[BUG] ZeRO Stage 3 is consuming more memory than Stage 2 when using DeepSpeed-Chat

Describe the bug

When using DeepSpeed-Chat, ZeRO Stage 3 appears to consume more memory than Stage 2. While fine-tuning the Pythia-6.9b model with a context length of approximately 1200 and a LoRA_rank of 128, I observed the following memory behavior:

After deepspeed.initialize:
Stage 2: 16.6 GB
Stage 3: 6.6 GB

Immediately after the first forward pass:
Stage 2: 32.4 GB
Stage 3: 36.6 GB

This indicates a noticeable increase in memory consumption for Stage 3 compared to Stage 2 during the forward pass.

To Reproduce
Steps to reproduce the behavior:
Excuting DeepSpeed-Chat (supervised finetuning) with Pythia-6.9b, LoRA_rank=128

Expected behavior
Stage 3 is supposed to consume less memory than Stage 2

System info (please complete the following information):

• OS: NAME=20.04.5 LTS (Focal Fossa)
• GPU count and types = 4*RTX6000
• Python version = 3.10
• DeepSpeed = 0.10.2

Launcher context
deepspeed.launcher.launch

Full log

2 Time to load fused_adam op: 0.5769875049591064 seconds
3 [2023-09-12 00:26:08,683] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed info: version=0.10.2, git-hash=unknown, git-branch=unknown
4 [2023-09-12 00:26:08,683] [INFO] [comm.py:662:init_distributed] Distributed backend already initialized
5 [2023-09-12 00:26:08,874] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed Flops Profiler Enabled: False
6 [2023-09-12 00:26:08,908] [INFO] [logging.py:96:log_dist] [Rank 0] Removing param_group that has no 'params' in the client Optimizer
7 [2023-09-12 00:26:08,909] [INFO] [logging.py:96:log_dist] [Rank 0] Using client Optimizer as basic optimizer
8 [2023-09-12 00:26:08,983] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed Basic Optimizer = FusedAdam
9 [2023-09-12 00:26:08,983] [INFO] [utils.py:54:is_zero_supported_optimizer] Checking ZeRO support for optimizer=FusedAdam type=<class 'deepspeed.ops.adam.fused_adam.FusedAdam'>
10 [2023-09-12 00:26:08,983] [INFO] [logging.py:96:log_dist] [Rank 0] Creating fp16 ZeRO stage 3 optimizer, MiCS is enabled False, Hierarchical params gather False
11 [2023-09-12 00:26:08,984] [INFO] [logging.py:96:log_dist] [Rank 0] Creating torch.float16 ZeRO stage 3 optimizer
12 [2023-09-12 00:26:09,304] [INFO] [utils.py:803:see_memory_usage] Stage 3 initialize beginning
13 [2023-09-12 00:26:09,307] [INFO] [utils.py:804:see_memory_usage] MA 4.03 GB Max_MA 4.39 GB CA 6.77 GB Max_CA 7 GB
14 [2023-09-12 00:26:09,308] [INFO] [utils.py:811:see_memory_usage] CPU Virtual Memory: used = 45.66 GB, percent = 11.6%
15 [2023-09-12 00:26:09,316] [INFO] [stage3.py:126:init] Reduce bucket size 500,000,000
16 [2023-09-12 00:26:09,316] [INFO] [stage3.py:127:init] Prefetch bucket size 30000000
17 Using /home/ubuntu/.cache/torch_extensions/py310_cu118 as PyTorch extensions root...
18 Detected CUDA files, patching ldflags
19 Emitting ninja build file /home/ubuntu/.cache/torch_extensions/py310_cu118/fused_adam/build.ninja...
20 Building extension module fused_adam...
21 Allowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N)
22 Loading extension module fused_adam...
23 [2023-09-12 00:26:09,668] [INFO] [utils.py:803:see_memory_usage] DeepSpeedZeRoOffload initialize [begin]
24 [2023-09-12 00:26:09,670] [INFO] [utils.py:804:see_memory_usage] MA 4.03 GB Max_MA 4.03 GB CA 6.77 GB Max_CA 7 GB
25 [2023-09-12 00:26:09,671] [INFO] [utils.py:811:see_memory_usage] CPU Virtual Memory: used = 45.66 GB, percent = 11.6%
26 Parameter Offload: Total persistent parameters: 794624 in 194 params
