Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

CUDA out of memory error on training #21

Open
AftabHussain opened this issue Apr 19, 2022 · 3 comments
Open

CUDA out of memory error on training #21

AftabHussain opened this issue Apr 19, 2022 · 3 comments

Comments

@AftabHussain
Copy link

AftabHussain commented Apr 19, 2022 •
edited

I was trying to train Polycoder using the preconfigured dataset, from the checkpoint checkpoints-2-7B, I used the following command as per the instructions in the repo (only changing the configs as appropriate):

sudo python ./deepy.py train.py -d configs 2-7B.yml local_setup.yml

which gave the following error:

RuntimeError: CUDA out of memory. Tried to allocate 1.86 GiB (GPU 0; 23.70 GiB total capacity; 20.49 GiB already allocated; 1.74 GiB free; 20.50 GiB reserved in total by PyTorch)

Interestingly, the full 25 Gigs of our GPU is free, as per nvidia-smi.

I tried updating the batch size, and the the only location I found to update batch size in the config files was train_micro_batch_size_per_gpu: 8, in 2-7B.yml.

It was 8, I changed it to 4, and then also to 1, but in both cases got the same error.

I am running all this in docker, as per the containerized setup instructions.

Appreciate any help!
@AftabHussain
Copy link
Author

AftabHussain commented Apr 19, 2022 •
edited

I also updated train_batch_size by updating gradient_accumulation_steps to 1, to give me a train_batch_size of 1 (where train_batch_size = micro_batch_per_gpu * gradient_acc_step * world_size) but still got the same error. Here is the full log:

