与 DeepSpeed 相同的训练时间（尽管批量大小增加）

Question

问题：

嗨！我正在使用 huggingface（不带 Trainer API）和使用 deepspeed 来微调 Electra 模型。应用 Deepspeed 后，我可以增加训练模型的批量大小（64 -> 128，但 OOM 为 256），因此我预计训练时间会减少。 但是，即使我在代码中应用了 deepspeed，训练时间也是相同的。我不确定 Deepspeed 是否按照我的配置工作，如果没有，我想知道如何才能使其正常工作（更快的训练）

我正在 ColabProPlus 中训练模型，规格就像这个
（我每次都分配到相同的GPU - Tesla P100-PCIE-16GB GPU）
（我没有包括处理器 1~7 的规格，因为它们与处理器 0 相同）

"""
from psutil import *
print(cpu_count())
print(cpu_stats())
!cat /proc/cpuinfo
!df -h
"""

outputs:

8
scpustats(ctx_switches=2878445, interrupts=1351912, soft_interrupts=1239827, syscalls=0)
processor   : 0
vendor_id   : GenuineIntel
cpu family  : 6
model       : 63
model name  : Intel(R) Xeon(R) CPU @ 2.30GHz
stepping    : 0
microcode   : 0x1
cpu MHz     : 2299.998
cache size  : 46080 KB
physical id : 0
siblings    : 8
core id     : 0
cpu cores   : 4
apicid      : 0
initial apicid  : 0
fpu     : yes
fpu_exception   : yes
cpuid level : 13
wp      : yes
flags       : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc cpuid tsc_known_freq pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm abm invpcid_single ssbd ibrs ibpb stibp fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid xsaveopt arat md_clear arch_capabilities
bugs        : cpu_meltdown spectre_v1 spectre_v2 spec_store_bypass l1tf mds swapgs
bogomips    : 4599.99
clflush size    : 64
cache_alignment : 64
address sizes   : 46 bits physical, 48 bits virtual
power management:


Filesystem      Size  Used Avail Use% Mounted on
overlay         167G   43G  125G  26% /
tmpfs            64M     0   64M   0% /dev
shm              25G     0   25G   0% /dev/shm
/dev/root       2.0G  1.2G  817M  59% /sbin/docker-init
tmpfs            26G   28K   26G   1% /var/colab
/dev/sda1       174G   46G  128G  27% /opt/bin/.nvidia
tmpfs            26G     0   26G   0% /proc/acpi
tmpfs            26G     0   26G   0% /proc/scsi
tmpfs            26G     0   26G   0% /sys/firmware
drive           167G   49G  119G  30% /content/drive

我在初始化运行时时运行此命令以使用 deepspeed

!apt-get install libaio-dev
!pip install deepspeed triton==1.0.0

，然后，我的 ds_report 在运行训练代码之前显示这一点。

--------------------------------------------------
DeepSpeed C++/CUDA extension op report
--------------------------------------------------
NOTE: Ops not installed will be just-in-time (JIT) compiled at
      runtime if needed. Op compatibility means that your system
      meet the required dependencies to JIT install the op.
--------------------------------------------------
JIT compiled ops requires ninja
ninja .................. [OKAY]
--------------------------------------------------
op name ................ installed .. compatible
--------------------------------------------------
cpu_adam ............... [NO] ....... [OKAY]
cpu_adagrad ............ [NO] ....... [OKAY]
fused_adam ............. [NO] ....... [OKAY]
fused_lamb ............. [NO] ....... [OKAY]
sparse_attn ............ [NO] ....... [OKAY]
transformer ............ [NO] ....... [OKAY]
stochastic_transformer . [NO] ....... [OKAY]
async_io ............... [NO] ....... [OKAY]
utils .................. [NO] ....... [OKAY]
quantizer .............. [NO] ....... [OKAY]
transformer_inference .. [NO] ....... [OKAY]
--------------------------------------------------
DeepSpeed general environment info:
torch install path ............... ['/usr/local/lib/python3.7/dist-packages/torch']
torch version .................... 1.10.0+cu111
torch cuda version ............... 11.1
torch hip version ................ None
nvcc version ..................... 11.1
deepspeed install path ........... ['/usr/local/lib/python3.7/dist-packages/deepspeed']
deepspeed info ................... 0.6.0, unknown, unknown
deepspeed wheel compiled w. ...... torch 1.10, cuda 11.1, hip 0.0