27 [2023-09-12 00:26:10,345] [INFO] [utils.py:803:see_memory_usage] DeepSpeedZeRoOffload initialize [end]
28 [2023-09-12 00:26:10,348] [INFO] [utils.py:804:see_memory_usage] MA 3.65 GB Max_MA 4.03 GB CA 6.77 GB Max_CA 7 GB
29 [2023-09-12 00:26:10,348] [INFO] [utils.py:811:see_memory_usage] CPU Virtual Memory: used = 45.67 GB, percent = 11.6%
30 [2023-09-12 00:26:10,667] [INFO] [utils.py:803:see_memory_usage] Before creating fp16 partitions
31 [2023-09-12 00:26:10,669] [INFO] [utils.py:804:see_memory_usage] MA 3.65 GB Max_MA 3.65 GB CA 6.77 GB Max_CA 7 GB
32 [2023-09-12 00:26:10,670] [INFO] [utils.py:811:see_memory_usage] CPU Virtual Memory: used = 45.67 GB, percent = 11.6%
33 [2023-09-12 00:26:11,405] [INFO] [utils.py:803:see_memory_usage] After creating fp16 partitions: 1
34 [2023-09-12 00:26:11,408] [INFO] [utils.py:804:see_memory_usage] MA 3.65 GB Max_MA 3.65 GB CA 3.92 GB Max_CA 7 GB
35 [2023-09-12 00:26:11,408] [INFO] [utils.py:811:see_memory_usage] CPU Virtual Memory: used = 45.68 GB, percent = 11.6%
36 [2023-09-12 00:26:11,700] [INFO] [utils.py:803:see_memory_usage] Before creating fp32 partitions
37 [2023-09-12 00:26:11,702] [INFO] [utils.py:804:see_memory_usage] MA 3.65 GB Max_MA 3.65 GB CA 3.92 GB Max_CA 4 GB
38 [2023-09-12 00:26:11,702] [INFO] [utils.py:811:see_memory_usage] CPU Virtual Memory: used = 45.68 GB, percent = 11.6%
39 [2023-09-12 00:26:11,996] [INFO] [utils.py:803:see_memory_usage] After creating fp32 partitions
40 [2023-09-12 00:26:11,999] [INFO] [utils.py:804:see_memory_usage] MA 3.9 GB Max_MA 4.03 GB CA 4.3 GB Max_CA 4 GB
41 [2023-09-12 00:26:11,999] [INFO] [utils.py:811:see_memory_usage] CPU Virtual Memory: used = 45.68 GB, percent = 11.6%
42 [2023-09-12 00:26:12,291] [INFO] [utils.py:803:see_memory_usage] Before initializing optimizer states
43 [2023-09-12 00:26:12,294] [INFO] [utils.py:804:see_memory_usage] MA 3.9 GB Max_MA 3.9 GB CA 4.3 GB Max_CA 4 GB
44 [2023-09-12 00:26:12,295] [INFO] [utils.py:811:see_memory_usage] CPU Virtual Memory: used = 45.68 GB, percent = 11.6%
45 [2023-09-12 00:26:12,590] [INFO] [utils.py:803:see_memory_usage] After initializing optimizer states
46 [2023-09-12 00:26:12,592] [INFO] [utils.py:804:see_memory_usage] MA 4.4 GB Max_MA 4.65 GB CA 5.05 GB Max_CA 5 GB
47 [2023-09-12 00:26:12,593] [INFO] [utils.py:811:see_memory_usage] CPU Virtual Memory: used = 45.68 GB, percent = 11.6%
48 [2023-09-12 00:26:12,594] [INFO] [stage3.py:445:_setup_for_real_optimizer] optimizer state initialized
49 [2023-09-12 00:26:13,108] [INFO] [utils.py:803:see_memory_usage] After initializing ZeRO optimizer
50 [2023-09-12 00:26:13,111] [INFO] [utils.py:804:see_memory_usage] MA 5.46 GB Max_MA 5.47 GB CA 5.98 GB Max_CA 6 GB
51 [2023-09-12 00:26:13,111] [INFO] [utils.py:811:see_memory_usage] CPU Virtual Memory: used = 45.68 GB, percent = 11.6%
52 [2023-09-12 00:26:13,112] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed Final Optimizer = FusedAdam
53 [2023-09-12 00:26:13,112] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed using client LR scheduler
54 [2023-09-12 00:26:13,113] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed LR Scheduler = <torch.optim.lr_scheduler.LambdaLR object at 0x7f487c2cddb0>
55 [2023-09-12 00:26:13,113] [INFO] [logging.py:96:log_dist] [Rank 0] step=0, skipped=0, lr=[0.0005], mom=[(0.9, 0.95)]
56 [2023-09-12 00:26:13,135] [INFO] [config.py:963:print] DeepSpeedEngine configuration:
57 [2023-09-12 00:26:13,135] [INFO] [config.py:967:print] activation_checkpointing_config {
58 "partition_activations": true,
