[SDXL] save_file bug on DeepSpeed

Hi,

I met a bug on save_file when I used DeepSpeed.
How should I fix it?

```
override steps. steps for 1 epochs is / 指定エポックまでのステップ数: 18165
[2023-07-09 14:35:23,072] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed info: version=0.9.5, git-hash=unknown, git-branch=unknown
[2023-07-09 14:35:25,118] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed Flops Profiler Enabled: False
[2023-07-09 14:35:25,120] [INFO] [logging.py:96:log_dist] [Rank 0] Removing param_group that has no 'params' in the client Optimizer
[2023-07-09 14:35:25,120] [INFO] [logging.py:96:log_dist] [Rank 0] Using client Optimizer as basic optimizer
[2023-07-09 14:35:25,294] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed Basic Optimizer = AdamW
[2023-07-09 14:35:25,294] [INFO] [utils.py:54:is_zero_supported_optimizer] Checking ZeRO support for optimizer=AdamW type=
[2023-07-09 14:35:25,294] [INFO] [logging.py:96:log_dist] [Rank 0] Creating torch.bfloat16 ZeRO stage 2 optimizer
[2023-07-09 14:35:25,294] [INFO] [stage_1_and_2.py:133:__init__] Reduce bucket size 500,000,000
[2023-07-09 14:35:25,294] [INFO] [stage_1_and_2.py:134:__init__] Allgather bucket size 500,000,000
[2023-07-09 14:35:25,294] [INFO] [stage_1_and_2.py:135:__init__] CPU Offload: False
[2023-07-09 14:35:25,294] [INFO] [stage_1_and_2.py:136:__init__] Round robin gradient partitioning: False
Rank: 0 partition count [2] and sizes[(1283731842, False)]
Rank: 1 partition count [2] and sizes[(1283731842, False)]
[2023-07-09 14:35:28,817] [INFO] [utils.py:785:see_memory_usage] Before initializing optimizer states
[2023-07-09 14:35:28,817] [INFO] [utils.py:786:see_memory_usage] MA 9.73 GB Max_MA 12.12 GB CA 12.12 GB Max_CA 12 GB
[2023-07-09 14:35:28,817] [INFO] [utils.py:793:see_memory_usage] CPU Virtual Memory: used = 42.89 GB, percent = 34.1%
[2023-07-09 14:35:29,565] [INFO] [utils.py:785:see_memory_usage] After initializing optimizer states
[2023-07-09 14:35:29,565] [INFO] [utils.py:786:see_memory_usage] MA 19.29 GB Max_MA 33.64 GB CA 36.04 GB Max_CA 36 GB
[2023-07-09 14:35:29,565] [INFO] [utils.py:793:see_memory_usage] CPU Virtual Memory: used = 42.89 GB, percent = 34.1%
[2023-07-09 14:35:29,565] [INFO] [stage_1_and_2.py:488:__init__] optimizer state initialized
[2023-07-09 14:35:30,307] [INFO] [utils.py:785:see_memory_usage] After initializing ZeRO optimizer
[2023-07-09 14:35:30,308] [INFO] [utils.py:786:see_memory_usage] MA 19.29 GB Max_MA 19.29 GB CA 36.04 GB Max_CA 36 GB
[2023-07-09 14:35:30,308] [INFO] [utils.py:793:see_memory_usage] CPU Virtual Memory: used = 42.89 GB, percent = 34.1%
[2023-07-09 14:35:30,318] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed Final Optimizer = AdamW
[2023-07-09 14:35:30,318] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed using client LR scheduler
[2023-07-09 14:35:30,318] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed LR Scheduler = None
[2023-07-09 14:35:30,318] [INFO] [logging.py:96:log_dist] [Rank 0] step=0, skipped=0, lr=[5e-06], mom=[(0.9, 0.999)]
[2023-07-09 14:35:30,320] [INFO] [config.py:960:print] DeepSpeedEngine configuration:
[2023-07-09 14:35:30,320] [INFO] [config.py:964:print] activation_checkpointing_config {
"partition_activations": false,
"contiguous_memory_optimization": false,
"cpu_checkpointing": false,
"number_checkpoints": null,
"synchronize_checkpoint_boundary": false,
"profile": false
}
[2023-07-09 14:35:30,320] [INFO] [config.py:964:print] aio_config ................... {'block_size': 1048576, 'queue_depth': 8, 'thread_count': 1, 'single_submit': False, 'overlap_events': True}
[2023-07-09 14:35:30,320] [INFO] [config.py:964:print] amp_enabled .................. False
[2023-07-09 14:35:30,320] [INFO] [config.py:964:print] amp_params ................... False
[2023-07-09 14:35:30,320] [INFO] [config.py:964:print] autotuning_config ............ {
"enabled": false,
"start_step": null,
"end_step": null,
"metric_path": null,
"arg_mappings": null,
"metric": "throughput",
"model_info": null,
"results_dir": "autotuning_results",
"exps_dir": "autotuning_exps",
"overwrite": true,
"fast": true,
"start_profile_step": 3,
"end_profile_step": 5,
"tuner_type": "gridsearch",
"tuner_early_stopping": 5,
"tuner_num_trials": 50,
"model_info_path": null,
"mp_size": 1,
"max_train_batch_size": null,
"min_train_batch_size": 1,
"max_train_micro_batch_size_per_gpu": 1.024000e+03,
"min_train_micro_batch_size_per_gpu": 1,