我的deepspeed配置是这样的。

"ds_config":{
        "stage2": {
            "train_micro_batch_size_per_gpu": 128,
            "fp16": {
                "enabled": true,
                "loss_scale": 0,
                "loss_scale_window": 1000,
                "initial_scale_power": 16,
                "hysteresis": 2,
                "min_loss_scale": 1
                    },

            "optimizer": {
                "type": "AdamW",
                "params": {
                    "lr": 5e-05,
                    "betas": [
                        0.9,
                        0.999
                        ],
                    "eps": 1e-8
                        }
                      },

            "scheduler": {
                "type": "WarmupLR",
                "params": {
                    "warmup_min_lr": 0,
                    "warmup_max_lr": 5e-5,
                    "warmup_num_steps": 0
                        }
                        },

            "zero_optimization": {
                "stage": 2,
                "allgather_partitions": true,
                "allgather_bucket_size": 5e8,
                "overlap_comm": true,
                "reduce_scatter": true,
                "reduce_bucket_size": 5e8,
                "contiguous_gradients": true,
                "cpu_offload": true,
                "offload_optimizer": {
                    "device": "cpu",
                    "pin_memory": true,
                    "fast_init": true
                                }
                          },

            "gradient_accumulation_steps": 1,
            "gradient_clipping": 1.0
        }
}

我的火车代码与评论中链接的代码几乎相同，只添加了几行来使用 deepspeed，如下所示。

# https://github.com/monologg/KoELECTRA/blob/master/finetune/run_seq_cls.py

...
import deepspeed
...

def train():
    ...
    ...

    model, optimizer, _, _ = deepspeed.initialize(model=model,
                              config_params=args['ds_config']['stage2'],
                              model_parameters=model.parameters())
    ...
    ...
    for epoch in mb:
        epoch_iterator = progress_bar(train_dataloader, parent=mb)
        for step, batch in enumerate(epoch_iterator):
            model.train()
            batch = tuple(t.to(args.device) for t in batch)
            inputs = {
                "input_ids": batch[0],
                "attention_mask": batch[1],
                "labels": batch[3]
            }
            
            outputs = model(**inputs)
            loss = outputs.loss
            model.backward(loss)

            tr_loss += loss.item()
            if (step + 1) % args.gradient_accumulation_steps == 0 or (
                    len(train_dataloader) <= args.gradient_accumulation_steps
                    and (step + 1) == len(train_dataloader)
            ):
                #torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
                #optimizer.step()
                model.step()
                ...

def main():
    ....
    deepspeed.init_distributed("nccl")
    ....

