chatglm3-6b 用DeepSpeed微调报错exits with return code = -7，但用单卡微调没问题。参考了issue#1683的方法也没用

[wumin86]

Reminder

• I have read the README and searched the existing issues.

Reproduction

训练命令如下，其中参考了issue#1683的方法加上export NCCL_P2P_LEVEL=NVL也没用
#!/bin/bash
module load anaconda cudnn/8.6.0.163_cuda11.x compilers/cuda/11.8 compilers/gcc/11.3.0
source activate myLLM
export PYTHONUNBUFFERED=1
export NCCL_P2P_LEVEL=NVL
deepspeed --num_gpus 4 --master_port=9901 src/train_bash.py
--deepspeed ds_config.json
--stage sft
--model_name_or_path ChatGLM3-6B
--do_train
--dataset allSemiDataNewMerge
--template chatglm3
--finetuning_type lora
--lora_rank 32
--lora_target all
--cutoff_len 1024
--output_dir outputs_ChatGLM3_lora_nopt_sft
--overwrite_cache
--per_device_train_batch_size 1
--gradient_accumulation_steps 1
--lr_scheduler_type cosine
--logging_steps 10
--save_steps 50
--save_total_limit 3
--learning_rate 5e-5
--num_train_epochs 1
--plot_loss
--seed 42
--bf16
--ddp_timeout 30000
--report_to tensorboard
--ddp_find_unused_parameters False
--gradient_checkpointing True \

ds_config.json 文件如下

{
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"zero_allow_untested_optimizer": true,
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"initial_scale_power": 16,
"loss_scale_window": 1000,
"hysteresis": 2,
"min_loss_scale": 1
},
"zero_optimization": {
"stage": 2,
"allgather_partitions": true,
"allgather_bucket_size": 5e8,
"reduce_scatter": true,
"reduce_bucket_size": 5e8,
"overlap_comm": false,
"contiguous_gradients": true
}
}

日志如下：

Running tokenizer on dataset: 100%|██████████| 68658/68658 [04:18<00:00, 203.98 examples/s]
Running tokenizer on dataset: 100%|██████████| 68658/68658 [04:19<00:00, 201.93 examples/s]
Running tokenizer on dataset: 100%|██████████| 68658/68658 [04:19<00:00, 264.99 examples/s]
/home/bingxing2/home/scx9042/.conda/envs/myLLM/lib/python3.10/site-packages/transformers/training_args.py:1751: FutureWarning: --push_to_hub_token is deprecated and will be removed in version 5 of 🤗 Transformers. Use --hub_token instead.
warnings.warn(

