52 lines
1.7 KiB
Bash
52 lines
1.7 KiB
Bash
export WANDB_BASE_URL="https://api.wandb.ai"
|
|
export WANDB_MODE=online
|
|
export TOKENIZERS_PARALLELISM=false
|
|
# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
|
|
|
|
DATA_DIR=[your data dir]
|
|
VALIDATION_DATASET_FILE=[your validation dataset file]
|
|
NUM_GPUS=4
|
|
# export CUDA_VISIBLE_DEVICES=4,5
|
|
# IP=[MASTER NODE IP]
|
|
|
|
# Make sure that num_latent_t is a multiple of sp_size
|
|
torchrun --nnodes 1 --nproc_per_node $NUM_GPUS\
|
|
fastvideo/training/wan_training_pipeline.py\
|
|
--model_path Wan-AI/Wan2.1-T2V-1.3B-Diffusers \
|
|
--inference_mode False\
|
|
--pretrained_model_name_or_path Wan-AI/Wan2.1-T2V-1.3B-Diffusers \
|
|
--data_path "$DATA_DIR"\
|
|
--validation_dataset_file "$VALIDATION_DATASET_FILE"\
|
|
--train_batch_size=4 \
|
|
--num_latent_t 20 \
|
|
--sp_size 4 \
|
|
--tp_size 1 \
|
|
--hsdp_replicate_dim 1 \
|
|
--hsdp_shard_dim 4 \
|
|
--num_gpus $NUM_GPUS \
|
|
--train_sp_batch_size 1\
|
|
--dataloader_num_workers 10\
|
|
--gradient_accumulation_steps=1 \
|
|
--max_train_steps=5000 \
|
|
--learning_rate=1e-6\
|
|
--mixed_precision="bf16"\
|
|
--training_state_checkpointing_steps 6000 \
|
|
--validation_steps 200\
|
|
--validation_sampling_steps "2,4,8" \
|
|
--log_validation \
|
|
--checkpoints_total_limit 3\
|
|
--ema_start_step 0\
|
|
--training_cfg_rate 0.0\
|
|
--output_dir="$DATA_DIR/outputs/wan_finetune"\
|
|
--tracker_project_name wan_finetune \
|
|
--num_height 480 \
|
|
--num_width 832 \
|
|
--num_frames 81 \
|
|
--validation_guidance_scale "6.0" \
|
|
--num_euler_timesteps 50 \
|
|
--multi_phased_distill_schedule "4000-1" \
|
|
--weight_decay 0.01 \
|
|
--not_apply_cfg_solver \
|
|
--dit_precision "fp32" \
|
|
--max_grad_norm 1.0 \
|
|
--enable_gradient_checkpointing_type "full" |