59 "contiguous_memory_optimization": true,
60 "cpu_checkpointing": true,
61 "number_checkpoints": 5,
62 "synchronize_checkpoint_boundary": false,
63 "profile": false
64 }
65 [2023-09-12 00:26:13,136] [INFO] [config.py:967:print] aio_config ................... {'block_size': 1048576, 'queue_depth': 8, 'thread_count': 1, 'single_submit': False, 'overlap_events': True}
66 [2023-09-12 00:26:13,136] [INFO] [config.py:967:print] amp_enabled .................. False
67 [2023-09-12 00:26:13,136] [INFO] [config.py:967:print] amp_params ................... False
68 [2023-09-12 00:26:13,139] [INFO] [config.py:967:print] autotuning_config ............ {
69 "enabled": false,
70 "start_step": null,
71 "end_step": null,
72 "metric_path": null,
73 "arg_mappings": null,
74 "metric": "throughput",
75 "model_info": null,
76 "results_dir": "autotuning_results",
77 "exps_dir": "autotuning_exps",
78 "overwrite": true,
79 "fast": true,
80 "start_profile_step": 3,
81 "end_profile_step": 5,
82 "tuner_type": "gridsearch",
83 "tuner_early_stopping": 5,
84 "tuner_num_trials": 50,
85 "model_info_path": null,
86 "mp_size": 1,
87 "max_train_batch_size": null,
88 "min_train_batch_size": 1,
89 "max_train_micro_batch_size_per_gpu": 1.024000e+03,
90 "min_train_micro_batch_size_per_gpu": 1,
91 "num_tuning_micro_batch_sizes": 3
92 }
93 [2023-09-12 00:26:13,139] [INFO] [config.py:967:print] bfloat16_enabled ............. False
94 [2023-09-12 00:26:13,139] [INFO] [config.py:967:print] checkpoint_parallel_write_pipeline False
95 [2023-09-12 00:26:13,140] [INFO] [config.py:967:print] checkpoint_tag_validation_enabled True
96 [2023-09-12 00:26:13,140] [INFO] [config.py:967:print] checkpoint_tag_validation_fail False
97 [2023-09-12 00:26:13,140] [INFO] [config.py:967:print] comms_config ................. <deepspeed.comm.config.DeepSpeedCommsConfig object at 0x7f44d015c9a0>
98 [2023-09-12 00:26:13,140] [INFO] [config.py:967:print] communication_data_type ...... None
99 [2023-09-12 00:26:13,140] [INFO] [config.py:967: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}}
100 [2023-09-12 00:26:13,141] [INFO] [config.py:967:print] curriculum_enabled_legacy .... False
101 [2023-09-12 00:26:13,141] [INFO] [config.py:967:print] curriculum_params_legacy ..... False
102 [2023-09-12 00:26:13,141] [INFO] [config.py:967: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}}}}
103 [2023-09-12 00:26:13,141] [INFO] [config.py:967:print] data_efficiency_enabled ...... False
104 [2023-09-12 00:26:13,142] [INFO] [config.py:967:print] dataloader_drop_last ......... False
105 [2023-09-12 00:26:13,142] [INFO] [config.py:967:print] disable_allgather ............ False
106 [2023-09-12 00:26:13,142] [INFO] [config.py:967:print] dump_state ................... False
107 [2023-09-12 00:26:13,142] [INFO] [config.py:967:print] dynamic_loss_scale_args ...... {'init_scale': 65536, 'scale_window': 100, 'delayed_shift': 2, 'consecutive_hysteresis': False, 'min_scale': 1}
108 [2023-09-12 00:26:13,143] [INFO] [config.py:967:print] eigenvalue_enabled ........... False
109 [2023-09-12 00:26:13,143] [INFO] [config.py:967:print] eigenvalue_gas_boundary_resolution 1
110 [2023-09-12 00:26:13,143] [INFO] [config.py:967:print] eigenvalue_layer_name ........ bert.encoder.layer
111 [2023-09-12 00:26:13,144] [INFO] [config.py:967:print] eigenvalue_layer_num ......... 0
112 [2023-09-12 00:26:13,144] [INFO] [config.py:967:print] eigenvalue_max_iter .......... 100
113 [2023-09-12 00:26:13,144] [INFO] [config.py:967:print] eigenvalue_stability ......... 1e-06
