[BUG] Same train time with DeepSpeed (despite increased batch size)

Describe the bug and Expected behavior
Hi! I really appreciate your awesome work! I'm fine-tuning Electra model with using huggingface without Trainer API and with using deepspeed. After I applied deepspeed, I could increase the batch size (64 -> 128, but OOM with 256) of training model so I expected train time would decrease. However, even though I applied deepspeed in my code, the train time is the same. I am not sure if deepspeed is working as my configuration and if not, I wonder how I can make it work well (faster training)

I am training the model in ColabProPlus and Environment info is as belows.
( I have been allocated with the same GPU every time)

System info (please complete the following information):

• OS: Ubuntu 18.04.5 LTS
• GPU count and types : 1 GPU, Tesla P100-PCIE-16GB (Colab Pro Plus)
• Python version : 3.7.12
• 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

To Reproduce
I run this command to use deepspeed JIT mode when I initialize runtime in colab.

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

then, my ds_report shows as below before running train code.

ds_report output

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

Deepspeed config
my deepspeed configuration is like this.

"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
        }
}

Train code
My train code is almost same with code of the link in comment and added just few lines to use deepspeed as below.

# 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")
    ....

Output Logs
This is my output log when I run the training.

[2022-03-11 02:05:14,971] [INFO] [logging.py:69:log_dist] [Rank 0] DeepSpeed info: version=0.6.0, git-hash=unknown, git-branch=unknown
03/11/2022 02:05:14 - INFO - torch.distributed.distributed_c10d -   Added key: store_based_barrier_key:2 to store for rank: 0
03/11/2022 02:05:14 - INFO - torch.distributed.distributed_c10d -   Rank 0: Completed store-based barrier for key:store_based_barrier_key:2 with 1 nodes.
[2022-03-11 02:05:15,039] [INFO] [engine.py:278:__init__] DeepSpeed Flops Profiler Enabled: False
Using /root/.cache/torch_extensions/py37_cu111 as PyTorch extensions root...
Creating extension directory /root/.cache/torch_extensions/py37_cu111/cpu_adam...
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)
[1/3] /usr/local/cuda/bin/nvcc  -DTORCH_EXTENSION_NAME=cpu_adam -DTORCH_API_INCLUDE_EXTENSION_H -DPYBIND11_COMPILER_TYPE=\"_gcc\" -DPYBIND11_STDLIB=\"_libstdcpp\" -DPYBIND11_BUILD_ABI=\"_cxxabi1011\" -I/usr/local/lib/python3.7/dist-packages/deepspeed/ops/csrc/includes -I/usr/local/cuda/include -isystem /usr/local/lib/python3.7/dist-packages/torch/include -isystem /usr/local/lib/python3.7/dist-packages/torch/include/torch/csrc/api/include -isystem /usr/local/lib/python3.7/dist-packages/torch/include/TH -isystem /usr/local/lib/python3.7/dist-packages/torch/include/THC -isystem /usr/local/cuda/include -isystem /usr/include/python3.7m -D_GLIBCXX_USE_CXX11_ABI=0 -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_60,code=compute_60 -gencode=arch=compute_60,code=sm_60 --compiler-options '-fPIC' -O3 --use_fast_math -std=c++14 -U__CUDA_NO_HALF_OPERATORS__ -U__CUDA_NO_HALF_CONVERSIONS__ -U__CUDA_NO_HALF2_OPERATORS__ -gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_60,code=compute_60 -c /usr/local/lib/python3.7/dist-packages/deepspeed/ops/csrc/common/custom_cuda_kernel.cu -o custom_cuda_kernel.cuda.o 
