43 lines
1.0 KiB
Bash
43 lines
1.0 KiB
Bash
#!/bin/bash
|
|
|
|
model_name=flux-dev-fill-lora
|
|
train_data_root='configs/data/visualcloze.yaml'
|
|
gpu_num=8
|
|
node_num=1
|
|
batch_size=16
|
|
micro_batch_size=2
|
|
lr=1e-4
|
|
precision=bf16
|
|
lora_rank=256
|
|
snr_type=lognorm
|
|
training_type="lora"
|
|
grid_resolution=384
|
|
exp_name=visualcloze_${node_num}x${gpu_num}_bs${batch_size}_mbs${micro_batch_size}_rank${lora_rank}_lr${lr}_${grid_resolution}
|
|
results_dir=./output/${exp_name}
|
|
mkdir -p ${results_dir}
|
|
|
|
torchrun --nproc_per_node=${gpu_num} --nnodes=${node_num} --master_port 29339 train.py \
|
|
--global_bs ${batch_size} \
|
|
--micro_bs ${micro_batch_size} \
|
|
--data_path ${train_data_root} \
|
|
--results_dir ${results_dir} \
|
|
--lr ${lr} \
|
|
--grad_clip 2.0 \
|
|
--grid_resolution ${grid_resolution} \
|
|
--lora_rank ${lora_rank} \
|
|
--data_parallel fsdp \
|
|
--max_steps 1000000 \
|
|
--ckpt_every 5000 \
|
|
--log_every 1 \
|
|
--precision ${precision} \
|
|
--grad_precision fp32 \
|
|
--global_seed 20240826 \
|
|
--num_workers 4 \
|
|
--snr_type ${snr_type} \
|
|
--training_type ${training_type} \
|
|
--debug \
|
|
--load_t5 \
|
|
--load_clip \
|
|
--model_name ${model_name} \
|
|
--checkpointing
|