114 [2023-09-12 00:26:13,144] [INFO] [config.py:967:print] eigenvalue_tol ............... 0.01
115 [2023-09-12 00:26:13,145] [INFO] [config.py:967:print] eigenvalue_verbose ........... False
116 [2023-09-12 00:26:13,145] [INFO] [config.py:967:print] elasticity_enabled ........... False
117 [2023-09-12 00:26:13,145] [INFO] [config.py:967:print] flops_profiler_config ........ {
118 "enabled": false,
119 "recompute_fwd_factor": 0.0,
120 "profile_step": 1,
121 "module_depth": -1,
122 "top_modules": 1,
123 "detailed": true,
124 "output_file": null
125 }
126 [2023-09-12 00:26:13,145] [INFO] [config.py:967:print] fp16_auto_cast ............... False
127 [2023-09-12 00:26:13,146] [INFO] [config.py:967:print] fp16_enabled ................. True
128 [2023-09-12 00:26:13,146] [INFO] [config.py:967:print] fp16_master_weights_and_gradients False
129 [2023-09-12 00:26:13,146] [INFO] [config.py:967:print] global_rank .................. 0
130 [2023-09-12 00:26:13,146] [INFO] [config.py:967:print] grad_accum_dtype ............. None
131 [2023-09-12 00:26:13,147] [INFO] [config.py:967:print] gradient_accumulation_steps .. 16
132 [2023-09-12 00:26:13,147] [INFO] [config.py:967:print] gradient_clipping ............ 1.0
133 [2023-09-12 00:26:13,147] [INFO] [config.py:967:print] gradient_predivide_factor .... 1.0
134 [2023-09-12 00:26:13,148] [INFO] [config.py:967:print] hybrid_engine ................ enabled=False max_out_tokens=1250 inference_tp_size=1 release_inference_cache=False pin_parameters=True tp_gather_partition_size=8
135 [2023-09-12 00:26:13,148] [INFO] [config.py:967:print] initial_dynamic_scale ........ 65536
136 [2023-09-12 00:26:13,148] [INFO] [config.py:967:print] load_universal_checkpoint .... False
137 [2023-09-12 00:26:13,149] [INFO] [config.py:967:print] loss_scale ................... 0
138 [2023-09-12 00:26:13,149] [INFO] [config.py:967:print] memory_breakdown ............. False
139 [2023-09-12 00:26:13,149] [INFO] [config.py:967:print] mics_hierarchial_params_gather False
140 [2023-09-12 00:26:13,149] [INFO] [config.py:967:print] mics_shard_size .............. -1
141 [2023-09-12 00:26:13,150] [INFO] [config.py:967: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
142 [2023-09-12 00:26:13,150] [INFO] [config.py:967:print] nebula_config ................ {
143 "enabled": false,
144 "persistent_storage_path": null,
145 "persistent_time_interval": 100,
146 "num_of_version_in_retention": 2,
147 "enable_nebula_load": true,
148 "load_path": null
149 }
150 [2023-09-12 00:26:13,150] [INFO] [config.py:967:print] optimizer_legacy_fusion ...... False
151 [2023-09-12 00:26:13,151] [INFO] [config.py:967:print] optimizer_name ............... None
152 [2023-09-12 00:26:13,151] [INFO] [config.py:967:print] optimizer_params ............. None
153 [2023-09-12 00:26:13,151] [INFO] [config.py:967:print] pipeline ..................... {'stages': 'auto', 'partition': 'best', 'seed_layers': False, 'activation_checkpoint_interval': 0}
154 [2023-09-12 00:26:13,151] [INFO] [config.py:967:print] pld_enabled .................. False
155 [2023-09-12 00:26:13,151] [INFO] [config.py:967:print] pld_params ................... False
156 [2023-09-12 00:26:13,152] [INFO] [config.py:967:print] prescale_gradients ........... False
157 [2023-09-12 00:26:13,152] [INFO] [config.py:967:print] scheduler_name ............... None
158 [2023-09-12 00:26:13,152] [INFO] [config.py:967:print] scheduler_params ............. None
159 [2023-09-12 00:26:13,152] [INFO] [config.py:967:print] sparse_attention ............. None
