update

fastvideo/v1/tests/slurm/training/wan_t2v_1.3b/finetune.sh

@@ -0,0 +1,51 @@
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
+
+DATA_DIR="data/crush-smol_processed_main_t2v/latents/combined_parquet_dataset"
+VALIDATION_DIR="data/crush-smol_processed_main_t2v/latents/validation_parquet_dataset"
+NUM_GPUS=4
+# 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 $NUM_GPUS\
+    fastvideo/v1/training/wan_training_pipeline.py\
+    --model_path Wan-AI/Wan2.1-T2V-1.3B-Diffusers \
+    --inference_mode False\
+    --pretrained_model_name_or_path Wan-AI/Wan2.1-T2V-1.3B-Diffusers \
+    --data_path "$DATA_DIR"\
+    --validation_preprocessed_path "$VALIDATION_DIR"\
+    --train_batch_size=1 \
+    --num_latent_t 8 \
+    --sp_size 4 \
+    --tp_size 4 \
+    --hsdp_replicate_dim 1 \
+    --hsdp_shard_dim 4 \
+    --num_gpus $NUM_GPUS \
+    --train_sp_batch_size 1\
+    --dataloader_num_workers 1\
+    --gradient_accumulation_steps=8 \
+    --max_train_steps=5000 \
+    --learning_rate=1e-5\
+    --mixed_precision="bf16"\
+    --checkpointing_steps=6000 \
+    --validation_steps 50\
+    --validation_sampling_steps "50" \
+    --log_validation \
+    --checkpoints_total_limit 3\
+    --allow_tf32\
+    --ema_start_step 0\
+    --cfg 0.0\
+    --output_dir="$DATA_DIR/outputs/wan_finetune"\
+    --tracker_project_name wan_finetune \
+    --num_height 480 \
+    --num_width 832 \
+    --num_frames  77 \
+    --validation_guidance_scale "1.0" \
+    --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

fastvideo/v1/tests/slurm/training/wan_t2v_1.3b/preprocess.sh

@@ -0,0 +1,24 @@
+# export WANDB_MODE="offline"
+GPU_NUM=1 # 2,4,8
+MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+MODEL_TYPE="wan"
+DATA_MERGE_PATH="data/crush-smol/merge.txt"
+OUTPUT_DIR="data/crush-smol_processed_main_t2v/latents"
+VALIDATION_PATH="examples/training/finetune/wan_t2v_1.3b/crush_smol/validation.json"
+
+torchrun --nproc_per_node=$GPU_NUM \
+    fastvideo/v1/pipelines/preprocess/v1_preprocess.py \
+    --model_path $MODEL_PATH \
+    --data_merge_path $DATA_MERGE_PATH \
+    --preprocess_video_batch_size 8 \
+    --max_height 480 \
+    --max_width 832 \
+    --num_frames 77 \
+    --dataloader_num_workers 0 \
+    --output_dir=$OUTPUT_DIR \
+    --model_type $MODEL_TYPE \
+    --train_fps 16 \
+    --validation_dataset_file $VALIDATION_PATH \
+    --samples_per_file 8 \
+    --flush_frequency 8 \
+    --preprocess_task "t2v"

pyproject.toml

@@ -19,7 +19,7 @@ dependencies = [
 
     # Machine Learning & Transformers
     "transformers>=4.46.1", "tokenizers>=0.20.1", "sentencepiece==0.2.0",
-    "timm==1.0.11", "peft==0.13.2", "diffusers>=0.33.1", "bitsandbytes",
+    "timm==1.0.11", "peft==0.15.0", "diffusers>=0.33.1", "bitsandbytes",
     "torch==2.7.1", "torchvision",
 
     # Acceleration & Optimization