Files
hao-ai-lab-FastVideo/scripts/train_t2v.sh
T
Peiyuan Zhang 8ec82cbc37 Generate synthetic dataset
Delete Open Sora Plan Modeling

amend log validation

Change name to fast video

Remove OSP modeling

clean up name changing

remove files

 Deleted unnecessary files

commit first

commit first

training

ok

Update generate_synthetic.sh and deepspeed_zero2_config.yaml

14

debug

overfitting ....

debugging ..

Debug successful!

Add zero3

OK

update optimizer

load

small bug

random seed args

sp enable & still has bug

rename

update inference sp code / can run / still has bug / don't output normal mp4

switch to deepspeed dummyoptim

fix some bugs; still output green

SP inference done!

fix typos in readme and latent dataset debug file
2024-10-28 00:44:02 +00:00

66 lines
1.9 KiB
Bash

# export WANDB_MODE="offline"
export ENTITY="linbin"
export HF_DATASETS_OFFLINE=1
export TRANSFORMERS_OFFLINE=1
export PDSH_RCMD_TYPE=ssh
# NCCL setting
export GLOO_SOCKET_IFNAME=bond0
export NCCL_SOCKET_IFNAME=bond0
export NCCL_IB_HCA=mlx5_10:1,mlx5_11:1,mlx5_12:1,mlx5_13:1
export NCCL_IB_GID_INDEX=3
export NCCL_IB_TC=162
export NCCL_IB_TIMEOUT=22
export NCCL_PXN_DISABLE=0
export NCCL_IB_QPS_PER_CONNECTION=4
export NCCL_ALGO=Ring
export OMP_NUM_THREADS=1
export MKL_NUM_THREADS=1
# export NCCL_ALGO=Tree
accelerate launch \
--config_file scripts/accelerate_configs/deepspeed_zero2_config.yaml \
opensora/train/train_t2v_diffusers.py \
--model OpenSoraT2V-ROPE-L/122 \
--text_encoder_name google/mt5-xxl \
--cache_dir "./cache_dir" \
--dataset t2v \
--data "scripts/train_data/merge_data.txt" \
--ae CausalVAEModel_D4_4x8x8 \
--ae_path "/storage/dataset/test140k" \
--sample_rate 1 \
--num_frames 29 \
--max_height 480 \
--max_width 640 \
--interpolation_scale_t 1.0 \
--interpolation_scale_h 1.0 \
--interpolation_scale_w 1.0 \
--attention_mode xformers \
--gradient_checkpointing \
--train_batch_size=1 \
--dataloader_num_workers 10 \
--gradient_accumulation_steps=1 \
--max_train_steps=1000000 \
--learning_rate=1e-4 \
--lr_scheduler="constant" \
--lr_warmup_steps=0 \
--mixed_precision="bf16" \
--report_to="wandb" \
--checkpointing_steps=250 \
--allow_tf32 \
--text_max_length 512 \
--use_image_num 0 \
--tile_sample_stride_height 0.125 \
--snr_gamma 5.0 \
--use_ema \
--ema_start_step 0 \
--cfg 0.1 \
--noise_offset 0.02 \
--use_rope \
--resume_from_checkpoint="latest" \
--ema_decay 0.999 \
--enable_tiling \
--speed_factor 1.0 \
--group_frame \
--sp_size 8 \
--train_sp_batch_size 2 \
--output_dir="bs16x8x1_29x480p_lr1e-4_snr5_ema999_opensora122_rope_mt5xxl_speed1.0"