mchorse@de65c7f6f3c6:/gpt-neox$ sudo python ./deepy.py train.py -d configs 2-7B.yml  local_setup.yml  
NeoXArgs.from_ymls() ['configs/2-7B.yml', 'configs/local_setup.yml']
INFO:root:NeoXArgs.calculate_derived() Total number of GPUs determined to be: 1
-------------------- arguments --------------------
  attention_config ................ ['global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global', 'global']updated
  attention_dropout ............... 0...........................updated
  batch_size ...................... 1...........................updated
  bias_gelu_fusion ................ True........................updated
  checkpoint_activations .......... True........................updated
  clip_grad ....................... 1.0.........................updated
  config_files .................... {'2-7B.yml': '# GPT-2 pretraining setup\n{\n   # parallelism settings ( you will want to change these based on your cluster setup, ideally scheduling pipeline stages\n   # across the node boundaries )\n   "pipe-parallel-size": 1,\n   "model-parallel-size": 1,\n\n   # model settings\n   "num-layers": 32,\n   "hidden-size": 2560,\n   "num-attention-heads": 32,\n   "seq-length": 2048,\n   "max-position-embeddings": 2048,\n   "norm": "layernorm",\n   "pos-emb": "rotary",\n   "no-weight-tying": true,\n\n   # these should provide some speedup but takes a while to build, set to true if desired\n   "scaled-upper-triang-masked-softmax-fusion": true,\n   "bias-gelu-fusion": true,\n\n   # optimizer settings\n   "zero_allow_untested_optimizer": true,\n   "optimizer": {\n     "type": "adam",\n     "params": {\n       "lr": 0.00016,\n       "betas": [0.9, 0.999],\n       "eps": 1.0e-8,\n     }\n   },\n   "zero_optimization": {\n    "stage": 1,\n    "allgather_partitions": True,\n    "allgather_bucket_size": 500000000,\n    "overlap_comm": True,\n    "reduce_scatter": True,\n    "reduce_bucket_size": 500000000,\n    "contiguous_gradients": True,\n    "cpu_offload": False\n  },\n\n   # batch / data settings # original batch size per gpu was 8, and gradient accumulation steps was 4\n   "train_micro_batch_size_per_gpu": 1,\n   "gradient_accumulation_steps": 1,\n   "data-impl": "mmap",\n   "split": "989,10,1",\n\n   # activation checkpointing\n   "checkpoint-activations": true,\n   "checkpoint-num-layers": 1,\n   "partition-activations": true,\n   "synchronize-each-layer": true,\n\n   # regularization\n   "gradient_clipping": 1.0,\n   "weight-decay": 0,\n   "hidden-dropout": 0,\n   "attention-dropout": 0,\n\n   # precision settings\n   "fp16": { \n     "fp16": true,\n     "enabled": true,\n     "loss_scale": 0,\n     "initial_scale_power": 16,\n     "loss_scale_window": 1000,\n     "hysteresis": 2,\n     "min_loss_scale": 1\n   },\n\n   # misc. training settings\n   "train-iters": 160000,\n   "lr-decay-iters": 160000,\n   "distributed-backend": "nccl",\n   "lr-decay-style": "cosine",\n   "warmup": 0.01,\n   "save-interval": 1000,\n   "eval-interval": 1000,\n   "eval-iters": 10,\n\n   # logging\n   "log-interval": 100,\n   "steps_per_print": 10,\n   "keep-last-n-checkpoints": 1,\n   "wall_clock_breakdown": true,\n}\n', 'local_setup.yml': '# Suggested data paths when using GPT-NeoX locally\n{\n  "data-path": "./data/enron/enron_text_document",\n  \n  # or for weighted datasets: \n  # "train-data-paths": ["data/enron/enron_text_document", "data/enron/enron_text_document"],\n  # "test-data-paths": ["data/enron/enron_text_document", "data/enron/enron_text_document"],\n  # "valid-data-paths": ["data/enron/enron_text_document", "data/enron/enron_text_document"],\n  # "train-data-weights": [1., 2.],\n  # "test-data-weights": [2., 1.],\n  # "valid-data-weights": [0.5, 0.4],\n\n  # If weight_by_num_documents is True, Builds dataset weights from a multinomial distribution over groups of data according to the number of documents in each group. \n  # WARNING: setting this to True will override any user provided weights\n  # "weight_by_num_documents": false,\n  # "weighted_sampler_alpha": 0.3,\n\n  "vocab-file": "data/code-vocab.json",\n  "merge-file": "data/code-merges.txt",\n\n  "save": "checkpoints",\n  "load": "checkpoints",\n  "checkpoint_validation_with_forward_pass": False,\n  \n  "tensorboard-dir": "tensorboard",\n  "log-dir": "logs",\n  "use_wandb": True,\n  "wandb_host": "https://api.wandb.ai",\n  "wandb_project": "neox"\n}\n'}updated
  data_impl ....................... mmap........................updated
  data_path ....................... ./data/enron/enron_text_documentupdated
  dynamic_loss_scale .............. True........................updated
  eval_iters ...................... 10..........................updated
  fp16 ............................ {'fp16': True, 'enabled': True, 'loss_scale': 0, 'initial_scale_power': 16, 'loss_scale_window': 1000, 'hysteresis': 2, 'min_loss_scale': 1}updated
  gas ............................. 1...........................updated
  global_num_gpus ................. 1...........................updated
  gradient_clipping ............... 1.0.........................updated
  hidden_dropout .................. 0...........................updated
  hidden_size ..................... 2560........................updated
  is_pipe_parallel ................ True........................updated
  keep_last_n_checkpoints ......... 1...........................updated