"num_tuning_micro_batch_sizes": 3
}
[2023-07-09 14:35:30,320] [INFO] [config.py:964:print] bfloat16_enabled ............. True
[2023-07-09 14:35:30,320] [INFO] [config.py:964:print] checkpoint_parallel_write_pipeline False
[2023-07-09 14:35:30,320] [INFO] [config.py:964:print] checkpoint_tag_validation_enabled True
[2023-07-09 14:35:30,320] [INFO] [config.py:964:print] checkpoint_tag_validation_fail False
[2023-07-09 14:35:30,320] [INFO] [config.py:964:print] comms_config .................
[2023-07-09 14:35:30,320] [INFO] [config.py:964:print] communication_data_type ...... None
[2023-07-09 14:35:30,320] [INFO] [config.py:964:print] compression_config ........... {'weight_quantization': {'shared_parameters': {'enabled': False, 'quantizer_kernel': False, 'schedule_offset': 0, 'quantize_groups': 1, 'quantize_verbose': False, 'quantization_type': 'symmetric', 'quantize_weight_in_forward': False, 'rounding': 'nearest', 'fp16_mixed_quantize': False, 'quantize_change_ratio': 0.001}, 'different_groups': {}}, 'activation_quantization': {'shared_parameters': {'enabled': False, 'quantization_type': 'symmetric', 'range_calibration': 'dynamic', 'schedule_offset': 1000}, 'different_groups': {}}, 'sparse_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'row_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'head_pruning': {'shared_parameters': {'enabled': False, 'method': 'topk', 'schedule_offset': 1000}, 'different_groups': {}}, 'channel_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'layer_reduction': {'enabled': False}}
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] curriculum_enabled_legacy .... False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] curriculum_params_legacy ..... False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] data_efficiency_config ....... {'enabled': False, 'seed': 1234, 'data_sampling': {'enabled': False, 'num_epochs': 1000, 'num_workers': 0, 'curriculum_learning': {'enabled': False}}, 'data_routing': {'enabled': False, 'random_ltd': {'enabled': False, 'layer_token_lr_schedule': {'enabled': False}}}}
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] data_efficiency_enabled ...... False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] dataloader_drop_last ......... False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] disable_allgather ............ False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] dump_state ................... False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] dynamic_loss_scale_args ...... None
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] eigenvalue_enabled ........... False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] eigenvalue_gas_boundary_resolution 1
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] eigenvalue_layer_name ........ bert.encoder.layer
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] eigenvalue_layer_num ......... 0
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] eigenvalue_max_iter .......... 100
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] eigenvalue_stability ......... 1e-06
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] eigenvalue_tol ............... 0.01
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] eigenvalue_verbose ........... False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] elasticity_enabled ........... False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] flops_profiler_config ........ {
"enabled": false,
"recompute_fwd_factor": 0.0,
"profile_step": 1,
"module_depth": -1,
"top_modules": 1,
"detailed": true,
"output_file": null
}
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] fp16_auto_cast ............... None
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] fp16_enabled ................. False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] fp16_master_weights_and_gradients False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] global_rank .................. 0
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] grad_accum_dtype ............. None
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] gradient_accumulation_steps .. 16
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] gradient_clipping ............ 1.0
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] gradient_predivide_factor .... 1.0
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] hybrid_engine ................ enabled=False max_out_tokens=512 inference_tp_size=1 release_inference_cache=False pin_parameters=True tp_gather_partition_size=8
