Upload folder using huggingface_hub

configs/accelerate_config.yaml

@@ -1,10 +1,14 @@
-# Flux LoRA Fine-tuning Config
-# Run with: accelerate launch --config_file configs/accelerate_config.yaml scripts/training/train_flux_lora.py
+# Flux LoRA Fine-tuning Config - DeepSpeed ZeRO-3
+# Run with: accelerate launch --config_file configs/accelerate_config.yaml scripts/training/train_flux_deepspeed.py
 
 compute_environment: LOCAL_MACHINE
-distributed_type: MULTI_GPU
+distributed_type: DEEPSPEED
+deepspeed_config:
+  deepspeed_config_file: /home/adminuser/chungcat/configs/deepspeed_zero3.json
+  zero3_init_flag: true
+  zero3_save_16bit_model: true
 num_machines: 1
-num_processes: 8
+num_processes: 2
 mixed_precision: bf16
 gpu_ids: all
 use_cpu: false

configs/deepspeed_zero3.json

@@ -0,0 +1,19 @@
+{
+  "bf16": {
+    "enabled": true
+  },
+  "zero_optimization": {
+    "stage": 3,
+    "overlap_comm": true,
+    "contiguous_gradients": true,
+    "reduce_bucket_size": 5e7,
+    "stage3_prefetch_bucket_size": 4e7,
+    "stage3_param_persistence_threshold": 1e5,
+    "stage3_gather_16bit_weights_on_model_save": true,
+    "sub_group_size": 1e9
+  },
+  "gradient_accumulation_steps": 4,
+  "gradient_clipping": 1.0,
+  "train_micro_batch_size_per_gpu": 4,
+  "wall_clock_breakdown": false
+}