[2/3] c++ -MMD -MF cpu_adam.o.d -DTORCH_EXTENSION_NAME=cpu_adam -DTORCH_API_INCLUDE_EXTENSION_H -DPYBIND11_COMPILER_TYPE=\"_gcc\" -DPYBIND11_STDLIB=\"_libstdcpp\" -DPYBIND11_BUILD_ABI=\"_cxxabi1011\" -I/usr/local/lib/python3.7/dist-packages/deepspeed/ops/csrc/includes -I/usr/local/cuda/include -isystem /usr/local/lib/python3.7/dist-packages/torch/include -isystem /usr/local/lib/python3.7/dist-packages/torch/include/torch/csrc/api/include -isystem /usr/local/lib/python3.7/dist-packages/torch/include/TH -isystem /usr/local/lib/python3.7/dist-packages/torch/include/THC -isystem /usr/local/cuda/include -isystem /usr/include/python3.7m -D_GLIBCXX_USE_CXX11_ABI=0 -fPIC -std=c++14 -O3 -std=c++14 -g -Wno-reorder -L/usr/local/cuda/lib64 -lcudart -lcublas -g -march=native -fopenmp -D__AVX256__ -c /usr/local/lib/python3.7/dist-packages/deepspeed/ops/csrc/adam/cpu_adam.cpp -o cpu_adam.o 
[3/3] c++ cpu_adam.o custom_cuda_kernel.cuda.o -shared -lcurand -L/usr/local/lib/python3.7/dist-packages/torch/lib -lc10 -lc10_cuda -ltorch_cpu -ltorch_cuda_cu -ltorch_cuda_cpp -ltorch -ltorch_python -L/usr/local/cuda/lib64 -lcudart -o cpu_adam.so
Loading extension module cpu_adam...
Time to load cpu_adam op: 26.065569162368774 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
[2022-03-11 02:05:42,721] [INFO] [engine.py:1066:_configure_optimizer] Using DeepSpeed Optimizer param name adamw as basic optimizer
[2022-03-11 02:05:42,730] [INFO] [engine.py:1073:_configure_optimizer] DeepSpeed Basic Optimizer = DeepSpeedCPUAdam
[2022-03-11 02:05:42,731] [INFO] [utils.py:49:is_zero_supported_optimizer] Checking ZeRO support for optimizer=DeepSpeedCPUAdam type=<class 'deepspeed.ops.adam.cpu_adam.DeepSpeedCPUAdam'>
[2022-03-11 02:05:42,731] [INFO] [logging.py:69:log_dist] [Rank 0] Creating fp16 ZeRO stage 2 optimizer
[2022-03-11 02:05:42,731] [INFO] [stage_1_and_2.py:125:__init__] Reduce bucket size 500000000.0
[2022-03-11 02:05:42,731] [INFO] [stage_1_and_2.py:126:__init__] Allgather bucket size 500000000.0
[2022-03-11 02:05:42,731] [INFO] [stage_1_and_2.py:127:__init__] CPU Offload: True
[2022-03-11 02:05:42,731] [INFO] [stage_1_and_2.py:128:__init__] Round robin gradient partitioning: False
Using /root/.cache/torch_extensions/py37_cu111 as PyTorch extensions root...
Creating extension directory /root/.cache/torch_extensions/py37_cu111/utils...
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)
[1/2] c++ -MMD -MF flatten_unflatten.o.d -DTORCH_EXTENSION_NAME=utils -DTORCH_API_INCLUDE_EXTENSION_H -DPYBIND11_COMPILER_TYPE=\"_gcc\" -DPYBIND11_STDLIB=\"_libstdcpp\" -DPYBIND11_BUILD_ABI=\"_cxxabi1011\" -isystem /usr/local/lib/python3.7/dist-packages/torch/include -isystem /usr/local/lib/python3.7/dist-packages/torch/include/torch/csrc/api/include -isystem /usr/local/lib/python3.7/dist-packages/torch/include/TH -isystem /usr/local/lib/python3.7/dist-packages/torch/include/THC -isystem /usr/include/python3.7m -D_GLIBCXX_USE_CXX11_ABI=0 -fPIC -std=c++14 -c /usr/local/lib/python3.7/dist-packages/deepspeed/ops/csrc/utils/flatten_unflatten.cpp -o flatten_unflatten.o 
[2/2] c++ flatten_unflatten.o -shared -L/usr/local/lib/python3.7/dist-packages/torch/lib -lc10 -ltorch_cpu -ltorch -ltorch_python -o utils.so
Loading extension module utils...
Time to load utils op: 12.981101274490356 seconds
Rank: 0 partition count [1] and sizes[(113066686, False)] 
[2022-03-11 02:05:56,300] [INFO] [utils.py:824:see_memory_usage] Before initializing optimizer states
[2022-03-11 02:05:56,301] [INFO] [utils.py:829:see_memory_usage] MA 0.26 GB         Max_MA 0.47 GB         CA 0.74 GB         Max_CA 1 GB 
[2022-03-11 02:05:56,302] [INFO] [utils.py:834:see_memory_usage] CPU Virtual Memory:  used = 7.39 GB, percent = 14.5%
[2022-03-11 02:05:56,823] [INFO] [utils.py:824:see_memory_usage] After initializing optimizer states
[2022-03-11 02:05:56,824] [INFO] [utils.py:829:see_memory_usage] MA 0.26 GB         Max_MA 0.26 GB         CA 0.74 GB         Max_CA 1 GB 
[2022-03-11 02:05:56,824] [INFO] [utils.py:834:see_memory_usage] CPU Virtual Memory:  used = 8.66 GB, percent = 17.0%
[2022-03-11 02:05:56,825] [INFO] [stage_1_and_2.py:497:__init__] optimizer state initialized
[2022-03-11 02:05:56,963] [INFO] [utils.py:824:see_memory_usage] After initializing ZeRO optimizer
[2022-03-11 02:05:56,963] [INFO] [utils.py:829:see_memory_usage] MA 0.26 GB         Max_MA 0.26 GB         CA 0.74 GB         Max_CA 1 GB 