Running tokenizer on dataset: 100%|██████████| 68658/68658 [04:19<00:00, 264.85 examples/s]
/home/bingxing2/home/scx9042/.conda/envs/myLLM/lib/python3.10/site-packages/transformers/training_args.py:1751: FutureWarning: --push_to_hub_token is deprecated and will be removed in version 5 of 🤗 Transformers. Use --hub_token instead.
warnings.warn(
[WARNING|modeling_utils.py:2045] 2024-01-03 14:05:27,621 >> You are using an old version of the checkpointing format that is deprecated (We will also silently ignore gradient_checkpointing_kwargs in case you passed it).Please update to the new format on your modeling file. To use the new format, you need to completely remove the definition of the method _set_gradient_checkpointing in your model.
[WARNING|modeling_utils.py:2045] 2024-01-03 14:05:27,870 >> You are using an old version of the checkpointing format that is deprecated (We will also silently ignore gradient_checkpointing_kwargs in case you passed it).Please update to the new format on your modeling file. To use the new format, you need to completely remove the definition of the method _set_gradient_checkpointing in your model.
[2024-01-03 14:05:33,233] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed Flops Profiler Enabled: False
[2024-01-03 14:05:33,239] [INFO] [logging.py:96:log_dist] [Rank 0] Using client Optimizer as basic optimizer
[2024-01-03 14:05:33,239] [INFO] [logging.py:96:log_dist] [Rank 0] Removing param_group that has no 'params' in the basic Optimizer
[2024-01-03 14:05:33,263] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed Basic Optimizer = AdamW
[2024-01-03 14:05:33,263] [INFO] [utils.py:56:is_zero_supported_optimizer] Checking ZeRO support for optimizer=AdamW type=<class 'torch.optim.adamw.AdamW'>
[2024-01-03 14:05:33,264] [INFO] [logging.py:96:log_dist] [Rank 0] Creating torch.bfloat16 ZeRO stage 2 optimizer
[2024-01-03 14:05:33,264] [INFO] [stage_1_and_2.py:148:init] Reduce bucket size 500000000
[2024-01-03 14:05:33,264] [INFO] [stage_1_and_2.py:149:init] Allgather bucket size 500000000
[2024-01-03 14:05:33,264] [INFO] [stage_1_and_2.py:150:init] CPU Offload: False
[2024-01-03 14:05:33,264] [INFO] [stage_1_and_2.py:151:init] Round robin gradient partitioning: False
[2024-01-03 14:05:35,314] [INFO] [utils.py:791:see_memory_usage] Before initializing optimizer states
[2024-01-03 14:05:35,315] [INFO] [utils.py:792:see_memory_usage] MA 11.82 GB Max_MA 11.85 GB CA 11.87 GB Max_CA 12 GB
[2024-01-03 14:05:35,315] [INFO] [utils.py:799:see_memory_usage] CPU Virtual Memory: used = 76.4 GB, percent = 30.1%
[2024-01-03 14:05:35,519] [INFO] [utils.py:791:see_memory_usage] After initializing optimizer states
[2024-01-03 14:05:35,520] [INFO] [utils.py:792:see_memory_usage] MA 11.93 GB Max_MA 12.1 GB CA 12.15 GB Max_CA 12 GB
[2024-01-03 14:05:35,521] [INFO] [utils.py:799:see_memory_usage] CPU Virtual Memory: used = 76.52 GB, percent = 30.1%
[2024-01-03 14:05:35,521] [INFO] [stage_1_and_2.py:516:init] optimizer state initialized
[2024-01-03 14:05:35,706] [INFO] [utils.py:791:see_memory_usage] After initializing ZeRO optimizer
[2024-01-03 14:05:35,707] [INFO] [utils.py:792:see_memory_usage] MA 11.93 GB Max_MA 11.93 GB CA 12.15 GB Max_CA 12 GB
[2024-01-03 14:05:35,707] [INFO] [utils.py:799:see_memory_usage] CPU Virtual Memory: used = 76.51 GB, percent = 30.1%
[2024-01-03 14:05:35,710] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed Final Optimizer = AdamW
[2024-01-03 14:05:35,711] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed using client LR scheduler
[2024-01-03 14:05:35,711] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed LR Scheduler = None
[2024-01-03 14:05:35,711] [INFO] [logging.py:96:log_dist] [Rank 0] step=0, skipped=0, lr=[5e-05], mom=[(0.9, 0.999)]
[2024-01-03 14:05:35,714] [INFO] [config.py:984:print] DeepSpeedEngine configuration:
[2024-01-03 14:05:35,715] [INFO] [config.py:988:print] activation_checkpointing_config {
"partition_activations": false,
"contiguous_memory_optimization": false,
"cpu_checkpointing": false,
"number_checkpoints": null,
"synchronize_checkpoint_boundary": false,
"profile": false
}
[2024-01-03 14:05:35,715] [INFO] [config.py:988:print] aio_config ................... {'block_size': 1048576, 'queue_depth': 8, 'thread_count': 1, 'single_submit': False, 'overlap_events': True}
[2024-01-03 14:05:35,715] [INFO] [config.py:988:print] amp_enabled .................. False
[2024-01-03 14:05:35,715] [INFO] [config.py:988:print] amp_params ................... False
[2024-01-03 14:05:35,715] [INFO] [config.py:988: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
}
[2024-01-03 14:05:35,715] [INFO] [config.py:988:print] bfloat16_enabled ............. True
[2024-01-03 14:05:35,716] [INFO] [config.py:988:print] checkpoint_parallel_write_pipeline False
[2024-01-03 14:05:35,716] [INFO] [config.py:988:print] checkpoint_tag_validation_enabled True
[2024-01-03 14:05:35,716] [INFO] [config.py:988:print] checkpoint_tag_validation_fail False
[2024-01-03 14:05:35,716] [INFO] [config.py:988:print] comms_config ................. <deepspeed.comm.config.DeepSpeedCommsConfig object at 0x40047d4088e0>
[2024-01-03 14:05:35,716] [INFO] [config.py:988:print] communication_data_type ...... None
[2024-01-03 14:05:35,716] [INFO] [config.py:988: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}}
[2024-01-03 14:05:35,716] [INFO] [config.py:988:print] curriculum_enabled_legacy .... False
[2024-01-03 14:05:35,716] [INFO] [config.py:988:print] curriculum_params_legacy ..... False
[2024-01-03 14:05:35,716] [INFO] [config.py:988: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}}}}
[2024-01-03 14:05:35,716] [INFO] [config.py:988:print] data_efficiency_enabled ...... False
[2024-01-03 14:05:35,716] [INFO] [config.py:988:print] dataloader_drop_last ......... False
[2024-01-03 14:05:35,716] [INFO] [config.py:988:print] disable_allgather ............ False
[2024-01-03 14:05:35,716] [INFO] [config.py:988:print] dump_state ................... False
[2024-01-03 14:05:35,716] [INFO] [config.py:988:print] dynamic_loss_scale_args ...... None
[2024-01-03 14:05:35,716] [INFO] [config.py:988:print] eigenvalue_enabled ........... False
[2024-01-03 14:05:35,716] [INFO] [config.py:988:print] eigenvalue_gas_boundary_resolution 1
[2024-01-03 14:05:35,716] [INFO] [config.py:988:print] eigenvalue_layer_name ........ bert.encoder.layer
[2024-01-03 14:05:35,717] [INFO] [config.py:988:print] eigenvalue_layer_num ......... 0
[2024-01-03 14:05:35,717] [INFO] [config.py:988:print] eigenvalue_max_iter .......... 100