  load ............................ checkpoints.................updated
  log_dir ......................... logs........................updated
  log_interval .................... 100.........................updated
  lr .............................. 0.00016.....................updated
  lr_decay_iters .................. 160000......................updated
  lr_decay_style .................. cosine......................updated
  max_position_embeddings ......... 2048........................updated
  merge_file ...................... data/code-merges.txt........updated
  no_weight_tying ................. True........................updated
  num_attention_heads ............. 32..........................updated
  num_layers ...................... 32..........................updated
  optimizer ....................... {'type': 'adam', 'params': {'lr': 0.00016, 'betas': [0.9, 0.999], 'eps': 1e-08}}updated
  partition_activations ........... True........................updated
  pipe_parallel_size .............. 1...........................updated
  pos_emb ......................... rotary......................updated
  precision ....................... fp16........................updated
  save ............................ checkpoints.................updated
  save_interval ................... 1000........................updated
  scaled_upper_triang_masked_softmax_fusion  True...............updated
  seq_length ...................... 2048........................updated
  sparsity_config ................. {}..........................updated
  split ........................... 989,10,1....................updated
  synchronize_each_layer .......... True........................updated
  tensorboard_dir ................. tensorboard.................updated
  train_batch_size ................ 1...........................updated
  train_iters ..................... 160000......................updated
  train_micro_batch_size_per_gpu .. 1...........................updated
  use_wandb ....................... True........................updated
  user_script ..................... train.py....................updated
  vocab_file ...................... data/code-vocab.json........updated
  wall_clock_breakdown ............ True........................updated
  wandb_group ..................... DeMkeMCDWQ8LA6TGSz7jtQ_2b3bycopupdated
  weight_decay .................... 0...........................updated
  zero_allgather_bucket_size ...... 500000000...................updated
  zero_allow_untested_optimizer ... True........................updated
  zero_contiguous_gradients ....... True........................updated
  zero_optimization ............... {'stage': 1, 'allgather_partitions': True, 'allgather_bucket_size': 500000000, 'overlap_comm': True, 'reduce_scatter': True, 'reduce_bucket_size': 500000000, 'contiguous_gradients': True, 'cpu_offload': False}updated
  zero_reduce_bucket_size ......... 500000000...................updated
  zero_reduce_scatter ............. True........................updated
  zero_stage ...................... 1...........................updated
  activation ...................... gelu........................default
  adlr_autoresume ................. False.......................default
  adlr_autoresume_interval ........ 1000........................default
  amp ............................. None........................default
  apply_query_key_layer_scaling ... False.......................default
  attention_softmax_in_fp32 ....... False.......................default
  bias_dropout_fusion ............. False.......................default
  char_level_ppl .................. False.......................default
  checkpoint_in_cpu ............... False.......................default
  checkpoint_num_layers ........... 1...........................default
  checkpoint_validation_with_forward_pass  False................default
  contiguous_checkpointing ........ False.......................default
  deepscale ....................... False.......................default
  deepscale_config ................ None........................default
  deepspeed ....................... True........................default
  deepspeed_activation_checkpointing  True......................default
  deepspeed_mpi ................... False.......................default
  detect_nvlink_pairs ............. False.......................default
  distributed_backend ............. nccl........................default
  do_test ......................... None........................default
  do_train ........................ None........................default
  do_valid ........................ None........................default
  dump_state ...................... False.......................default
  eod_mask_loss ................... False.......................default
  eval_interval ................... 1000........................default
  eval_results_prefix ............. ............................default
  eval_tasks ...................... None........................default
  exclude ......................... None........................default
  exit_interval ................... None........................default
  finetune ........................ False.......................default
  flops_profiler .................. None........................default