[2022-03-11 02:05:56,964] [INFO] [utils.py:834:see_memory_usage] CPU Virtual Memory:  used = 8.66 GB, percent = 17.0%
[2022-03-11 02:05:56,964] [INFO] [logging.py:69:log_dist] [Rank 0] DeepSpeed Final Optimizer = adamw
[2022-03-11 02:05:56,964] [INFO] [engine.py:777:_configure_lr_scheduler] DeepSpeed using configured LR scheduler = WarmupLR
[2022-03-11 02:05:56,964] [INFO] [logging.py:69:log_dist] [Rank 0] DeepSpeed LR Scheduler = <deepspeed.runtime.lr_schedules.WarmupLR object at 0x7f4e91ebe7d0>
[2022-03-11 02:05:56,964] [INFO] [logging.py:69:log_dist] [Rank 0] step=0, skipped=0, lr=[5e-05], mom=[[0.9, 0.999]]
[2022-03-11 02:05:56,965] [INFO] [config.py:1058:print] DeepSpeedEngine configuration:
[2022-03-11 02:05:56,965] [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
}
[2022-03-11 02:05:56,965] [INFO] [config.py:1062:print]   aio_config ................... {'block_size': 1048576, 'queue_depth': 8, 'thread_count': 1, 'single_submit': False, 'overlap_events': True}
[2022-03-11 02:05:56,965] [INFO] [config.py:1062:print]   amp_enabled .................. False
[2022-03-11 02:05:56,965] [INFO] [config.py:1062:print]   amp_params ................... False
[2022-03-11 02:05:56,965] [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
}
[2022-03-11 02:05:56,965] [INFO] [config.py:1062:print]   bfloat16_enabled ............. False
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   checkpoint_tag_validation_enabled  True
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   checkpoint_tag_validation_fail  False
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   communication_data_type ...... None
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   curriculum_enabled ........... False
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   curriculum_params ............ False
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   dataloader_drop_last ......... False
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   disable_allgather ............ False
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   dump_state ................... False
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   dynamic_loss_scale_args ...... {'init_scale': 65536, 'scale_window': 1000, 'delayed_shift': 2, 'min_scale': 1}
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   eigenvalue_enabled ........... False
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   eigenvalue_gas_boundary_resolution  1
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   eigenvalue_layer_name ........ bert.encoder.layer
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   eigenvalue_layer_num ......... 0
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   eigenvalue_max_iter .......... 100
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   eigenvalue_stability ......... 1e-06
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   eigenvalue_tol ............... 0.01
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   eigenvalue_verbose ........... False
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   elasticity_enabled ........... False
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   flops_profiler_config ........ {
    "enabled": false, 
    "profile_step": 1, 
    "module_depth": -1, 
    "top_modules": 1, 
    "detailed": true, 
    "output_file": null
}
[2022-03-11 02:05:56,966] [INFO] [config.py:1062:print]   fp16_enabled ................. True
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   fp16_master_weights_and_gradients  False
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   fp16_mixed_quantize .......... False
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   global_rank .................. 0
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   gradient_accumulation_steps .. 1
