Merge pull request #1272 from modelscope/zero3-fix

Support DeepSpeed ZeRO 3

examples/flux/model_training/full/accelerate_config_zero3.yaml

@@ -0,0 +1,23 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  gradient_accumulation_steps: 1
+  offload_optimizer_device: none
+  offload_param_device: none
+  zero3_init_flag: true
+  zero3_save_16bit_model: true
+  zero_stage: 3
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/flux2/model_training/full/accelerate_config_zero3.yaml

@@ -0,0 +1,23 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  gradient_accumulation_steps: 1
+  offload_optimizer_device: none
+  offload_param_device: none
+  zero3_init_flag: true
+  zero3_save_16bit_model: true
+  zero_stage: 3
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/flux2/model_training/special/npu_training/FLUX.2-dev-Lora-NPU.sh

@@ -0,0 +1,36 @@
+export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
+export CPU_AFFINITY_CONF=1
+
+accelerate launch examples/flux2/model_training/train.py \
+  --dataset_base_path data/example_image_dataset \
+  --dataset_metadata_path data/example_image_dataset/metadata.csv \
+  --max_pixels 1048576 \
+  --dataset_repeat 1 \
+  --model_id_with_origin_paths "black-forest-labs/FLUX.2-dev:text_encoder/*.safetensors,black-forest-labs/FLUX.2-dev:vae/diffusion_pytorch_model.safetensors" \
+  --learning_rate 1e-4 \
+  --num_epochs 5 \
+  --remove_prefix_in_ckpt "pipe.dit." \
+  --output_path "./models/train/FLUX.2-dev-LoRA-splited-cache" \
+  --lora_base_model "dit" \
+  --lora_target_modules "to_q,to_k,to_v,add_q_proj,add_k_proj,add_v_proj,to_qkv_mlp_proj,to_out.0,to_add_out,linear_in,linear_out,single_transformer_blocks.0.attn.to_out,single_transformer_blocks.1.attn.to_out,single_transformer_blocks.2.attn.to_out,single_transformer_blocks.3.attn.to_out,single_transformer_blocks.4.attn.to_out,single_transformer_blocks.5.attn.to_out,single_transformer_blocks.6.attn.to_out,single_transformer_blocks.7.attn.to_out,single_transformer_blocks.8.attn.to_out,single_transformer_blocks.9.attn.to_out,single_transformer_blocks.10.attn.to_out,single_transformer_blocks.11.attn.to_out,single_transformer_blocks.12.attn.to_out,single_transformer_blocks.13.attn.to_out,single_transformer_blocks.14.attn.to_out,single_transformer_blocks.15.attn.to_out,single_transformer_blocks.16.attn.to_out,single_transformer_blocks.17.attn.to_out,single_transformer_blocks.18.attn.to_out,single_transformer_blocks.19.attn.to_out,single_transformer_blocks.20.attn.to_out,single_transformer_blocks.21.attn.to_out,single_transformer_blocks.22.attn.to_out,single_transformer_blocks.23.attn.to_out,single_transformer_blocks.24.attn.to_out,single_transformer_blocks.25.attn.to_out,single_transformer_blocks.26.attn.to_out,single_transformer_blocks.27.attn.to_out,single_transformer_blocks.28.attn.to_out,single_transformer_blocks.29.attn.to_out,single_transformer_blocks.30.attn.to_out,single_transformer_blocks.31.attn.to_out,single_transformer_blocks.32.attn.to_out,single_transformer_blocks.33.attn.to_out,single_transformer_blocks.34.attn.to_out,single_transformer_blocks.35.attn.to_out,single_transformer_blocks.36.attn.to_out,single_transformer_blocks.37.attn.to_out,single_transformer_blocks.38.attn.to_out,single_transformer_blocks.39.attn.to_out,single_transformer_blocks.40.attn.to_out,single_transformer_blocks.41.attn.to_out,single_transformer_blocks.42.attn.to_out,single_transformer_blocks.43.attn.to_out,single_transformer_blocks.44.attn.to_out,single_transformer_blocks.45.attn.to_out,single_transformer_blocks.46.attn.to_out,single_transformer_blocks.47.attn.to_out" \
+  --lora_rank 32 \
+  --use_gradient_checkpointing \
+  --dataset_num_workers 8 \
+  --task "sft:data_process"
+
+accelerate launch --config_file examples/flux2/model_training/full/accelerate_config_zero3.yaml examples/flux2/model_training/train.py \
+  --dataset_base_path "./models/train/FLUX.2-dev-LoRA-splited-cache" \
+  --max_pixels 1048576 \
+  --dataset_repeat 50 \
+  --model_id_with_origin_paths "black-forest-labs/FLUX.2-dev:transformer/*.safetensors" \
+  --learning_rate 1e-4 \
+  --num_epochs 5 \
+  --remove_prefix_in_ckpt "pipe.dit." \
+  --output_path "./models/train/FLUX.2-dev-LoRA-splited" \
+  --lora_base_model "dit" \
+  --lora_target_modules "to_q,to_k,to_v,add_q_proj,add_k_proj,add_v_proj,to_qkv_mlp_proj,to_out.0,to_add_out,linear_in,linear_out,single_transformer_blocks.0.attn.to_out,single_transformer_blocks.1.attn.to_out,single_transformer_blocks.2.attn.to_out,single_transformer_blocks.3.attn.to_out,single_transformer_blocks.4.attn.to_out,single_transformer_blocks.5.attn.to_out,single_transformer_blocks.6.attn.to_out,single_transformer_blocks.7.attn.to_out,single_transformer_blocks.8.attn.to_out,single_transformer_blocks.9.attn.to_out,single_transformer_blocks.10.attn.to_out,single_transformer_blocks.11.attn.to_out,single_transformer_blocks.12.attn.to_out,single_transformer_blocks.13.attn.to_out,single_transformer_blocks.14.attn.to_out,single_transformer_blocks.15.attn.to_out,single_transformer_blocks.16.attn.to_out,single_transformer_blocks.17.attn.to_out,single_transformer_blocks.18.attn.to_out,single_transformer_blocks.19.attn.to_out,single_transformer_blocks.20.attn.to_out,single_transformer_blocks.21.attn.to_out,single_transformer_blocks.22.attn.to_out,single_transformer_blocks.23.attn.to_out,single_transformer_blocks.24.attn.to_out,single_transformer_blocks.25.attn.to_out,single_transformer_blocks.26.attn.to_out,single_transformer_blocks.27.attn.to_out,single_transformer_blocks.28.attn.to_out,single_transformer_blocks.29.attn.to_out,single_transformer_blocks.30.attn.to_out,single_transformer_blocks.31.attn.to_out,single_transformer_blocks.32.attn.to_out,single_transformer_blocks.33.attn.to_out,single_transformer_blocks.34.attn.to_out,single_transformer_blocks.35.attn.to_out,single_transformer_blocks.36.attn.to_out,single_transformer_blocks.37.attn.to_out,single_transformer_blocks.38.attn.to_out,single_transformer_blocks.39.attn.to_out,single_transformer_blocks.40.attn.to_out,single_transformer_blocks.41.attn.to_out,single_transformer_blocks.42.attn.to_out,single_transformer_blocks.43.attn.to_out,single_transformer_blocks.44.attn.to_out,single_transformer_blocks.45.attn.to_out,single_transformer_blocks.46.attn.to_out,single_transformer_blocks.47.attn.to_out" \
+  --lora_rank 32 \
+  --use_gradient_checkpointing \
+  --dataset_num_workers 8 \
+  --initialize_model_on_cpu \
+  --task "sft:train"