这是我运行训练时的输出日志。

[2022-03-09 04:22:56,048] [WARNING] [runner.py:155:fetch_hostfile] Unable to find hostfile, will proceed with training with local resources only.
[2022-03-09 04:22:56,055] [INFO] [runner.py:438:main] cmd = /usr/bin/python3 -u -m deepspeed.launcher.launch --world_info=eyJsb2NhbGhvc3QiOiBbMF19 --master_addr=127.0.0.1 --master_port=29500 run_seq_cls2.py --dataset pm --mode all --config_dir ./electra/config
[2022-03-09 04:22:57,946] [INFO] [launch.py:96:main] 0 NCCL_VERSION=2.7.8
[2022-03-09 04:22:57,946] [INFO] [launch.py:103:main] WORLD INFO DICT: {'localhost': [0]}
[2022-03-09 04:22:57,946] [INFO] [launch.py:110:main] nnodes=1, num_local_procs=1, node_rank=0
[2022-03-09 04:22:57,946] [INFO] [launch.py:122:main] global_rank_mapping=defaultdict(<class 'list'>, {'localhost': [0]})
[2022-03-09 04:22:57,946] [INFO] [launch.py:123:main] dist_world_size=1
[2022-03-09 04:22:57,946] [INFO] [launch.py:125:main] Setting CUDA_VISIBLE_DEVICES=0
[2022-03-09 04:23:03,174] [INFO] [distributed.py:47:init_distributed] Initializing torch distributed with backend: nccl
03/09/2022 04:23:03 - INFO - torch.distributed.distributed_c10d -   Added key: store_based_barrier_key:1 to store for rank: 0
03/09/2022 04:23:03 - INFO - torch.distributed.distributed_c10d -   Rank 0: Completed store-based barrier for key:store_based_barrier_key:1 with 1 nodes.
03/09/2022 04:23:03 - INFO - __main__ -   Training/evaluation parameters AttrDict({'run_n_fold': 1, 'task': 'spend', 'data_dir': './electra/data', 'ckpt_dir': './electra/ckpt', 'evaluate_test_during_training': False, 'eval_all_checkpoints': False, 'save_optimizer': True, 'do_lower_case': False, 'do_train': True, 'do_eval': True, 'max_seq_len': 128, 'num_train_epochs': 100, 'weight_decay': 0.0, 'gradient_accumulation_steps': 1, 'adam_epsilon': 1e-08, 'warmup_proportion': 0, 'max_steps': -1, 'max_grad_norm': 1.0, 'no_cuda': False, 'model_type': 'koelectra-base-v3', 'model_name_or_path': 'monologg/koelectra-base-v3-discriminator', 'output_dir': 'koelectra-base-v3-spend-ckpt', 'seed': 42, 'train_batch_size': 128, 'eval_batch_size': 128, 'logging_steps': 3000, 'save_steps': 3000, 'logging_epochs': 1, 'save_epochs': 1, 'learning_rate': 5e-05, 'deepspeed': True, 'ds_config': {'stage2': {'train_micro_batch_size_per_gpu': 128, 'fp16': {'enabled': True, 'loss_scale': 0, 'loss_scale_window': 1000, 'initial_scale_power': 16, 'hysteresis': 2, 'min_loss_scale': 1}, 'optimizer': {'type': 'AdamW', 'params': {'lr': 5e-05, 'betas': [0.9, 0.999], 'eps': 1e-08}}, 'scheduler': {'type': 'WarmupLR', 'params': {'warmup_min_lr': 0, 'warmup_max_lr': 5e-05, 'warmup_num_steps': 0}}, 'zero_optimization': {'stage': 2, 'allgather_partitions': True, 'allgather_bucket_size': 500000000.0, 'overlap_comm': True, 'reduce_scatter': True, 'reduce_bucket_size': 500000000.0, 'contiguous_gradients': True, 'cpu_offload': True, 'offload_optimizer': {'device': 'cpu', 'pin_memory': True, 'fast_init': True}}, 'gradient_accumulation_steps': 1, 'gradient_clipping': 1.0}, 'stage3': {'fp16': {'enabled': 'auto', 'loss_scale': 0, 'loss_scale_window': 1000, 'initial_scale_power': 16, 'hysteresis': 2, 'min_loss_scale': 1}, 'optimizer': {'type': 'AdamW', 'params': {'lr': 'auto', 