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   gradient_clipping ............ 1.0
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   gradient_predivide_factor .... 1.0
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   initial_dynamic_scale ........ 65536
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   loss_scale ................... 0
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   memory_breakdown ............. False
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   optimizer_legacy_fusion ...... False
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   optimizer_name ............... adamw
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   optimizer_params ............. {'lr': 5e-05, 'betas': [0.9, 0.999], 'eps': 1e-08}
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   pipeline ..................... {'stages': 'auto', 'partition': 'best', 'seed_layers': False, 'activation_checkpoint_interval': 0}
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   pld_enabled .................. False
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   pld_params ................... False
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   prescale_gradients ........... False
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   quantize_change_rate ......... 0.001
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   quantize_groups .............. 1
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   quantize_offset .............. 1000
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   quantize_period .............. 1000
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   quantize_rounding ............ 0
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   quantize_start_bits .......... 16
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   quantize_target_bits ......... 8
[2022-03-11 02:05:56,967] [INFO] [config.py:1062:print]   quantize_training_enabled .... False
[2022-03-11 02:05:56,968] [INFO] [config.py:1062:print]   quantize_type ................ 0
[2022-03-11 02:05:56,968] [INFO] [config.py:1062:print]   quantize_verbose ............. False
[2022-03-11 02:05:56,968] [INFO] [config.py:1062:print]   scheduler_name ............... WarmupLR
[2022-03-11 02:05:56,968] [INFO] [config.py:1062:print]   scheduler_params ............. {'warmup_min_lr': 0, 'warmup_max_lr': 5e-05, 'warmup_num_steps': 0}
[2022-03-11 02:05:56,968] [INFO] [config.py:1062:print]   sparse_attention ............. None
[2022-03-11 02:05:56,968] [INFO] [config.py:1062:print]   sparse_gradients_enabled ..... False
[2022-03-11 02:05:56,968] [INFO] [config.py:1062:print]   steps_per_print .............. 10
[2022-03-11 02:05:56,968] [INFO] [config.py:1062:print]   tensorboard_enabled .......... False
[2022-03-11 02:05:56,968] [INFO] [config.py:1062:print]   tensorboard_job_name ......... DeepSpeedJobName
[2022-03-11 02:05:56,968] [INFO] [config.py:1062:print]   tensorboard_output_path ...... 
[2022-03-11 02:05:56,968] [INFO] [config.py:1062:print]   train_batch_size ............. 128
[2022-03-11 02:05:56,968] [INFO] [config.py:1062:print]   train_micro_batch_size_per_gpu  128
[2022-03-11 02:05:56,968] [INFO] [config.py:1062:print]   use_quantizer_kernel ......... False
[2022-03-11 02:05:56,968] [INFO] [config.py:1062:print]   wall_clock_breakdown ......... False
[2022-03-11 02:05:56,968] [INFO] [config.py:1062:print]   world_size ................... 1
[2022-03-11 02:05:56,968] [INFO] [config.py:1062:print]   zero_allow_untested_optimizer  False
[2022-03-11 02:05:56,969] [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
}
[2022-03-11 02:05:56,969] [INFO] [config.py:1062:print]   zero_enabled ................. True
[2022-03-11 02:05:56,969] [INFO] [config.py:1062:print]   zero_optimization_stage ...... 2
[2022-03-11 02:05:56,969] [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: 0.00044083595275878906 seconds