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] initial_dynamic_scale ........ 1
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] load_universal_checkpoint .... False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] loss_scale ................... 1.0
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] memory_breakdown ............. False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] mics_hierarchial_params_gather False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] mics_shard_size .............. -1
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] monitor_config ............... tensorboard=TensorBoardConfig(enabled=False, output_path='', job_name='DeepSpeedJobName') wandb=WandbConfig(enabled=False, group=None, team=None, project='deepspeed') csv_monitor=CSVConfig(enabled=False, output_path='', job_name='DeepSpeedJobName') enabled=False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] nebula_config ................ {
"enabled": false,
"persistent_storage_path": null,
"persistent_time_interval": 100,
"num_of_version_in_retention": 2,
"enable_nebula_load": true,
"load_path": null
}
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] optimizer_legacy_fusion ...... False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] optimizer_name ............... None
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] optimizer_params ............. None
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] pipeline ..................... {'stages': 'auto', 'partition': 'best', 'seed_layers': False, 'activation_checkpoint_interval': 0}
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] pld_enabled .................. False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] pld_params ................... False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] prescale_gradients ........... False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] scheduler_name ............... None
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] scheduler_params ............. None
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] sparse_attention ............. None
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] sparse_gradients_enabled ..... False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] steps_per_print .............. inf
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] train_batch_size ............. 32
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] train_micro_batch_size_per_gpu 1
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] use_node_local_storage ....... False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] wall_clock_breakdown ......... False
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] world_size ................... 2
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] zero_allow_untested_optimizer True
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] zero_config .................. stage=2 contiguous_gradients=True reduce_scatter=True reduce_bucket_size=500,000,000 allgather_partitions=True allgather_bucket_size=500,000,000 overlap_comm=False load_from_fp32_weights=True elastic_checkpoint=False offload_param=DeepSpeedZeroOffloadParamConfig(device='none', nvme_path=None, buffer_count=5, buffer_size=100,000,000, max_in_cpu=1,000,000,000, pin_memory=False) offload_optimizer=DeepSpeedZeroOffloadOptimizerConfig(device='none', nvme_path=None, buffer_count=4, pin_memory=False, pipeline=False, pipeline_read=False, pipeline_write=False, fast_init=False) sub_group_size=1,000,000,000 cpu_offload_param=None cpu_offload_use_pin_memory=None cpu_offload=None prefetch_bucket_size=50,000,000 param_persistence_threshold=100,000 model_persistence_threshold=sys.maxsize max_live_parameters=1,000,000,000 max_reuse_distance=1,000,000,000 gather_16bit_weights_on_model_save=False stage3_gather_fp16_weights_on_model_save=False ignore_unused_parameters=True legacy_stage1=False round_robin_gradients=False mics_shard_size=-1 mics_hierarchical_params_gather=False memory_efficient_linear=True
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] zero_enabled ................. True
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] zero_force_ds_cpu_optimizer .. True
[2023-07-09 14:35:30,321] [INFO] [config.py:964:print] zero_optimization_stage ...... 2
