wip

examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/finetune_i2v.sh

@@ -9,17 +9,19 @@ MODEL_PATH="weizhou03/Wan2.1-Fun-1.3B-InP-Diffusers"
 DATA_DIR="data/crush-smol_processed_i2v_1_3b_inp/combined_parquet_dataset/"
 VALIDATION_DATASET_FILE="examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/validation.json"
 NUM_GPUS=8
+export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
 
 # Training arguments
 training_args=(
   --tracker_project_name "wan_i2v_finetune"
-  --output_dir "$DATA_DIR/outputs/wan_i2v_finetune"
+  --output_dir "outputs/wan_i2v_finetune"
+  --wandb_run_name "wan_i2v_finetune"
   --max_train_steps 2000
-  --train_batch_size 4
+  --train_batch_size 1
   --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
+  --gradient_accumulation_steps 2
   --num_latent_t 8
   --num_height 480
   --num_width 832
@@ -30,10 +32,10 @@ training_args=(
 # Parallel arguments
 parallel_args=(
   --num_gpus $NUM_GPUS
-  --sp_size 4
-  --tp_size 4
-  --hsdp_replicate_dim 2
-  --hsdp_shard_dim 4
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1
+  --hsdp_shard_dim 8
 )
 
 # Model arguments
@@ -59,7 +61,7 @@ validation_args=(
 
 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 2e-5
+  --learning_rate 1e-5
   --mixed_precision "bf16"
   --checkpointing_steps 2000
   --weight_decay 1e-4