[Training] [5/n] Add single gpu training pipeline (#447)

Co-authored-by: JerryZhou54 <zhouw.jerry2017@outlook.com> Co-authored-by: Wei Zhou <69577934+JerryZhou54@users.noreply.github.com> Co-authored-by: Kevin Lin <42618777+kevin314@users.noreply.github.com> Co-authored-by: “BrianChen1129” <yongqich@umich.edu>
2025-05-29 11:49:46 -07:00
parent 8e18dc9f71
commit 007e237e69
13 changed files with 1138 additions and 61 deletions
@@ -0,0 +1,49 @@
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
+
+DATA_DIR=data/HD-Mixkit-Finetune-Wan/combined_parquet_dataset
+VALIDATION_DIR=data/HD-Mixkit-Finetune-Wan/validation_parquet_dataset
+NUM_GPUS=1
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
+torchrun --nnodes 1 --nproc_per_node $NUM_GPUS\
+    fastvideo/v1/training/wan_training_pipeline.py\
+    --model_path Wan-AI/Wan2.1-T2V-1.3B-Diffusers \
+    --inference_mode False\
+    --pretrained_model_name_or_path Wan-AI/Wan2.1-T2V-1.3B-Diffusers \
+    --cache_dir "/home/ray/.cache"\
+    --data_path "$DATA_DIR"\
+    --validation_prompt_dir "$VALIDATION_DIR"\
+    --train_batch_size=1\
+    --num_latent_t 4 \
+    --sp_size $NUM_GPUS \
+    --tp_size $NUM_GPUS \
+    --train_sp_batch_size 1\
+    --dataloader_num_workers 5\
+    --gradient_accumulation_steps=1\
+    --max_train_steps=120 \
+    --learning_rate=1e-6\
+    --mixed_precision="bf16"\
+    --checkpointing_steps=10 \
+    --validation_steps 20\
+    --validation_sampling_steps "2,4,8" \
+    --log_validation \
+    --checkpoints_total_limit 3\
+    --allow_tf32\
+    --ema_start_step 0\
+    --cfg 0.0\
+    --output_dir="$DATA_DIR/outputs/wan_finetune"\
+    --tracker_project_name wan_finetune \
+    --num_height 480 \
+    --num_width 832 \
+    --num_frames  81 \
+    --shift 3 \
+    --validation_guidance_scale "1.0" \
+    --num_euler_timesteps 50 \
+    --multi_phased_distill_schedule "4000-1" \
+    --weight_decay 0.01 \
+    --not_apply_cfg_solver \
+    --master_weight_type "bf16"