  fp16_lm_cross_entropy ........... False.......................default
  fp32_allreduce .................. False.......................default
  git_hash ........................ 98683ae.....................default
  gmlp_attn_dim ................... 64..........................default
  gpt_j_residual .................. False.......................default
  gradient_accumulation_steps ..... 1...........................default
  gradient_noise_scale_cpu_offload  False.......................default
  gradient_noise_scale_n_batches .. 5...........................default
  gradient_predivide_factor ....... 1.0.........................default
  hostfile ........................ None........................default
  hysteresis ...................... 2...........................default
  include ......................... None........................default
  init_method ..................... normal......................default
  init_method_std ................. 0.02........................default
  iteration ....................... None........................default
  launcher ........................ pdsh........................default
  layernorm_epsilon ............... 1e-05.......................default
  lazy_mpu_init ................... False.......................default
  local_rank ...................... None........................default
  log_grad_norm ................... False.......................default
  log_gradient_noise_scale ........ False.......................default
  log_optimizer_states ............ False.......................default
  log_param_norm .................. False.......................default
  loss_scale ...................... None........................default
  loss_scale_window ............... 1000.0......................default
  make_vocab_size_divisible_by .... 128.........................default
  master_addr ..................... None........................default
  master_port ..................... 29500.......................default
  maximum_tokens .................. 64..........................default
  min_lr .......................... 0.0.........................default
  min_scale ....................... 1.0.........................default
  mmap_warmup ..................... False.......................default
  model_parallel_size ............. 1...........................default
  no_load_optim ................... False.......................default
  no_load_rng ..................... False.......................default
  no_save_optim ................... False.......................default
  no_save_rng ..................... False.......................default
  norm ............................ layernorm...................default
  num_gpus ........................ None........................default
  num_nodes ....................... -1..........................default
  num_samples ..................... 0...........................default
  num_unique_layers ............... None........................default
  num_workers ..................... 2...........................default
  onnx_safe ....................... False.......................default
  optimizer_type .................. adam........................default
  output_layer_init_method ........ scaled_normal...............default
  output_layer_parallelism ........ row.........................default
  override_lr_scheduler ........... False.......................default
  padded_vocab_size ............... None........................default
  param_sharing_style ............. grouped.....................default
  pipe_partition_method ........... type:transformer|mlp........default
  prescale_gradients .............. False.......................default
  profile_backward ................ False.......................default
  rank ............................ None........................default
  recompute ....................... False.......................default
  rms_norm_epsilon ................ 1e-08.......................default
  rotary_emb_base ................. 10000.......................default
  rotary_pct ...................... 1.0.........................default
  rpe_max_distance ................ 128.........................default
  rpe_num_buckets ................. 32..........................default
  sample_input_file ............... None........................default
  sample_output_file .............. None........................default
  scaled_masked_softmax_fusion .... False.......................default
  scalenorm_epsilon ............... 1e-08.......................default
  scheduler ....................... None........................default
  seed ............................ 1234........................default
  short_seq_prob .................. 0.1.........................default
  soft_prompt_tuning .............. None........................default
  sparse_gradients ................ False.......................default
  steps_per_print ................. 10..........................default
  temperature ..................... 0.0.........................default
  test_data_paths ................. None........................default
  test_data_weights ............... None........................default
  text_gen_type ................... None........................default
  tokenizer_type .................. GPT2BPETokenizer............default