160 [2023-09-12 00:26:13,153] [INFO] [config.py:967:print] sparse_gradients_enabled ..... False
161 [2023-09-12 00:26:13,153] [INFO] [config.py:967:print] steps_per_print .............. 10
162 [2023-09-12 00:26:13,153] [INFO] [config.py:967:print] train_batch_size ............. 64
163 [2023-09-12 00:26:13,154] [INFO] [config.py:967:print] train_micro_batch_size_per_gpu 1
164 [2023-09-12 00:26:13,154] [INFO] [config.py:967:print] use_node_local_storage ....... False
165 [2023-09-12 00:26:13,154] [INFO] [config.py:967:print] wall_clock_breakdown ......... False
166 [2023-09-12 00:26:13,155] [INFO] [config.py:967:print] world_size ................... 4
167 [2023-09-12 00:26:13,155] [INFO] [config.py:967:print] zero_allow_untested_optimizer False
168 [2023-09-12 00:26:13,155] [INFO] [config.py:967:print] zero_config .................. stage=3 contiguous_gradients=True reduce_scatter=True reduce_bucket_size=500,000,000 allgather_partitions=True allgather_bucket_size=500,000,000 overlap_comm=True load_from_fp32_weights=True elastic_checkpoint=False offload_param=DeepSpeedZeroOffloadParamConfig(device='none', nvme_path=None, buffer_count=5, buffer_size=100,000,000, max_in_cpu=1,000,000,000, pin_memory=False) offload_optimizer=DeepSpeedZeroOffloadOptimizerConfig(device='none', nvme_path=None, buffer_count=4, pin_memory=False, pipeline=False, pipeline_read=False, pipeline_write=False, fast_init=False) sub_group_size=1,000,000,000 cpu_offload_param=None cpu_offload_use_pin_memory=None cpu_offload=None prefetch_bucket_size=30000000 param_persistence_threshold=10000 model_persistence_threshold=sys.maxsize max_live_parameters=30000000 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=False pipeline_loading_checkpoint=False override_module_apply=True
169 [2023-09-12 00:26:13,155] [INFO] [config.py:967:print] zero_enabled ................. True
170 [2023-09-12 00:26:13,155] [INFO] [config.py:967:print] zero_force_ds_cpu_optimizer .. True
171 [2023-09-12 00:26:13,156] [INFO] [config.py:967:print] zero_optimization_stage ...... 3
172 [2023-09-12 00:26:13,156] [INFO] [config.py:953:print_user_config] json = {
173 "train_batch_size": 64,
174 "train_micro_batch_size_per_gpu": 1,
175 "steps_per_print": 10,
176 "zero_optimization": {
177 "stage": 3,
178 "offload_param": {
179 "device": "none"
180 },
181 "offload_optimizer": {
182 "device": "none"
183 },
184 "stage3_param_persistence_threshold": 1.000000e+04,
185 "stage3_max_live_parameters": 3.000000e+07,
186 "stage3_prefetch_bucket_size": 3.000000e+07,
187 "memory_efficient_linear": false
188 },
189 "fp16": {
190 "enabled": true,
191 "loss_scale_window": 100
192 },
193 "gradient_clipping": 1.0,
194 "prescale_gradients": false,
195 "wall_clock_breakdown": false,
196 "hybrid_engine": {
197 "enabled": false,
198 "max_out_tokens": 1.250000e+03,
199 "inference_tp_size": 1,
200 "release_inference_cache": false,
201 "pin_parameters": true,
202 "tp_gather_partition_size": 8
203 },
204 "activation_checkpointing": {
205 "partition_activations": true,
206 "cpu_checkpointing": true,
207 "contiguous_memory_optimization": true,
208 "number_checkpoints": 5
209 }
210 }

Start by reproducing DeepSpeed-Chat supervised fine-tuning with Pythia-6.9b, LoRA rank 128, and the stated four RTX6000 setup, comparing memory after initialization and the first forward pass. Inspect the ZeRO Stage 3 path in deepspeed/runtime/zero/stage3.py and the logged configuration in config.py. Done means the cause of the Stage 3 increase is confirmed and the expected memory behavior is verified.