'betas': 'auto', 'eps': 'auto', 'weight_decay': 'auto'}}, 'scheduler': {'type': 'WarmupLR', 'params': {'warmup_min_lr': 'auto', 'warmup_max_lr': 'auto', 'warmup_num_steps': 'auto'}}, 'zero_optimization': {'stage': 3, 'offload_optimizer': {'device': 'cpu', 'pin_memory': True}, 'offload_param': {'device': 'cpu', 'pin_memory': True}, 'overlap_comm': True, 'contiguous_gradients': True, 'sub_group_size': 100000000000000.0, 'reduce_bucket_size': 'auto', 'stage3_prefetch_bucket_size': 'auto', 'stage3_param_persistence_threshold': 'auto', 'stage3_max_live_parameters': 1000000000.0, 'stage3_max_reuse_distance': 1000000000.0, 'stage3_gather_fp16_weights_on_model_save': True}, 'gradient_accumulation_steps': 'auto', 'gradient_clipping': 'auto', 'steps_per_print': 2000, 'train_batch_size': 'auto', 'train_micro_batch_size_per_gpu': 'auto', 'wall_clock_breakdown': False}}})
03/09/2022 04:23:12 - INFO - processor.seq_cls -   Loading features from cached file ./electra/data/cached_spend_koelectra-base-v3-discriminator_128
[2022-03-09 04:23:55,158] [INFO] [logging.py:69:log_dist] [Rank 0] DeepSpeed info: version=0.6.0, git-hash=unknown, git-branch=unknown
03/09/2022 04:23:55 - INFO - torch.distributed.distributed_c10d -   Added key: store_based_barrier_key:2 to store for rank: 0
03/09/2022 04:23:55 - INFO - torch.distributed.distributed_c10d -   Rank 0: Completed store-based barrier for key:store_based_barrier_key:2 with 1 nodes.
[2022-03-09 04:23:55,225] [INFO] [engine.py:278:__init__] DeepSpeed Flops Profiler Enabled: False
Using /root/.cache/torch_extensions/py37_cu111 as PyTorch extensions root...
Detected CUDA files, patching ldflags
Emitting ninja build file /root/.cache/torch_extensions/py37_cu111/cpu_adam/build.ninja...
Building extension module cpu_adam...
Allowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N)
ninja: no work to do.
Loading extension module cpu_adam...
Time to load cpu_adam op: 3.290276527404785 seconds
Adam Optimizer #0 is created with AVX2 arithmetic capability.
Config: alpha=0.000050, betas=(0.900000, 0.999000), weight_decay=0.000000, adam_w=1
[2022-03-09 04:24:00,352] [INFO] [engine.py:1066:_configure_optimizer] Using DeepSpeed Optimizer param name adamw as basic optimizer
[2022-03-09 04:24:00,363] [INFO] [engine.py:1073:_configure_optimizer] DeepSpeed Basic Optimizer = DeepSpeedCPUAdam
[2022-03-09 04:24:00,363] [INFO] [utils.py:49:is_zero_supported_optimizer] Checking ZeRO support for optimizer=DeepSpeedCPUAdam type=<class 'deepspeed.ops.adam.cpu_adam.DeepSpeedCPUAdam'>
[2022-03-09 04:24:00,363] [INFO] [logging.py:69:log_dist] [Rank 0] Creating fp16 ZeRO stage 2 optimizer
[2022-03-09 04:24:00,364] [INFO] [stage_1_and_2.py:125:__init__] Reduce bucket size 500000000.0
[2022-03-09 04:24:00,364] [INFO] [stage_1_and_2.py:126:__init__] Allgather bucket size 500000000.0
[2022-03-09 04:24:00,364] [INFO] [stage_1_and_2.py:127:__init__] CPU Offload: True
[2022-03-09 04:24:00,364] [INFO] [stage_1_and_2.py:128:__init__] Round robin gradient partitioning: False
Using /root/.cache/torch_extensions/py37_cu111 as PyTorch extensions root...
Emitting ninja build file /root/.cache/torch_extensions/py37_cu111/utils/build.ninja...
Building extension module utils...
Allowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N)
ninja: no work to do.
Loading extension module utils...
Time to load utils op: 0.5493736267089844 seconds
Rank: 0 partition count [1] and sizes[(113066686, False)] 
[2022-03-09 04:24:01,501] [INFO] [utils.py:824:see_memory_usage] Before initializing optimizer states
[2022-03-09 04:24:01,502] [INFO] [utils.py:829:see_memory_usage] MA 0.26 GB         Max_MA 0.47 GB         CA 0.74 GB         Max_CA 1 GB 
[2022-03-09 04:24:01,502] [INFO] [utils.py:834:see_memory_usage] CPU Virtual Memory:  used = 7.89 GB, percent = 15.5%
[2022-03-09 04:24:02,174] [INFO] [utils.py:824:see_memory_usage] After initializing optimizer states
[2022-03-09 04:24:02,176] [INFO] [utils.py:829:see_memory_usage] MA 0.26 GB         Max_MA 0.26 GB         CA 0.74 GB         Max_CA 1 GB 
[2022-03-09 04:24:02,177] [INFO] [utils.py:834:see_memory_usage] CPU Virtual Memory:  used = 9.17 GB, percent = 18.0%
[2022-03-09 04:24:02,177] [INFO] [stage_1_and_2.py:497:__init__] optimizer state initialized
[2022-03-09 04:24:02,350] [INFO] [utils.py:824:see_memory_usage] After initializing ZeRO optimizer
[2022-03-09 04:24:02,351] [INFO] [utils.py:829:see_memory_usage] MA 0.26 GB         Max_MA 0.26 GB         CA 0.74 GB         Max_CA 1 GB 
[2022-03-09 04:24:02,351] [INFO] [utils.py:834:see_memory_usage] CPU Virtual Memory:  used = 9.17 GB, percent = 18.0%
[2022-03-09 04:24:02,352] [INFO] [logging.py:69:log_dist] [Rank 0] DeepSpeed Final Optimizer = adamw
[2022-03-09 04:24:02,352] [INFO] [engine.py:777:_configure_lr_scheduler] DeepSpeed using configured LR scheduler = WarmupLR
[2022-03-09 04:24:02,352] [INFO] [logging.py:69:log_dist] [Rank 0] DeepSpeed LR Scheduler = <deepspeed.runtime.lr_schedules.WarmupLR object at 0x7f507ff31f50>
[2022-03-09 04:24:02,352] [INFO] [logging.py:69:log_dist] [Rank 0] step=0, skipped=0, lr=[5e-05], mom=[[0.9, 0.999]]
[2022-03-09 04:24:02,353] [INFO] [config.py:1058:print] DeepSpeedEngine configuration:
[2022-03-09 04:24:02,353] [INFO] [config.py:1062:print]   activation_checkpointing_config  {
    "partition_activations": false, 
    "contiguous_memory_optimization": false, 
    "cpu_checkpointing": false, 
    "number_checkpoints": null, 
    "synchronize_checkpoint_boundary": false, 
    "profile": false
}
[2022-03-09 04:24:02,353] [INFO] [config.py:1062:print]   aio_config ................... {'block_size': 1048576, 'queue_depth': 8, 'thread_count': 1, 'single_submit': False, 'overlap_events': True}
[2022-03-09 04:24:02,353] [INFO] [config.py:1062:print]   amp_enabled .................. False
[2022-03-09 04:24:02,354] [INFO] [config.py:1062:print]   amp_params ................... False
[2022-03-09 04:24:02,354] [INFO] [config.py:1062:print]   autotuning_config ............ {
    "enabled": false, 
    "start_step": null, 
    "end_step": null, 
    "metric_path": null, 
    "arg_mappings": null, 
    "metric": "throughput", 
    "model_info": null, 
    "results_dir": null, 
    "exps_dir": null, 
    "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
}
[2022-03-09 04:24:02,354] [INFO] [config.py:1062:print]   bfloat16_enabled ............. False
[2022-03-09 04:24:02,354] [INFO] [config.py:1062:print]   checkpoint_tag_validation_enabled  True
