OOM despite ZeRO stage 3

Author: kchu02Created Jul 13, 2022Updated Jul 23, 2025

Following the example in HelloDeepSpeed, yet I still have to CUDA OOM despite moving all the way to stage 3 on the configuration below.

deepspeed train_bert_ds.py --checkpoint_dir . --num_layers 24 --h_dim 4096

I added offload options to support larger model but it was still OOM on my 16GB V100.

ds_config = {
    "train_micro_batch_size_per_gpu": batch_size,
    "optimizer": {
        "type": "Adam",
        "params": {
            "lr": 1e-4
        }
    },
    "fp16": {
        "enabled": True
    },
    "zero_optimization": {
       "stage": 3,
       "offload_optimizer": {
            "device": "cpu",
            "pin_memory": True
        },
        "offload_param": {
            "device": "cpu",
            "pin_memory": True
        }
    }
}

Is there anything missing here? Please let me know if there is a documentation on config where the model is too large to be loaded onto a single GPU.

Thank you.

Version info

python --version
Python 3.9.12

python -c "import torch; print(torch.__version__)"
1.12.0+cu116

python -c "import deepspeed; print(deepspeed.__version__)"
0.6.5

Log output

[2022-07-13 21:03:49,576] [WARNING] [runner.py:159:fetch_hostfile] Unable to find hostfile, will proceed with training with local resources only.
[2022-07-13 21:03:49,576] [INFO] [runner.py:457:main] cmd = /working/anaconda3/envs/tmp/bin/python -u -m deepspeed.launcher.launch --world_info=eyJsb2NhbGhvc3QiOiBbMCwgMSwgMiwgM119 --master_addr=127.0.0.1 --master_port=29500 train_bert_ds.py --checkpoint_dir . --num_layers 24 --h_dim 4096
[2022-07-13 21:03:50,730] [INFO] [launch.py:103:main] WORLD INFO DICT: {'localhost': [0, 1, 2, 3]}
[2022-07-13 21:03:50,731] [INFO] [launch.py:109:main] nnodes=1, num_local_procs=4, node_rank=0
[2022-07-13 21:03:50,731] [INFO] [launch.py:122:main] global_rank_mapping=defaultdict(<class 'list'>, {'localhost': [0, 1, 2, 3]})
[2022-07-13 21:03:50,731] [INFO] [launch.py:123:main] dist_world_size=4
[2022-07-13 21:03:50,731] [INFO] [launch.py:125:main] Setting CUDA_VISIBLE_DEVICES=0,1,2,3
[2022-07-13 21:04:39,659] [INFO] [logging.py:69:log_dist] [Rank -1] DeepSpeed info: version=0.6.5, git-hash=unknown, git-branch=unknown
[2022-07-13 21:04:39,905] [INFO] [logging.py:69:log_dist] [Rank -1] DeepSpeed info: version=0.6.5, git-hash=unknown, git-branch=unknown
[2022-07-13 21:04:39,909] [INFO] [distributed.py:48:init_distributed] Initializing torch distributed with backend: nccl
[2022-07-13 21:04:40,010] [INFO] [logging.py:69:log_dist] [Rank -1] DeepSpeed info: version=0.6.5, git-hash=unknown, git-branch=unknown
[2022-07-13 21:04:40,289] [INFO] [logging.py:69:log_dist] [Rank -1] DeepSpeed info: version=0.6.5, git-hash=unknown, git-branch=unknown
[2022-07-13 21:04:46,103] [INFO] [engine.py:278:__init__] DeepSpeed Flops Profiler Enabled: False
Using /home/user/.cache/torch_extensions/py39_cu116 as PyTorch extensions root...
Using /home/user/.cache/torch_extensions/py39_cu116 as PyTorch extensions root...
Using /home/user/.cache/torch_extensions/py39_cu116 as PyTorch extensions root...
Using /home/user/.cache/torch_extensions/py39_cu116 as PyTorch extensions root...
Creating extension directory /home/user/.cache/torch_extensions/py39_cu116/cpu_adam...Creating extension directory /home/user/.cache/torch_extensions/py39_cu116/cpu_adam...Creating extension directory /home/user/.cache/torch_extensions/py39_cu116/cpu_adam...Creating extension directory /home/user/.cache/torch_extensions/py39_cu116/cpu_adam...



