diffusion_clip=args.diffusion_clip,
|
|
diffusion_clip_value=args.diffusion_clip_value,
|
|
)
|
|
|
|
# GRPO loss computation
|
|
adv = torch.clamp(
|
|
sample["advantages"][:, j],
|
|
-args.adv_clip_max,
|
|
args.adv_clip_max,
|
|
)
|
|
|
|
ratio = torch.exp(log_prob - sample["log_probs"][:, j])
|
|
|
|
unclipped_loss = -adv * ratio
|
|
clipped_loss = -adv * torch.clamp(
|
|
ratio,
|
|
1.0 - args.clip_range,
|
|
1.0 + args.clip_range,
|
|
)
|
|
policy_loss = torch.mean(torch.maximum(unclipped_loss, clipped_loss))
|
|
policy_loss = policy_loss / gradient_accumulation_steps_total
|
|
|
|
if args.grpo_beta > 0 and ref_prev_sample_mean is not None:
|
|
# Per-step Gaussian KL between the policy and reference SDE transitions.
|
|
# Policy and ref share the same timestep/scheduler, so the transition
|
|
# std is identical and model-independent:
|
|
# sigma_step = std_dev_t * sqrt(-dt) = std_dev_t * sqrt_dt
|
|
# (sd3_sde_with_logprob.py L71 adds std_dev_t*sqrt(-dt)*noise, and its
|
|
# L74 log_prob denominator is exactly 2*(std_dev_t*sqrt(-dt))**2).
|
|
# Equal-variance Gaussians give KL = (mu - mu_ref)^2 / (2*sigma_step^2),
|
|
# so the denominator MUST be 2*(std_dev_t*sqrt_dt)**2 to stay consistent
|
|
# with the log_prob used in the ratio above. (flow_grpo's 2*std_dev_t**2
|
|
# drops the (-dt) factor and contradicts its own log_prob -- do not copy.)
|
|
# Latent is 5-D (B,C,1,H,W) -> average over dim=(1,2,3,4); std_dev_t and
|
|
# sqrt_dt are (B,1,1,1,1) and squeeze to match the (B,) numerator.
|
|
kl_loss = ((prev_sample_mean - ref_prev_sample_mean) ** 2).mean(dim=(1,2,3,4)) / (2 * (std_dev_t * sqrt_dt).squeeze() ** 2 + 1e-8)
|
|
kl_loss = torch.mean(kl_loss)
|
|
kl_loss = kl_loss / gradient_accumulation_steps_total
|
|
loss = policy_loss + args.grpo_beta * kl_loss
|
|
else:
|
|
kl_loss = torch.tensor(0.0, device=accelerator.device)
|
|
loss = policy_loss
|
|
|
|
approx_kl = 0.5 * torch.mean((log_prob - sample["log_probs"][:, j]) ** 2)
|
|
clipfrac = torch.mean((torch.abs(ratio - 1.0) > args.clip_range).float())
|
|
info["approx_kl"].append(approx_kl)
|
|
info["clipfrac"].append(clipfrac)
|
|
info["policy_loss"].append(policy_loss)
|
|
info["kl_loss"].append(kl_loss)
|
|
info["loss"].append(loss)
|
|
|
|
accelerator.backward(loss)
|
|
train_loss += loss.detach().item() * gradient_accumulation_steps_total
|
|
|
|
if should_sync:
|
|
accelerator.clip_grad_norm_(trainable_params, args.max_grad_norm)
|
|
optimizer.step()
|
|
lr_scheduler.step()
|
|
optimizer.zero_grad()
|
|
|
|
info_log = {k: torch.mean(torch.stack([v.float() if isinstance(v, torch.Tensor) else torch.tensor(v) for v in vals])) for k, vals in info.items() if vals}
|
|
logs = {
|
|
"step_loss": loss.detach().item(),
|
|
"lr": lr_scheduler.get_last_lr()[0],
|
|
"reward": gathered_rewards.mean().item(),
|
|
"kl": info_log["approx_kl"].item() if "approx_kl" in info_log else 0.0,
|
|
"clip": info_log["clipfrac"].item() if "clipfrac" in info_log else 0.0,
|
|
}
|
|
progress_bar.set_postfix(**logs)
|
|
accelerator.log(
|
|
{
|
|
"train_loss": train_loss,
|
|
"reward_mean": gathered_rewards.mean().item(),
|
|
"reward_std": gathered_rewards.std().item(),
|
|
**{k: v.item() for k, v in info_log.items()},
|
|
},
|
|
step=global_step,
|
|
)
|
|
train_loss = 0.0
|
|
info = {"approx_kl": [], "clipfrac": [], "policy_loss": [], "kl_loss": [], "loss": []}
|
|
|
|
transformer3d.eval()
|
|
|
|
# Move ref model back to CPU after training phase in low_vram mode
|
|
if args.low_vram and args.grpo_beta > 0:
|
|
ref_transformer3d.to('cpu')
|
|
torch.cuda.empty_cache()
|
|
|
|
# Update progress and global_step
|
|
progress_bar.update(1)
|
|
global_step += 1
|
|
|
|
# Reset sample buffer for next collection round
|
|
all_samples = []
|
|
|
|
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
|
|
removing_checkpoints = checkpoints[0:num_to_remove]
|
|
logger.info(f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints")
|
|
logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
|
|
for removing_checkpoint in removing_checkpoints:
|
|
removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
|
|
shutil.rmtree(removing_checkpoint)
|
|
gc.collect()
|
|
torch.cuda.empty_cache()
|
|
torch.cuda.ipc_collect()