[2024-01-03 14:05:35,717] [INFO] [config.py:988:print] eigenvalue_stability ......... 1e-06
[2024-01-03 14:05:35,717] [INFO] [config.py:988:print] eigenvalue_tol ............... 0.01
[2024-01-03 14:05:35,717] [INFO] [config.py:988:print] eigenvalue_verbose ........... False
[2024-01-03 14:05:35,717] [INFO] [config.py:988:print] elasticity_enabled ........... False
[2024-01-03 14:05:35,717] [INFO] [config.py:988: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
}
[2024-01-03 14:05:35,717] [INFO] [config.py:988:print] fp16_auto_cast ............... None
[2024-01-03 14:05:35,717] [INFO] [config.py:988:print] fp16_enabled ................. False
[2024-01-03 14:05:35,717] [INFO] [config.py:988:print] fp16_master_weights_and_gradients False
[2024-01-03 14:05:35,717] [INFO] [config.py:988:print] global_rank .................. 0
[2024-01-03 14:05:35,717] [INFO] [config.py:988:print] grad_accum_dtype ............. None
[2024-01-03 14:05:35,717] [INFO] [config.py:988:print] gradient_accumulation_steps .. 1
[2024-01-03 14:05:35,717] [INFO] [config.py:988:print] gradient_clipping ............ 1.0
[2024-01-03 14:05:35,717] [INFO] [config.py:988:print] gradient_predivide_factor .... 1.0
[2024-01-03 14:05:35,718] [INFO] [config.py:988:print] graph_harvesting ............. False
[2024-01-03 14:05:35,718] [INFO] [config.py:988: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
[2024-01-03 14:05:35,718] [INFO] [config.py:988:print] initial_dynamic_scale ........ 1
[2024-01-03 14:05:35,718] [INFO] [config.py:988:print] load_universal_checkpoint .... False
[2024-01-03 14:05:35,718] [INFO] [config.py:988:print] loss_scale ................... 1.0
[2024-01-03 14:05:35,718] [INFO] [config.py:988:print] memory_breakdown ............. False
[2024-01-03 14:05:35,718] [INFO] [config.py:988:print] mics_hierarchial_params_gather False
[2024-01-03 14:05:35,718] [INFO] [config.py:988:print] mics_shard_size .............. -1
[2024-01-03 14:05:35,718] [INFO] [config.py:988: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
[2024-01-03 14:05:35,718] [INFO] [config.py:988: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
}
[2024-01-03 14:05:35,718] [INFO] [config.py:988:print] optimizer_legacy_fusion ...... False
[2024-01-03 14:05:35,718] [INFO] [config.py:988:print] optimizer_name ............... None
[2024-01-03 14:05:35,718] [INFO] [config.py:988:print] optimizer_params ............. None
[2024-01-03 14:05:35,718] [INFO] [config.py:988:print] pipeline ..................... {'stages': 'auto', 'partition': 'best', 'seed_layers': False, 'activation_checkpoint_interval': 0, 'pipe_partitioned': True, 'grad_partitioned': True}
[2024-01-03 14:05:35,718] [INFO] [config.py:988:print] pld_enabled .................. False
[2024-01-03 14:05:35,718] [INFO] [config.py:988:print] pld_params ................... False
[2024-01-03 14:05:35,718] [INFO] [config.py:988:print] prescale_gradients ........... False
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] scheduler_name ............... None
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] scheduler_params ............. None
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] seq_parallel_communication_data_type torch.float32
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] sparse_attention ............. None
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] sparse_gradients_enabled ..... False
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] steps_per_print .............. inf
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] train_batch_size ............. 4
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] train_micro_batch_size_per_gpu 1
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] use_data_before_expert_parallel_ False
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] use_node_local_storage ....... False
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] wall_clock_breakdown ......... False
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] weight_quantization_config ... None
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] world_size ................... 4
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] zero_allow_untested_optimizer True
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] zero_config .................. stage=2 contiguous_gradients=True reduce_scatter=True reduce_bucket_size=500000000 use_multi_rank_bucket_allreduce=True allgather_partitions=True allgather_bucket_size=500000000 overlap_comm=False load_from_fp32_weights=True elastic_checkpoint=False offload_param=None offload_optimizer=None 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 zero_hpz_partition_size=1 zero_quantized_weights=False zero_quantized_nontrainable_weights=False zero_quantized_gradients=False mics_shard_size=-1 mics_hierarchical_params_gather=False memory_efficient_linear=True pipeline_loading_checkpoint=False override_module_apply=True
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] zero_enabled ................. True
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] zero_force_ds_cpu_optimizer .. True
[2024-01-03 14:05:35,719] [INFO] [config.py:988:print] zero_optimization_stage ...... 2
[2024-01-03 14:05:35,720] [INFO] [config.py:974:print_user_config] json = {
"train_batch_size": 4,
"train_micro_batch_size_per_gpu": 1,
"gradient_accumulation_steps": 1,
"gradient_clipping": 1.0,
"zero_allow_untested_optimizer": true,
"fp16": {
"enabled": false,
"loss_scale": 0,
"initial_scale_power": 16,
"loss_scale_window": 1000,
"hysteresis": 2,
"min_loss_scale": 1
},
"zero_optimization": {
"stage": 2,
"allgather_partitions": true,
"allgather_bucket_size": 5.000000e+08,
"reduce_scatter": true,
"reduce_bucket_size": 5.000000e+08,
"overlap_comm": false,
"contiguous_gradients": true
},
"steps_per_print": inf,
"bf16": {
"enabled": true
}
}
[INFO|trainer.py:1706] 2024-01-03 14:05:35,720 >> ***** Running training *****
[INFO|trainer.py:1707] 2024-01-03 14:05:35,720 >> Num examples = 68,658
[INFO|trainer.py:1708] 2024-01-03 14:05:35,720 >> Num Epochs = 1
[INFO|trainer.py:1709] 2024-01-03 14:05:35,720 >> Instantaneous batch size per device = 1
[INFO|trainer.py:1712] 2024-01-03 14:05:35,720 >> Total train batch size (w. parallel, distributed & accumulation) = 4
[INFO|trainer.py:1713] 2024-01-03 14:05:35,720 >> Gradient Accumulation steps = 1
[INFO|trainer.py:1714] 2024-01-03 14:05:35,720 >> Total optimization steps = 17,165
[INFO|trainer.py:1715] 2024-01-03 14:05:35,724 >> Number of trainable parameters = 59,293,696