[2022-03-09 04:24:02,354] [INFO] [config.py:1062:print]   checkpoint_tag_validation_fail  False
[2022-03-09 04:24:02,354] [INFO] [config.py:1062:print]   communication_data_type ...... None
[2022-03-09 04:24:02,354] [INFO] [config.py:1062:print]   curriculum_enabled ........... False
[2022-03-09 04:24:02,354] [INFO] [config.py:1062:print]   curriculum_params ............ False
[2022-03-09 04:24:02,354] [INFO] [config.py:1062:print]   dataloader_drop_last ......... False
[2022-03-09 04:24:02,354] [INFO] [config.py:1062:print]   disable_allgather ............ False
[2022-03-09 04:24:02,355] [INFO] [config.py:1062:print]   dump_state ................... False
[2022-03-09 04:24:02,355] [INFO] [config.py:1062:print]   dynamic_loss_scale_args ...... {'init_scale': 65536, 'scale_window': 1000, 'delayed_shift': 2, 'min_scale': 1}
[2022-03-09 04:24:02,355] [INFO] [config.py:1062:print]   eigenvalue_enabled ........... False
[2022-03-09 04:24:02,355] [INFO] [config.py:1062:print]   eigenvalue_gas_boundary_resolution  1
[2022-03-09 04:24:02,355] [INFO] [config.py:1062:print]   eigenvalue_layer_name ........ bert.encoder.layer
[2022-03-09 04:24:02,355] [INFO] [config.py:1062:print]   eigenvalue_layer_num ......... 0
[2022-03-09 04:24:02,355] [INFO] [config.py:1062:print]   eigenvalue_max_iter .......... 100
[2022-03-09 04:24:02,355] [INFO] [config.py:1062:print]   eigenvalue_stability ......... 1e-06
[2022-03-09 04:24:02,355] [INFO] [config.py:1062:print]   eigenvalue_tol ............... 0.01
[2022-03-09 04:24:02,355] [INFO] [config.py:1062:print]   eigenvalue_verbose ........... False
[2022-03-09 04:24:02,355] [INFO] [config.py:1062:print]   elasticity_enabled ........... False
[2022-03-09 04:24:02,355] [INFO] [config.py:1062:print]   flops_profiler_config ........ {
    "enabled": false, 
    "profile_step": 1, 
    "module_depth": -1, 
    "top_modules": 1, 
    "detailed": true, 
    "output_file": null
}
[2022-03-09 04:24:02,356] [INFO] [config.py:1062:print]   fp16_enabled ................. True
[2022-03-09 04:24:02,356] [INFO] [config.py:1062:print]   fp16_master_weights_and_gradients  False
[2022-03-09 04:24:02,356] [INFO] [config.py:1062:print]   fp16_mixed_quantize .......... False
[2022-03-09 04:24:02,356] [INFO] [config.py:1062:print]   global_rank .................. 0
[2022-03-09 04:24:02,356] [INFO] [config.py:1062:print]   gradient_accumulation_steps .. 1
[2022-03-09 04:24:02,356] [INFO] [config.py:1062:print]   gradient_clipping ............ 1.0
[2022-03-09 04:24:02,356] [INFO] [config.py:1062:print]   gradient_predivide_factor .... 1.0
[2022-03-09 04:24:02,356] [INFO] [config.py:1062:print]   initial_dynamic_scale ........ 65536
[2022-03-09 04:24:02,356] [INFO] [config.py:1062:print]   loss_scale ................... 0
[2022-03-09 04:24:02,356] [INFO] [config.py:1062:print]   memory_breakdown ............. False
[2022-03-09 04:24:02,356] [INFO] [config.py:1062:print]   optimizer_legacy_fusion ...... False
[2022-03-09 04:24:02,357] [INFO] [config.py:1062:print]   optimizer_name ............... adamw
[2022-03-09 04:24:02,357] [INFO] [config.py:1062:print]   optimizer_params ............. {'lr': 5e-05, 'betas': [0.9, 0.999], 'eps': 1e-08}