examples/flux2/model_training/special/npu_training/FLUX.2-klein-9B-NPU.sh

@@ -0,0 +1,34 @@
+# This script is tested on 8*910B(NPU)
+export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
+export CPU_AFFINITY_CONF=1
+
+accelerate launch --config_file examples/flux2/model_training/full/accelerate_config.yaml examples/flux2/model_training/train.py \
+  --dataset_base_path data/example_image_dataset \
+  --dataset_metadata_path data/example_image_dataset/metadata.csv \
+  --max_pixels 1048576 \
+  --dataset_repeat 50 \
+  --model_id_with_origin_paths "black-forest-labs/FLUX.2-klein-9B:text_encoder/*.safetensors,black-forest-labs/FLUX.2-klein-9B:transformer/*.safetensors,black-forest-labs/FLUX.2-klein-9B:vae/diffusion_pytorch_model.safetensors" \
+  --tokenizer_path "black-forest-labs/FLUX.2-klein-9B:tokenizer/" \
+  --learning_rate 1e-5 \
+  --num_epochs 2 \
+  --remove_prefix_in_ckpt "pipe.dit." \
+  --output_path "./models/train/FLUX.2-klein-9B_full" \
+  --trainable_models "dit" \
+  --use_gradient_checkpointing
+
+# Edit
+# accelerate launch --config_file examples/flux2/model_training/full/accelerate_config.yaml examples/flux2/model_training/train.py \
+#   --dataset_base_path data/example_image_dataset \
+#   --dataset_metadata_path data/example_image_dataset/metadata_qwen_imgae_edit_multi.json \
+#   --data_file_keys "image,edit_image" \
+#   --extra_inputs "edit_image" \
+#   --max_pixels 1048576 \
+#   --dataset_repeat 50 \
+#   --model_id_with_origin_paths "black-forest-labs/FLUX.2-klein-9B:text_encoder/*.safetensors,black-forest-labs/FLUX.2-klein-9B:transformer/*.safetensors,black-forest-labs/FLUX.2-klein-9B:vae/diffusion_pytorch_model.safetensors" \
+#   --tokenizer_path "black-forest-labs/FLUX.2-klein-9B:tokenizer/" \
+#   --learning_rate 1e-5 \
+#   --num_epochs 2 \
+#   --remove_prefix_in_ckpt "pipe.dit." \
+#   --output_path "./models/train/FLUX.2-klein-9B_full" \
+#   --trainable_models "dit" \
+#   --use_gradient_checkpointing