[2023-07-09 14:35:30,321] [INFO] [config.py:950:print_user_config] json = {
"train_batch_size": 32,
"train_micro_batch_size_per_gpu": 1,
"gradient_accumulation_steps": 16,
"zero_optimization": {
"stage": 2,
"offload_optimizer": {
"device": "none"
},
"offload_param": {
"device": "none"
},
"stage3_gather_16bit_weights_on_model_save": false
},
"gradient_clipping": 1.0,
"steps_per_print": inf,
"bf16": {
"enabled": true
},
"fp16": {
"enabled": false
},
"zero_allow_untested_optimizer": true
}
running training / 学習開始
num examples / サンプル数: 4650083
num batches per epoch / 1epochのバッチ数: 290639
num epochs / epoch数: 1
batch size per device / バッチサイズ: 1
total train batch size (with parallel & distributed & accumulation) / 総バッチサイズ（並列学習、勾配合計含む）: 32
gradient accumulation steps / 勾配を合計するステップ数 = 16
total optimization steps / 学習ステップ数: 18165
steps: 0%| | 0/18165 [00:00
train(args)
File "/mnt/my_raid/github/sd-scripts/sdxl_train.py", line 477, in train
sdxl_train_util.save_sd_model_on_epoch_end_or_stepwise(
File "/mnt/my_raid/github/sd-scripts/library/sdxl_train_util.py", line 306, in save_sd_model_on_epoch_end_or_stepwise
train_util.save_sd_model_on_epoch_end_or_stepwise_common(
File "/mnt/my_raid/github/sd-scripts/library/train_util.py", line 3561, in save_sd_model_on_epoch_end_or_stepwise_common
sd_saver(ckpt_file, epoch_no, global_step)
File "/mnt/my_raid/github/sd-scripts/library/sdxl_train_util.py", line 290, in sd_saver
sdxl_model_util.save_stable_diffusion_checkpoint(
File "/mnt/my_raid/github/sd-scripts/library/sdxl_model_util.py", line 312, in save_stable_diffusion_checkpoint
save_file(state_dict, output_file)
File "/mnt/my_raid/github/sd-scripts/venv/lib/python3.9/site-packages/safetensors/torch.py", line 232, in save_file
serialize_file(_flatten(tensors), filename, metadata=metadata)
File "/mnt/my_raid/github/sd-scripts/venv/lib/python3.9/site-packages/safetensors/torch.py", line 394, in _flatten
raise RuntimeError(
RuntimeError:
Some tensors share memory, this will lead to duplicate memory on disk and potential differences when loading them again: [{
```

```
A potential way to correctly save your model is to use `save_model`.
More information at https://huggingface.co/docs/safetensors/torch_shared_tensors

steps: 1%| | 100/18165 [07:46<23:24:42, 4.67s/it, loss=0.112]
WARNING:torch.distributed.elastic.multiprocessing.api:Sending process 22397 closing signal SIGTERM
ERROR:torch.distributed.elastic.multiprocessing.api:failed (exitcode: 1) local_rank: 0 (pid: 22396) of binary: /mnt/my_raid/github/sd-scripts/venv/bin/python
Traceback (most recent call last):
File "/mnt/my_raid/github/sd-scripts/venv/bin/accelerate", line 8, in
sys.exit(main())
File "/mnt/my_raid/github/sd-scripts/venv/lib/python3.9/site-packages/accelerate/commands/accelerate_cli.py", line 45, in main
args.func(args)
File "/mnt/my_raid/github/sd-scripts/venv/lib/python3.9/site-packages/accelerate/commands/launch.py", line 903, in launch_command
deepspeed_launcher(args)
File "/mnt/my_raid/github/sd-scripts/venv/lib/python3.9/site-packages/accelerate/commands/launch.py", line 648, in deepspeed_launcher
distrib_run.run(args)
File "/mnt/my_raid/github/sd-scripts/venv/lib/python3.9/site-packages/torch/distributed/run.py", line 785, in run
elastic_launch(
File "/mnt/my_raid/github/sd-scripts/venv/lib/python3.9/site-packages/torch/distributed/launcher/api.py", line 134, in __call__
return launch_agent(self._config, self._entrypoint, list(args))
File "/mnt/my_raid/github/sd-scripts/venv/lib/python3.9/site-packages/torch/distributed/launcher/api.py", line 250, in launch_agent
raise ChildFailedError(
torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
============================================================
sdxl_train.py FAILED
------------------------------------------------------------
Failures:

------------------------------------------------------------
Root Cause (first observed failure):
[0]:
time : 2023-07-09_14:43:24
host : balthasar
rank : 0 (local_rank: 0)
exitcode : 1 (pid: 22396)
error_file:
traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
============================================================
```

Best

The report identifies a save_file failure while using DeepSpeed, but names no source file or test. First locate the save_file call and reproduce the failure with the shown DeepSpeed setup; done means saving completes successfully and the resulting checkpoint can be used.

Written by the indexing model from the issue text.