[2022-03-09 04:24:02,357] [INFO] [config.py:1062:print]   pipeline ..................... {'stages': 'auto', 'partition': 'best', 'seed_layers': False, 'activation_checkpoint_interval': 0}
[2022-03-09 04:24:02,357] [INFO] [config.py:1062:print]   pld_enabled .................. False
[2022-03-09 04:24:02,357] [INFO] [config.py:1062:print]   pld_params ................... False
[2022-03-09 04:24:02,357] [INFO] [config.py:1062:print]   prescale_gradients ........... False
[2022-03-09 04:24:02,357] [INFO] [config.py:1062:print]   quantize_change_rate ......... 0.001
[2022-03-09 04:24:02,357] [INFO] [config.py:1062:print]   quantize_groups .............. 1
[2022-03-09 04:24:02,357] [INFO] [config.py:1062:print]   quantize_offset .............. 1000
[2022-03-09 04:24:02,358] [INFO] [config.py:1062:print]   quantize_period .............. 1000
[2022-03-09 04:24:02,358] [INFO] [config.py:1062:print]   quantize_rounding ............ 0
[2022-03-09 04:24:02,358] [INFO] [config.py:1062:print]   quantize_start_bits .......... 16
[2022-03-09 04:24:02,358] [INFO] [config.py:1062:print]   quantize_target_bits ......... 8
[2022-03-09 04:24:02,358] [INFO] [config.py:1062:print]   quantize_training_enabled .... False
[2022-03-09 04:24:02,358] [INFO] [config.py:1062:print]   quantize_type ................ 0
[2022-03-09 04:24:02,358] [INFO] [config.py:1062:print]   quantize_verbose ............. False
[2022-03-09 04:24:02,358] [INFO] [config.py:1062:print]   scheduler_name ............... WarmupLR
[2022-03-09 04:24:02,358] [INFO] [config.py:1062:print]   scheduler_params ............. {'warmup_min_lr': 0, 'warmup_max_lr': 5e-05, 'warmup_num_steps': 0}
[2022-03-09 04:24:02,358] [INFO] [config.py:1062:print]   sparse_attention ............. None
[2022-03-09 04:24:02,358] [INFO] [config.py:1062:print]   sparse_gradients_enabled ..... False
[2022-03-09 04:24:02,358] [INFO] [config.py:1062:print]   steps_per_print .............. 10
[2022-03-09 04:24:02,359] [INFO] [config.py:1062:print]   tensorboard_enabled .......... False
[2022-03-09 04:24:02,359] [INFO] [config.py:1062:print]   tensorboard_job_name ......... DeepSpeedJobName
[2022-03-09 04:24:02,359] [INFO] [config.py:1062:print]   tensorboard_output_path ...... 
[2022-03-09 04:24:02,359] [INFO] [config.py:1062:print]   train_batch_size ............. 128
[2022-03-09 04:24:02,359] [INFO] [config.py:1062:print]   train_micro_batch_size_per_gpu  128
[2022-03-09 04:24:02,359] [INFO] [config.py:1062:print]   use_quantizer_kernel ......... False
[2022-03-09 04:24:02,359] [INFO] [config.py:1062:print]   wall_clock_breakdown ......... False
[2022-03-09 04:24:02,359] [INFO] [config.py:1062:print]   world_size ................... 1
[2022-03-09 04:24:02,359] [INFO] [config.py:1062:print]   zero_allow_untested_optimizer  False
[2022-03-09 04:24:02,360] [INFO] [config.py:1062:print]   zero_config .................. {
    "stage": 2, 
    "contiguous_gradients": true, 
    "reduce_scatter": true, 
    "reduce_bucket_size": 5.000000e+08, 
    "allgather_partitions": true, 
    "allgather_bucket_size": 5.000000e+08, 
    "overlap_comm": true, 
    "load_from_fp32_weights": true, 
    "elastic_checkpoint": false, 
    "offload_param": null, 