Detected CUDA files, patching ldflags
Emitting ninja build file /home/user/.cache/torch_extensions/py39_cu116/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=\"_cxxabi1013\" -I/working/anaconda3/envs/tmp/lib/python3.9/site-packages/deepspeed/ops/csrc/includes -I/usr/local/cuda/include -isystem /working/anaconda3/envs/tmp/lib/python3.9/site-packages/torch/include -isystem /working/anaconda3/envs/tmp/lib/python3.9/site-packages/torch/include/torch/csrc/api/include -isystem /working/anaconda3/envs/tmp/lib/python3.9/site-packages/torch/include/TH -isystem /working/anaconda3/envs/tmp/lib/python3.9/site-packages/torch/include/THC -isystem /usr/local/cuda/include -isystem /working/anaconda3/envs/tmp/include/python3.9 -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_70,code=compute_70 -gencode=arch=compute_70,code=sm_70 --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_70,code=sm_70 -gencode=arch=compute_70,code=compute_70 -c /working/anaconda3/envs/tmp/lib/python3.9/site-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=\"_cxxabi1013\" -I/working/anaconda3/envs/tmp/lib/python3.9/site-packages/deepspeed/ops/csrc/includes -I/usr/local/cuda/include -isystem /working/anaconda3/envs/tmp/lib/python3.9/site-packages/torch/include -isystem /working/anaconda3/envs/tmp/lib/python3.9/site-packages/torch/include/torch/csrc/api/include -isystem /working/anaconda3/envs/tmp/lib/python3.9/site-packages/torch/include/TH -isystem /working/anaconda3/envs/tmp/lib/python3.9/site-packages/torch/include/THC -isystem /usr/local/cuda/include -isystem /working/anaconda3/envs/tmp/include/python3.9 -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 /working/anaconda3/envs/tmp/lib/python3.9/site-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/working/anaconda3/envs/tmp/lib/python3.9/site-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...
Loading extension module cpu_adam...Loading extension module cpu_adam...

Loading extension module cpu_adam...
Time to load cpu_adam op: 63.22985577583313 seconds
Time to load cpu_adam op: 63.2216637134552 secondsTime to load cpu_adam op: 63.241652727127075 seconds

Time to load cpu_adam op: 63.24095416069031 seconds
Adam Optimizer #0 is created with AVX2 arithmetic capability.
Config: alpha=0.000100, betas=(0.900000, 0.999000), weight_decay=0.000000, adam_w=1
Adam Optimizer #0 is created with AVX2 arithmetic capability.
Config: alpha=0.000100, betas=(0.900000, 0.999000), weight_decay=0.000000, adam_w=1
Adam Optimizer #0 is created with AVX2 arithmetic capability.
Config: alpha=0.000100, betas=(0.900000, 0.999000), weight_decay=0.000000, adam_w=1
Adam Optimizer #0 is created with AVX2 arithmetic capability.
Config: alpha=0.000100, betas=(0.900000, 0.999000), weight_decay=0.000000, adam_w=1
[2022-07-13 21:05:51,130] [INFO] [engine.py:1100:_configure_optimizer] Using DeepSpeed Optimizer param name adam as basic optimizer
Using /home/user/.cache/torch_extensions/py39_cu116 as PyTorch extensions root...
Using /home/user/.cache/torch_extensions/py39_cu116 as PyTorch extensions root...
Using /home/user/.cache/torch_extensions/py39_cu116 as PyTorch extensions root...
[2022-07-13 21:05:51,164] [INFO] [engine.py:1108:_configure_optimizer] DeepSpeed Basic Optimizer = DeepSpeedCPUAdam
[2022-07-13 21:05:51,164] [INFO] [utils.py:52:is_zero_supported_optimizer] Checking ZeRO support for optimizer=DeepSpeedCPUAdam type=<class 'deepspeed.ops.adam.cpu_adam.DeepSpeedCPUAdam'>
[2022-07-13 21:05:51,164] [INFO] [logging.py:69:log_dist] [Rank 0] Creating fp16 ZeRO stage 3 optimizer
[2022-07-13 21:05:51,164] [INFO] [engine.py:1410:_configure_zero_optimizer] Initializing ZeRO Stage 3
[2022-07-13 21:05:51,174] [INFO] [stage3.py:275:__init__] Reduce bucket size 500000000
[2022-07-13 21:05:51,174] [INFO] [stage3.py:276:__init__] Prefetch bucket size 50000000
Using /home/user/.cache/torch_extensions/py39_cu116 as PyTorch extensions root...
Emitting ninja build file /home/user/.cache/torch_extensions/py39_cu116/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: 0.9306187629699707 seconds
Loading extension module utils...Loading extension module utils...