|
|
# Keep the checkpoint out of the progress bar rate: a minute-long save would
|
|
# otherwise land in the next step's interval and be shown as a slow step. The
|
|
# save also stages the whole state in host RAM (safetensors materializes every
|
|
# tensor as bytes) and leaves the freed blocks in the allocator caches, so the
|
|
# cache flushes run inside the same window.
|
|
with progress_bar.paused():
|
|
if not args.save_state:
|
|
if args.use_peft_lora:
|
|
safetensor_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}.safetensors")
|
|
network_state_dict = get_peft_model_state_dict(accelerator.unwrap_model(transformer3d))
|
|
save_model(safetensor_save_path, network_state_dict)
|
|
safetensor_kohya_format_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}-compatible_with_comfyui.safetensors")
|
|
network_state_dict_kohya = convert_peft_lora_to_kohya_lora(network_state_dict)
|
|
save_model(safetensor_kohya_format_save_path, network_state_dict_kohya)
|
|
logger.info(f"Saved safetensor to {safetensor_save_path}")
|
|
else:
|
|
safetensor_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}.safetensors")
|
|
save_model(safetensor_save_path, accelerator.unwrap_model(network))
|
|
logger.info(f"Saved safetensor to {safetensor_save_path}")
|
|
else:
|
|
accelerator_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
|
|
accelerator.save_state(accelerator_save_path)
|
|
logger.info(f"Saved state to {accelerator_save_path}")
|
|
gc.collect()
|
|
torch.cuda.empty_cache()
|
|
torch.cuda.ipc_collect()
|
|
|
|
if args.validation_prompts is not None and global_step % args.validation_steps == 0:
|
|
with progress_bar.paused():
|
|
log_validation(
|
|
vae,
|
|
text_encoder,
|
|
tokenizer,
|
|
transformer3d,
|
|
network,
|
|
args,
|
|
accelerator,
|
|
weight_dtype,
|
|
global_step,
|
|
)
|
|
|
|
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,
|
|
transformer3d,
|
|
network,
|
|
args,
|
|
accelerator,
|
|
weight_dtype,
|
|
global_step,
|
|
)
|
|
|
|
if global_step >= args.max_train_steps:
|
|
break
|
|
|
|
# Close the bar before the end-of-run checkpoint: tqdm keeps redrawing a live bar whenever
|
|
# something else writes to the console. PauseAwareTqdm.close() rebases the closing line onto
|
|
# the smoothed rate, so the worker warm-up and the first dataloader fetch do not dilute it.
|
|
progress_bar.close()
|
|
|
|
# Create the pipeline using the trained modules and save it.
|
|
accelerator.wait_for_everyone()
|
|
if args.use_deepspeed or args.use_fsdp or accelerator.is_main_process:
|
|
gc.collect()
|
|
torch.cuda.empty_cache()
|
|
torch.cuda.ipc_collect()
|
|
if not args.save_state:
|
|
if args.use_peft_lora:
|
|
safetensor_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}.safetensors")
|
|
network_state_dict = get_peft_model_state_dict(accelerator.unwrap_model(transformer3d))
|
|
save_model(safetensor_save_path, network_state_dict)
|
|
|
|
safetensor_kohya_format_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}-compatible_with_comfyui.safetensors")
|
|
network_state_dict_kohya = convert_peft_lora_to_kohya_lora(network_state_dict)
|
|
save_model(safetensor_kohya_format_save_path, network_state_dict_kohya)
|
|
logger.info(f"Saved safetensor to {safetensor_save_path}")
|
|
else:
|
|
safetensor_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}.safetensors")
|
|
save_model(safetensor_save_path, accelerator.unwrap_model(network))
|
|
logger.info(f"Saved safetensor to {safetensor_save_path}")
|
|
else:
|
|
accelerator_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
|
|
accelerator.save_state(accelerator_save_path)
|
|
logger.info(f"Saved state to {accelerator_save_path}")
|
|
|
|
accelerator.end_training()
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|