training_scripts/tp-maliketh

#!/usr/bin/env zsh
#
# >>> conda initialize >>>
# !! Contents within this block are managed by 'conda init' !!
__conda_setup="$('/home/kade/miniconda3/bin/conda' 'shell.zsh' 'hook' 2> /dev/null)"
if [ $? -eq 0 ]; then
    eval "$__conda_setup"
else
    if [ -f "/home/kade/miniconda3/etc/profile.d/conda.sh" ]; then
        . "/home/kade/miniconda3/etc/profile.d/conda.sh"
    else
        export PATH="/home/kade/miniconda3/bin:$PATH"
    fi
fi
unset __conda_setup
# <<< conda initialize <

conda activate sdscripts

NAME="maliketh-v2s2000"
TRAINING_DIR="/home/kade/datasets/maliketh"
OUTPUT_DIR="/home/kade/output_dir"

# Extract the number of steps from the NAME
STEPS=$(echo $NAME | grep -oE '[0-9]+$')

# If no number is found at the end of NAME, set a default value
if [ -z "$STEPS" ]; then
    STEPS=4096
    echo "No step count found in NAME. Using default value of \e[35m$STEPS\e[0m"
else
    echo "Extracted \e[35m$STEPS\e[0m steps from NAME"
fi

# alpha=1 @ dim=16 is the same lr than alpha=4 @ dim=256
# --min_snr_gamma=1
args=(
    # ⚠️  TODO: Benchmark...
    --debiased_estimation_loss
    # ⚠️  TODO: What does this do? Does it even work?
    --max_token_length=225
    # Keep Tokens
    --keep_tokens=1
    --keep_tokens_separator="|||"
    # Model
    --pretrained_model_name_or_path=/home/kade/ComfyUI/models/checkpoints/ponyDiffusionV6XL_v6StartWithThisOne.safetensors
    # Output, logging
    --output_dir="$OUTPUT_DIR/$NAME"
    --output_name="$NAME"
    --log_prefix="$NAME-"
    --log_with=tensorboard
    --logging_dir="$OUTPUT_DIR/logs"
    --seed=1728871242
    # Dataset
    --train_data_dir="$TRAINING_DIR"
    --dataset_repeats=1
    --resolution="1024,1024"
    --enable_bucket
    --bucket_reso_steps=64
    --min_bucket_reso=256
    --max_bucket_reso=2048
    --flip_aug
    --shuffle_caption
    --cache_latents
    --cache_latents_to_disk
    --max_data_loader_n_workers=8
    --persistent_data_loader_workers
    # Network config
    --network_dim=100000
    # ⚠️  TODO: Plot
    --network_alpha=0.0625
    --network_module="lycoris.kohya"
    --network_args
               "preset=full"
               "conv_dim=100000"
               "decompose_both=False"
               "conv_alpha=1"
               "rank_dropout=0"
               "module_dropout=0"
               "use_tucker=True"
               "use_scalar=False"
               "rank_dropout_scale=False"
               "algo=lokr"
               "bypass_mode=False"
               "factor=16"
               "dora_wd=True"
               "train_norm=False"
    --network_dropout=0
    # Optimizer config
    --optimizer_type=ClybW
    --train_batch_size=14
    #--gradient_accumulation_steps=1
    --max_grad_norm=1
    --gradient_checkpointing
    #--scale_weight_norms=1
    # LR Scheduling
    --max_train_steps=$STEPS
    --lr_warmup_steps=0
    --learning_rate=0.0001
    --unet_lr=0.0002
    --text_encoder_lr=0.0001
    --lr_scheduler="cosine"
    --lr_scheduler_args="num_cycles=0.375"
    # Noise
    --multires_noise_iterations=12
    --multires_noise_discount=0.4
    #--min_snr_gamma=1
    # Optimization, details
    --no_half_vae
    --sdpa
    --mixed_precision="bf16"
    # Saving
    --save_model_as="safetensors"
    --save_precision="fp16"
    --save_every_n_steps=100
    # Saving States
    #--save_state
    # Either resume from a saved state
    #--resume="$OUTPUT_DIR/wolflink-vfucks400" # Resume from saved state
    #--skip_until_initial_step
    # Or from a checkpoint
    #--network_weights="$OUTPUT_DIR/wolflink-vfucks400/wolflink-vfucks400-step00000120.safetensors" # Resume from checkpoint (not needed with state, i think)
    #--initial_step=120
    # Sampling
    --sample_every_n_steps=1
    --sample_prompts="$TRAINING_DIR/sample-prompts.txt"
    --sample_sampler="euler_a"
    --sample_at_first
    --caption_extension=".txt"
)
cd ~/source/repos/sd-scripts
#accelerate launch --num_cpu_threads_per_process=2
python "./sdxl_train_network.py" "${args[@]}"
cd ~