Loading extension module utils...
Time to load utils op: 0.9696803092956543 seconds
Time to load utils op: 0.9813756942749023 seconds
Time to load utils op: 1.0037789344787598 seconds
[2022-07-13 21:05:59,485] [INFO] [stage3.py:567:_setup_for_real_optimizer] optimizer state initialized
Using /home/user/.cache/torch_extensions/py39_cu116 as PyTorch extensions root...
Using /home/user/.cache/torch_extensions/py39_cu116 as PyTorch extensions root...
No modifications detected for re-loaded extension module utils, skipping build step...
Loading extension module utils...Using /home/user/.cache/torch_extensions/py39_cu116 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: 0.03724932670593262 seconds
Time to load utils op: 0.037149906158447266 seconds
No modifications detected for re-loaded extension module utils, skipping build step...
Loading extension module utils...
Time to load utils op: 0.037335872650146484 seconds
[2022-07-13 21:06:00,737] [INFO] [utils.py:828:see_memory_usage] After initializing ZeRO optimizer
[2022-07-13 21:06:00,738] [INFO] [utils.py:829:see_memory_usage] MA 0.93 GB         Max_MA 3.64 GB         CA 5.35 GB         Max_CA 5 GB 
[2022-07-13 21:06:00,738] [INFO] [utils.py:837:see_memory_usage] CPU Virtual Memory:  used = 56.33 GB, percent = 23.5%
[2022-07-13 21:06:00,738] [INFO] [logging.py:69:log_dist] [Rank 0] DeepSpeed Final Optimizer = adam
[2022-07-13 21:06:00,738] [INFO] [engine.py:795:_configure_lr_scheduler] DeepSpeed using client LR scheduler
[2022-07-13 21:06:00,738] [INFO] [logging.py:69:log_dist] [Rank 0] DeepSpeed LR Scheduler = None
[2022-07-13 21:06:00,739] [INFO] [logging.py:69:log_dist] [Rank 0] step=0, skipped=0, lr=[0.0001], mom=[(0.9, 0.999)]
[2022-07-13 21:06:00,739] [INFO] [config.py:1059:print] DeepSpeedEngine configuration:
[2022-07-13 21:06:00,740] [INFO] [config.py:1063:print]   activation_checkpointing_config  {
    "partition_activations": false, 
    "contiguous_memory_optimization": false, 
    "cpu_checkpointing": false, 
    "number_checkpoints": null, 
    "synchronize_checkpoint_boundary": false, 
    "profile": false
}
[2022-07-13 21:06:00,740] [INFO] [config.py:1063:print]   aio_config ................... {'block_size': 1048576, 'queue_depth': 8, 'thread_count': 1, 'single_submit': False, 'overlap_events': True}
[2022-07-13 21:06:00,740] [INFO] [config.py:1063:print]   amp_enabled .................. False
[2022-07-13 21:06:00,740] [INFO] [config.py:1063:print]   amp_params ................... False
[2022-07-13 21:06:00,740] [INFO] [config.py:1063: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
}
[2022-07-13 21:06:00,740] [INFO] [config.py:1063:print]   bfloat16_enabled ............. False
[2022-07-13 21:06:00,740] [INFO] [config.py:1063:print]   checkpoint_tag_validation_enabled  True
[2022-07-13 21:06:00,740] [INFO] [config.py:1063:print]   checkpoint_tag_validation_fail  False
[2022-07-13 21:06:00,740] [INFO] [config.py:1063:print]   communication_data_type ...... None
[2022-07-13 21:06:00,740] [INFO] [config.py:1063:print]   curriculum_enabled ........... False
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   curriculum_params ............ False
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   dataloader_drop_last ......... False
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   disable_allgather ............ False
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   dump_state ................... False
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   dynamic_loss_scale_args ...... None
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   eigenvalue_enabled ........... False
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   eigenvalue_gas_boundary_resolution  1
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   eigenvalue_layer_name ........ bert.encoder.layer
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   eigenvalue_layer_num ......... 0
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   eigenvalue_max_iter .......... 100
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   eigenvalue_stability ......... 1e-06
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   eigenvalue_tol ............... 0.01
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   eigenvalue_verbose ........... False
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   elasticity_enabled ........... False
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   flops_profiler_config ........ {
    "enabled": false, 
    "profile_step": 1, 
    "module_depth": -1, 
    "top_modules": 1, 
    "detailed": true, 
    "output_file": null
}
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   fp16_enabled ................. True