0%| | 0/17165 [00:00<?, ?it/s][2024-01-03 14:05:36,090] [INFO] [launch.py:315:sigkill_handler] Killing subprocess 2726715
[2024-01-03 14:05:36,090] [INFO] [launch.py:315:sigkill_handler] Killing subprocess 2726716
[2024-01-03 14:05:36,302] [INFO] [launch.py:315:sigkill_handler] Killing subprocess 2726717
[2024-01-03 14:05:36,598] [INFO] [launch.py:315:sigkill_handler] Killing subprocess 2726718
[2024-01-03 14:05:36,716] [ERROR] [launch.py:321:sigkill_handler] ['/home/bingxing2/home/scx9042/.conda/envs/myLLM/bin/python', '-u', 'src/train_bash.py', '--local_rank=3', '--deepspeed', 'ds_config.json', '--stage', 'sft', '--model_name_or_path', 'ChatGLM3-6B', '--do_train', '--dataset', 'allSemiDataNewMerge', '--template', 'chatglm3', '--finetuning_type', 'lora', '--lora_rank', '32', '--lora_target', 'all', '--cutoff_len', '1024', '--output_dir', 'outputs_ChatGLM3_lora_nopt_sft_haha', '--overwrite_cache', '--per_device_train_batch_size', '1', '--gradient_accumulation_steps', '1', '--lr_scheduler_type', 'cosine', '--logging_steps', '10', '--save_steps', '50', '--save_total_limit', '3', '--learning_rate', '5e-5', '--num_train_epochs', '1', '--plot_loss', '--seed', '42', '--bf16', '--ddp_timeout', '30000', '--report_to', 'tensorboard', '--ddp_find_unused_parameters', 'False', '--gradient_checkpointing', 'True'] exits with return code = -7

Expected behavior

No response

System Info

No response

Others

No response

[hiyouga]

deepspeedai/DeepSpeed#4002

[wumin86]

microsoft/DeepSpeed#4002

这些解决方法是增加docker共享内存，但我在超算云上跑的，没用docker，怎么解决？

[hiyouga]

多申请一些内存

[luolanfeixue]

遇到相同的问题，docker共享内存还有600G+，cpu内存920G，cpu内存利用率也没出现峰值，感觉不是docker共享内存和cpu内存的问题。要怎么解决呢？