main_example.toml

# Output path for training runs. Each training run makes a new directory in here.
output_dir = '/data/diffusion_pipe_training_runs/852'

# Dataset config file.
dataset = '/data/shiv/diffusion-pipe/examples/852_dataset.toml'
# You can have separate eval datasets. Give them a name for Tensorboard metrics.
# eval_datasets = [
#     {name = 'something', config = 'path/to/eval_dataset.toml'},
# ]

# training settings

# I usually set this to a really high value because I don't know how long I want to train.
epochs = 1000
# Batch size of a single forward/backward pass for one GPU.
micro_batch_size_per_gpu = 16
# Pipeline parallelism degree. A single instance of the model is divided across this many GPUs.
pipeline_stages = 1
# Number of micro-batches sent through the pipeline for each training step.
# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
gradient_accumulation_steps = 8
# Grad norm clipping.
gradient_clipping = 1.0
# Learning rate warmup.
warmup_steps = 5

# Block swapping is supported for Wan, HunyuanVideo, Flux, and Chroma. This value controls the number
# of blocks kept offloaded to RAM. Increasing it lowers VRAM use, but has a performance penalty. The
# exactly performance penalty depends on the model and the type of training you are doing (e.g. images vs video).
# Block swapping only works for LoRA training, and requires pipeline_stages=1.
blocks_to_swap = 24 

# eval settings

eval_every_n_epochs = 1
eval_before_first_step = false
# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
eval_micro_batch_size_per_gpu = 1
eval_gradient_accumulation_steps = 1
# If using block swap, you can disable it for eval. Eval uses less memory, so depending on block swapping amount you can maybe get away with
# doing this, and then eval is much faster.
disable_block_swap_for_eval = true

# misc settings

# Probably want to set this a bit higher if you have a smaller dataset so you don't end up with a million saved models.
save_every_n_epochs = 1
# Can checkpoint the training state every n number of epochs or minutes. Set only one of these. You can resume from checkpoints using the --resume_from_checkpoint flag.
#checkpoint_every_n_epochs = 1
checkpoint_every_n_minutes = 60
# Always set to true unless you have a huge amount of VRAM.
# This can also be 'unsloth' to reduce VRAM even more, with a slight performance hit.
activation_checkpointing = true

# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
partition_method = 'parameters'
# Alternatively you can use 'manual' in combination with partition_split, which specifies the split points for dividing
# layers between GPUs. For example, with two GPUs, partition_split=[10] puts layers 0-9 on GPU 0, and the rest on GPU 1.
# With three GPUs, partition_split=[10, 20] puts layers 0-9 on GPU 0, layers 10-19 on GPU 1, and the rest on GPU 2.
# Length of partition_split must be pipeline_stages-1.
#partition_split = [N]

# dtype for saving the LoRA or model, if different from training dtype
save_dtype = 'bfloat16'
# If experiencing CUDA OOM errors with memory fragmentation, set this env var: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
caching_batch_size = 8 
# How often deepspeed logs to console.
steps_per_print = 100
# How to extract video clips for training from a single input video file.
# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
# number of frames for that bucket.
# single_beginning: one clip starting at the beginning of the video
# single_middle: one clip from the middle of the video (cutting off the start and end equally)
# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
# default is single_beginning
video_clip_mode = 'single_beginning'

# This is how you configure HunyuanVideo. Other models will be different. See docs/supported_models.md for
# details on the configuration and options for each model.
[model]
type = 'wan'
ckpt_path = '/data/shiv/diffusion-pipe/models/wan1.3'
# Base dtype used for all models.
dtype = 'bfloat16'
# Wan supports fp8 for the transformer when training LoRA.
# transformer_dtype = 'float8'
# How to sample timesteps to train on. Can be logit_normal or uniform.
timestep_sample_method = 'logit_normal'

# For models that support full fine tuning, simply delete or comment out the [adapter] table to FFT.
[adapter]
type = 'lora'
# Dtype for the LoRA weights you are training.
dtype = 'bfloat16'
# LoRA rank - determines capacity to learn (higher = more capacity, more VRAM usage)
rank = 48 

# Higher rank = more capacity to learn but uses more VRAM and may overfit
# The alpha value (scaling factor) is automatically set equal to rank in this codebase
# You can initialize the lora weights from a previously trained lora.
init_from_existing = '/data/diffusion_pipe_training_runs/852/20250411_21-03-36/epoch3'

[optimizer]
# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
# Look at train.py for other options. You could also easily edit the file and add your own.
type = 'adamw_optimi'
lr = 1e-5
betas = [0.9, 0.99]
weight_decay = 0.01