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   fp16_master_weights_and_gradients  False
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   fp16_mixed_quantize .......... False
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   global_rank .................. 0
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   gradient_accumulation_steps .. 1
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   gradient_clipping ............ 0.0
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   gradient_predivide_factor .... 1.0
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   initial_dynamic_scale ........ 4294967296
[2022-07-13 21:06:00,741] [INFO] [config.py:1063:print]   loss_scale ................... 0
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   memory_breakdown ............. False
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   optimizer_legacy_fusion ...... False
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   optimizer_name ............... adam
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   optimizer_params ............. {'lr': 0.0001}
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   pipeline ..................... {'stages': 'auto', 'partition': 'best', 'seed_layers': False, 'activation_checkpoint_interval': 0}
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   pld_enabled .................. False
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   pld_params ................... False
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   prescale_gradients ........... False
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   quantize_change_rate ......... 0.001
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   quantize_groups .............. 1
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   quantize_offset .............. 1000
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   quantize_period .............. 1000
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   quantize_rounding ............ 0
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   quantize_start_bits .......... 16
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   quantize_target_bits ......... 8
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   quantize_training_enabled .... False
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   quantize_type ................ 0
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   quantize_verbose ............. False
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   scheduler_name ............... None
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   scheduler_params ............. None
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   sparse_attention ............. None
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   sparse_gradients_enabled ..... False
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   steps_per_print .............. 10
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   tensorboard_enabled .......... False
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   tensorboard_job_name ......... DeepSpeedJobName
[2022-07-13 21:06:00,742] [INFO] [config.py:1063:print]   tensorboard_output_path ...... 
[2022-07-13 21:06:00,743] [INFO] [config.py:1063:print]   train_batch_size ............. 32
[2022-07-13 21:06:00,743] [INFO] [config.py:1063:print]   train_micro_batch_size_per_gpu  8
[2022-07-13 21:06:00,743] [INFO] [config.py:1063:print]   use_quantizer_kernel ......... False
[2022-07-13 21:06:00,743] [INFO] [config.py:1063:print]   wall_clock_breakdown ......... False
[2022-07-13 21:06:00,743] [INFO] [config.py:1063:print]   world_size ................... 4
[2022-07-13 21:06:00,743] [INFO] [config.py:1063:print]   zero_allow_untested_optimizer  False
[2022-07-13 21:06:00,743] [INFO] [config.py:1063:print]   zero_config .................. {
    "stage": 3, 
    "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": {
        "device": "cpu", 
        "nvme_path": null, 
        "buffer_count": 5, 
        "buffer_size": 1.000000e+08, 
        "max_in_cpu": 1.000000e+09, 
        "pin_memory": true
    }, 
    "offload_optimizer": {
        "device": "cpu", 
        "nvme_path": null, 
        "buffer_count": 4, 
        "pin_memory": true, 
        "pipeline_read": false, 
        "pipeline_write": false, 
        "fast_init": false, 
        "pipeline": 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
}
[2022-07-13 21:06:00,743] [INFO] [config.py:1063:print]   zero_enabled ..

Source: deepspeedai/DeepSpeedExamples

View original on GitHub View discussion on GitHub