examples/flux2/model_training/train.py

@@ -85,6 +85,7 @@ def flux2_parser():
     parser = add_general_config(parser)
     parser = add_image_size_config(parser)
     parser.add_argument("--tokenizer_path", type=str, default=None, help="Path to tokenizer.")
+    parser.add_argument("--initialize_model_on_cpu", default=False, action="store_true", help="Whether to initialize models on CPU.")
     return parser
 
 
@@ -126,7 +127,7 @@ if __name__ == "__main__":
         fp8_models=args.fp8_models,
         offload_models=args.offload_models,
         task=args.task,
-        device=accelerator.device,
+        device="cpu" if args.initialize_model_on_cpu else accelerator.device,
     )
     model_logger = ModelLogger(
         args.output_path,

examples/qwen_image/model_training/full/accelerate_config_zero3.yaml

@@ -0,0 +1,23 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  gradient_accumulation_steps: 1
+  offload_optimizer_device: none
+  offload_param_device: none
+  zero3_init_flag: true
+  zero3_save_16bit_model: true
+  zero_stage: 3
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/qwen_image/model_training/special/npu_training/Qwen-Image-Edit-2509-NPU.sh

@@ -0,0 +1,20 @@
+# This script was tested using zero3 and on 8*910B(NPU)
+export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
+export CPU_AFFINITY_CONF=1
+
+accelerate launch --config_file examples/qwen_image/model_training/full/accelerate_config_zero3.yaml examples/qwen_image/model_training/train.py \
+  --dataset_base_path data/example_image_dataset \
+  --dataset_metadata_path data/example_image_dataset/metadata_qwen_imgae_edit_multi.json \
+  --data_file_keys "image,edit_image" \
+  --extra_inputs "edit_image" \
+  --max_pixels 1048576 \
+  --dataset_repeat 50 \
+  --model_id_with_origin_paths "Qwen/Qwen-Image-Edit-2509:transformer/diffusion_pytorch_model*.safetensors,Qwen/Qwen-Image:text_encoder/model*.safetensors,Qwen/Qwen-Image:vae/diffusion_pytorch_model.safetensors" \
+  --learning_rate 1e-5 \
+  --num_epochs 2 \
+  --remove_prefix_in_ckpt "pipe.dit." \
+  --output_path "./models/train/Qwen-Image-Edit-2509_full" \
+  --trainable_models "dit" \
+  --use_gradient_checkpointing \
+  --find_unused_parameters \
+  --initialize_model_on_cpu

examples/qwen_image/model_training/train.py

@@ -101,6 +101,7 @@ def qwen_image_parser():
     parser.add_argument("--tokenizer_path", type=str, default=None, help="Path to tokenizer.")
     parser.add_argument("--processor_path", type=str, default=None, help="Path to the processor. If provided, the processor will be used for image editing.")
     parser.add_argument("--zero_cond_t", default=False, action="store_true", help="A special parameter introduced by Qwen-Image-Edit-2511. Please enable it for this model.")
+    parser.add_argument("--initialize_model_on_cpu", default=False, action="store_true", help="Whether to initialize models on CPU.")
     return parser
 
 
@@ -151,7 +152,7 @@ if __name__ == "__main__":
         fp8_models=args.fp8_models,
         offload_models=args.offload_models,
         task=args.task,
-        device=accelerator.device,
+        device="cpu" if args.initialize_model_on_cpu else accelerator.device,
         zero_cond_t=args.zero_cond_t,
     )
     model_logger = ModelLogger(

examples/wanvideo/model_training/full/accelerate_config_zero3.yaml

@@ -0,0 +1,23 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  gradient_accumulation_steps: 1
+  offload_optimizer_device: none
+  offload_param_device: none
+  zero3_init_flag: true
+  zero3_save_16bit_model: true
+  zero_stage: 3
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/z_image/model_training/full/accelerate_config_zero3.yaml

@@ -0,0 +1,23 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  gradient_accumulation_steps: 1
+  offload_optimizer_device: none
+  offload_param_device: none
+  zero3_init_flag: true
+  zero3_save_16bit_model: true
+  zero_stage: 3
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false