    "offload_optimizer": {
        "device": null, 
        "nvme_path": null, 
        "buffer_count": 4, 
        "pin_memory": false, 
        "pipeline_read": false, 
        "pipeline_write": false, 
        "fast_init": false
    }, 
    "sub_group_size": 1.000000e+09, 
    "prefetch_bucket_size": 5.000000e+07, 
    "param_persistence_threshold": 1.000000e+05, 
    "max_live_parameters": 1.000000e+09, 
    "max_reuse_distance": 1.000000e+09, 
    "gather_16bit_weights_on_model_save": false, 
    "ignore_unused_parameters": true, 
    "round_robin_gradients": false, 
    "legacy_stage1": false
}
[2022-03-09 04:24:02,360] [INFO] [config.py:1062:print]   zero_enabled ................. True
[2022-03-09 04:24:02,360] [INFO] [config.py:1062:print]   zero_optimization_stage ...... 2
[2022-03-09 04:24:02,360] [INFO] [config.py:1070:print]   json = {
    "train_micro_batch_size_per_gpu": 128, 
    "fp16": {
        "enabled": true, 
        "loss_scale": 0, 
        "loss_scale_window": 1000, 
        "initial_scale_power": 16, 
        "hysteresis": 2, 
        "min_loss_scale": 1
    }, 
    "optimizer": {
        "type": "AdamW", 
        "params": {
            "lr": 5e-05, 
            "betas": [0.9, 0.999], 
            "eps": 1e-08
        }
    }, 
    "scheduler": {
        "type": "WarmupLR", 
        "params": {
            "warmup_min_lr": 0, 
            "warmup_max_lr": 5e-05, 
            "warmup_num_steps": 0
        }
    }, 
    "zero_optimization": {
        "stage": 2, 
        "allgather_partitions": true, 
        "allgather_bucket_size": 5.000000e+08, 
        "overlap_comm": true, 
        "reduce_scatter": true, 
        "reduce_bucket_size": 5.000000e+08, 
        "contiguous_gradients": true, 
        "cpu_offload": true, 
        "offload_optimizer": {
            "device": "cpu", 
            "pin_memory": true, 
            "fast_init": true
        }
    }, 
    "gradient_accumulation_steps": 1, 
    "gradient_clipping": 1.0
}
Using /root/.cache/torch_extensions/py37_cu111 as PyTorch extensions root...
No modifications detected for re-loaded extension module utils, skipping build step...
Loading extension module utils...
Time to load utils op: 0.0008223056793212891 seconds
03/09/2022 04:24:02 - INFO - __main__ -   ***** Running training *****
03/09/2022 04:24:02 - INFO - __main__ -     Num examples = 234528
03/09/2022 04:24:02 - INFO - __main__ -     Num Epochs = 100
03/09/2022 04:24:02 - INFO - __main__ -     Total train batch size = 128
03/09/2022 04:24:02 - INFO - __main__ -     Gradient Accumulation steps = 1
03/09/2022 04:24:02 - INFO - __main__ -     Total optimization steps = 146600
03/09/2022 04:24:02 - INFO - __main__ -     Logging steps = 3000
03/09/2022 04:24:02 - INFO - __main__ -     Save steps = 3000
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
[2022-03-09 04:24:04,721] [INFO] [stage_1_and_2.py:1656:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 65536, reducing to 65536
<IPython.core.display.HTML object>
[2022-03-09 04:24:06,139] [INFO] [stage_1_and_2.py:1656:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 65536, reducing to 32768.0

我知道 Huggingface 和 Pytorch Trainer API 支持深度集成，但是，我想处理当前的代码并解决问题。提前感谢您的帮助:)