03/11/2022 02:05:56 - INFO - __main__ -   ***** Running training *****
03/11/2022 02:05:56 - INFO - __main__ -     Num examples = 234528
03/11/2022 02:05:56 - INFO - __main__ -     Num Epochs = 100
03/11/2022 02:05:56 - INFO - __main__ -     Total train batch size = 128
03/11/2022 02:05:56 - INFO - __main__ -     Gradient Accumulation steps = 1
03/11/2022 02:05:56 - INFO - __main__ -     Total optimization steps = 146600
03/11/2022 02:05:56 - INFO - __main__ -     Logging steps = 3000
03/11/2022 02:05:56 - INFO - __main__ -     Save steps = 3000
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
[2022-03-11 02:05:59,192] [INFO] [stage_1_and_2.py:1656:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 65536, reducing to 65536
/usr/local/lib/python3.7/dist-packages/torch/optim/lr_scheduler.py:134: UserWarning: Detected call of `lr_scheduler.step()` before `optimizer.step()`. In PyTorch 1.1.0 and later, you should call them in the opposite order: `optimizer.step()` before `lr_scheduler.step()`.  Failure to do this will result in PyTorch skipping the first value of the learning rate schedule. See more details at https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate
  "https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate", UserWarning)
<IPython.core.display.HTML object>
[2022-03-11 02:06:00,592] [INFO] [stage_1_and_2.py:1656:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 65536, reducing to 32768.0
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
[2022-03-11 02:06:13,536] [INFO] [logging.py:69:log_dist] [Rank 0] step=10, skipped=2, lr=[5e-05], mom=[[0.9, 0.999]]
[2022-03-11 02:06:13,537] [INFO] [timer.py:189:stop] 0/10, SamplesPerSec=79.2758999705874, MemAllocated=0.26GB, MaxMemAllocated=9.02GB
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
[2022-03-11 02:06:28,687] [INFO] [logging.py:69:log_dist] [Rank 0] step=20, skipped=2, lr=[5e-05], mom=[[0.9, 0.999]]
[2022-03-11 02:06:28,688] [INFO] [timer.py:189:stop] 0/20, SamplesPerSec=82.1916862587935, MemAllocated=0.26GB, MaxMemAllocated=9.02GB
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
[2022-03-11 02:06:43,833] [INFO] [logging.py:69:log_dist] [Rank 0] step=30, skipped=2, lr=[5e-05], mom=[[0.9, 0.999]]
[2022-03-11 02:06:43,834] [INFO] [timer.py:189:stop] 0/30, SamplesPerSec=83.07239837441124, MemAllocated=0.26GB, MaxMemAllocated=9.02GB
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
[2022-03-11 02:06:58,985] [INFO] [logging.py:69:log_dist] [Rank 0] step=40, skipped=2, lr=[5e-05], mom=[[0.9, 0.999]]
[2022-03-11 02:06:58,985] [INFO] [timer.py:189:stop] 0/40, SamplesPerSec=83.48609168106144, MemAllocated=0.26GB, MaxMemAllocated=9.02GB
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
[2022-03-11 02:07:14,130] [INFO] [logging.py:69:log_dist] [Rank 0] step=50, skipped=2, lr=[5e-05], mom=[[0.9, 0.999]]
[2022-03-11 02:07:14,130] [INFO] [timer.py:189:stop] 0/50, SamplesPerSec=83.73838805358588, MemAllocated=0.26GB, MaxMemAllocated=9.02GB
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
[2022-03-11 02:07:29,274] [INFO] [logging.py:69:log_dist] [Rank 0] step=60, skipped=2, lr=[5e-05], mom=[[0.9, 0.999]]
[2022-03-11 02:07:29,274] [INFO] [timer.py:189:stop] 0/60, SamplesPerSec=83.9044415557675, MemAllocated=0.26GB, MaxMemAllocated=9.02GB
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
<IPython.core.display.HTML object>
[2022-03-11 02:07:44,417] [INFO] [logging.py:69:log_dist] [Rank 0] step=70, skipped=2, lr=[5e-05], mom=[[0.9, 0.999]]
[2022-03-11 02:07:44,418] [INFO] [timer.py:189:stop] 0/70, SamplesPerSec=84.02348484813439, MemAllocated=0.26GB, MaxMemAllocated=9.02GB

Start with the linked KoELECTRA run_seq_cls.py example and the shown deepspeed.initialize call, then compare the ds_config, training loop, and output logs. Reproduce the timing with and without ZeRO stage 2 and document whether the reported configuration is active and what explains the unchanged training time.