  top_k ........................... 0...........................default
  top_p ........................... 0.0.........................default
  train_data_paths ................ None........................default
  train_data_weights .............. None........................default
  use_bnb_optimizer ............... False.......................default
  use_checkpoint_lr_scheduler ..... False.......................default
  use_cpu_initialization .......... False.......................default
  valid_data_paths ................ None........................default
  valid_data_weights .............. None........................default
  wandb_host ...................... https://api.wandb.ai........default
  wandb_project ................... neox........................default
  wandb_team ...................... None........................default
  warmup .......................... 0.01........................default
  weight_by_num_documents ......... False.......................default
  weighted_sampler_alpha .......... 0.3.........................default
  world_size ...................... None........................default
---------------- end of arguments ----------------
[2022-04-19 20:24:23,257] [WARNING] [runner.py:126:fetch_hostfile] Unable to find hostfile, will proceed with training with local resources only.
[2022-04-19 20:24:23,257] [INFO] [runner.py:366:main] cmd = /usr/bin/python -u -m deepspeed.launcher.launch --world_info=eyJsb2NhbGhvc3QiOiBbMF19 --master_addr=127.0.0.1 --master_port=29500 train.py --deepspeed_config {"train_batch_size": 1, "train_micro_batch_size_per_gpu": 1, "optimizer": {"type": "adam", "params": {"lr": 0.00016, "betas": [0.9, 0.999], "eps": 1e-08}}, "fp16": {"fp16": true, "enabled": true, "loss_scale": 0, "initial_scale_power": 16, "loss_scale_window": 1000, "hysteresis": 2, "min_loss_scale": 1}, "gradient_clipping": 1.0, "zero_optimization": {"stage": 1, "allgather_partitions": true, "allgather_bucket_size": 500000000, "overlap_comm": true, "reduce_scatter": true, "reduce_bucket_size": 500000000, "contiguous_gradients": true, "cpu_offload": false}, "wall_clock_breakdown": true, "zero_allow_untested_optimizer": true} --megatron_config {"train_batch_size": 1, "train_micro_batch_size_per_gpu": 1, "optimizer": {"type": "adam", "params": {"lr": 0.00016, "betas": [0.9, 0.999], "eps": 1e-08}}, "fp16": {"fp16": true, "enabled": true, "loss_scale": 0, "initial_scale_power": 16, "loss_scale_window": 1000, "hysteresis": 2, "min_loss_scale": 1}, "gradient_clipping": 1.0, "zero_optimization": {"stage": 1, "allgather_partitions": true, "allgather_bucket_size": 500000000, "overlap_comm": true, "reduce_scatter": true, "reduce_bucket_size": 500000000, "contiguous_gradients": true, "cpu_offload": false}, "wall_clock_breakdown": true, "zero_allow_untested_optimizer": true, "precision": "fp16", "num_layers": 32, "hidden_size": 2560, "num_attention_heads": 32, "seq_length": 2048, "max_position_embeddings": 2048, "pos_emb": "rotary", "no_weight_tying": true, "attention_config": ["global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global"], "sparsity_config": {}, "scaled_upper_triang_masked_softmax_fusion": true, "bias_gelu_fusion": true, "lr_decay_style": "cosine", "lr_decay_iters": 160000, "zero_stage": 1, "zero_reduce_scatter": true, "zero_contiguous_gradients": true, "zero_reduce_bucket_size": 500000000, "zero_allgather_bucket_size": 500000000, "lr": 0.00016, "data_path": "./data/enron/enron_text_document", "data_impl": "mmap", "save": "checkpoints", "config_files": {"2-7B.yml": "# GPT-2 pretraining setup\n{\n   # parallelism settings ( you will want to change these based on your cluster setup, ideally scheduling pipeline stages\n   # across the node boundaries )\n   \"pipe-parallel-size\": 1,\n   \"model-parallel-size\": 1,\n\n   # model settings\n   \"num-layers\": 32,\n   \"hidden-size\": 2560,\n   \"num-attention-heads\": 32,\n   \"seq-length\": 2048,\n   \"max-position-embeddings\": 2048,\n   \"norm\": \"layernorm\",\n   \"pos-emb\": \"rotary\",\n   \"no-weight-tying\": true,\n\n   # these should provide some speedup but takes a while to build, set to true if desired\n   \"scaled-upper-triang-masked-softmax-fusion\": true,\n   \"bias-gelu-fusion\": true,\n\n   # optimizer settings\n   \"zero_allow_untested_optimizer\": true,\n   \"optimizer\": {\n     \"type\": \"adam\",\n     \"params\": {\n       \"lr\": 0.00016,\n       \"betas\": [0.9, 0.999],\n       \"eps\": 1.0e-8,\n     }\n   },\n   \"zero_optimization\": {\n    \"stage\": 1,\n    \"allgather_partitions\": True,\n    \"allgather_bucket_size\": 500000000,\n    \"overlap_comm\": True,\n    \"reduce_scatter\": True,\n    \"reduce_bucket_size\": 500000000,\n    \"contiguous_gradients\": True,\n    \"cpu_offload\": False\n  },\n\n   # batch / data settings # original batch size per gpu was 8, and gradient accumulation steps was 4\n   \"train_micro_batch_size_per_gpu\": 1,\n   \"gradient_accumulation_steps\": 1,\n   \"data-impl\": \"mmap\",\n   \"split\": \"989,10,1\",\n\n   # activation checkpointing\n   \"checkpoint-activations\": true,\n   \"checkpoint-num-layers\": 1,\n   \"partition-activations\": true,\n   \"synchronize-each-layer\": true,\n\n   # regularization\n   \"gradient_clipping\": 1.0,\n   \"weight-decay\": 0,\n   \"hidden-dropout\": 0,\n   \"attention-dropout\": 0,\n\n   # precision settings\n   \"fp16\": { \n     \"fp16\": true,\n     \"enabled\": true,\n     \"loss_scale\": 0,\n     \"initial_scale_power\": 16,\n     \"loss_scale_window\": 1000,\n     \"hysteresis\": 2,\n     \"min_loss_scale\": 1\n   },\n\n   # misc. training settings\n   \"train-iters\": 160000,\n   \"lr-decay-iters\": 160000,\n   \"distributed-backend\": \"nccl\",\n   \"lr-decay-style\": \"cosine\",\n   \"warmup\": 0.01,\n   \"save-interval\": 1000,\n   \"eval-interval\": 1000,\n   \"eval-iters\": 10,\n\n   # logging\n   \"log-interval\": 100,\n   \"steps_per_print\": 10,\n   \"keep-last-n-checkpoints\": 1,\n   \"wall_clock_breakdown\": true,\n}\n", "local_setup.yml": "# Suggested data paths when using GPT-NeoX locally\n{\n  \"data-path\": \"./data/enron/enron_text_document\",\n  \n  # or for weighted datasets: \n  # \"train-data-paths\": [\"data/enron/enron_text_document\", \"data/enron/enron_text_document\"],\n  # \"test-data-paths\": [\"data/enron/enron_text_document\", \"data/enron/enron_text_document\"],\n  # \"valid-data-paths\": [\"data/enron/enron_text_document\", \"data/enron/enron_text_document\"],\n  # \"train-data-weights\": [1., 2.],\n  # \"test-data-weights\": [2., 1.],\n  # \"valid-data-weights\": [0.5, 0.4],\n\n  # If weight_by_num_documents is True, Builds dataset weights from a multinomial distribution over groups of data according to the number of documents in each group. \n  # WARNING: setting this to True will override any user provided weights\n  # \"weight_by_num_documents\": false,\n  # \"weighted_sampler_alpha\": 0.3,\n\n  \"vocab-file\": \"data/code-vocab.json\",\n  \"merge-file\": \"data/code-merges.txt\",\n\n  \"save\": \"checkpoints\",\n  \"load\": \"checkpoints\",\n  \"checkpoint_validation_with_forward_pass\": False,\n  \n  \"tensorboard-dir\": \"tensorboard\",\n  \"log-dir\": \"logs\",\n  \"use_wandb\": True,\n  \"wandb_host\": \"https://api.wandb.ai\",\n  \"wandb_project\": \"neox\"\n}\n"}, "load": "checkpoints", "save_interval": 1000, "batch_size": 1, "train_iters": 160000, "eval_iters": 10, "keep_last_n_checkpoints": 1, "split": "989,10,1", "vocab_file": "data/code-vocab.json", "merge_file": "data/code-merges.txt", "attention_dropout": 0, "hidden_dropout": 0, "weight_decay": 0, "checkpoint_activations": true, "synchronize_each_layer": true, "partition_activations": true, "gas": 1, "clip_grad": 1.0, "dynamic_loss_scale": true, "pipe_parallel_size": 1, "is_pipe_parallel": true, "use_wandb": true, "wandb_group": "DeMkeMCDWQ8LA6TGSz7jtQ_2b3bycop", "log_dir": "logs", "tensorboard_dir": "tensorboard", "log_interval": 100, "user_script": "train.py", "global_num_gpus": 1}
[2022-04-19 20:24:24,084] [INFO] [launch.py:82:main] WORLD INFO DICT: {'localhost': [0]}
[2022-04-19 20:24:24,085] [INFO] [launch.py:88:main] nnodes=1, num_local_procs=1, node_rank=0
[2022-04-19 20:24:24,085] [INFO] [launch.py:103:main] global_rank_mapping=defaultdict(<class 'list'>, {'localhost': [0]})
[2022-04-19 20:24:24,085] [INFO] [launch.py:104:main] dist_world_size=1
[2022-04-19 20:24:24,085] [INFO] [launch.py:112:main] Setting CUDA_VISIBLE_DEVICES=0
NeoXArgs.configure_distributed_args() using world size: 1 and model-parallel size: 1 
> building GPT2BPETokenizer tokenizer ...
 > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304)
WARNING: TensorBoard writing requested but is not available (are you using PyTorch 1.1.0 or later and do you have tensorboard installed?), no TensorBoard logs will be written.
> initializing torch distributed ...
[2022-04-19 20:24:26,400] [INFO] [distributed.py:46:init_distributed] Initializing torch distributed with backend: nccl
> initializing model parallel with size 1
MPU DP: [0]
MPU PP: [0]
MPU MP: [0]
> setting random seeds to 1234 ...
[2022-04-19 20:24:26,401] [INFO] [checkpointing.py:223:model_parallel_cuda_manual_seed] > initializing model parallel cuda seeds on global rank 0, model parallel rank 0, and data parallel rank 0 with model parallel seed: 3952 and data parallel seed: 1234
make: Entering directory '/gpt-neox/megatron/data'
make: Nothing to be done for 'default'.
make: Leaving directory '/gpt-neox/megatron/data'
building GPT2 model ...
SEED_LAYERS=False BASE_SEED=1234 SEED_FN=None
Using topology: {ProcessCoord(pipe=0, data=0, model=0): 0}
[2022-04-19 20:24:29,552] [INFO] [module.py:363:_partition_layers] Partitioning pipeline stages with method type:transformer|mlp
stage=0 layers=37
     0: EmbeddingPipe
     1: _pre_transformer_block
     2: ParallelTransformerLayerPipe
     3: ParallelTransformerLayerPipe
     4: ParallelTransformerLayerPipe
     5: ParallelTransformerLayerPipe
     6: ParallelTransformerLayerPipe
     7: ParallelTransformerLayerPipe
     8: ParallelTransformerLayerPipe
     9: ParallelTransformerLayerPipe
    10: ParallelTransformerLayerPipe
    11: ParallelTransformerLayerPipe
    12: ParallelTransformerLayerPipe
    13: ParallelTransformerLayerPipe
    14: ParallelTransformerLayerPipe
    15: ParallelTransformerLayerPipe
    16: ParallelTransformerLayerPipe
    17: ParallelTransformerLayerPipe
    18: ParallelTransformerLayerPipe
    19: ParallelTransformerLayerPipe
    20: ParallelTransformerLayerPipe
    21: ParallelTransformerLayerPipe
    22: ParallelTransformerLayerPipe
    23: ParallelTransformerLayerPipe
    24: ParallelTransformerLayerPipe
    25: ParallelTransformerLayerPipe
    26: ParallelTransformerLayerPipe
    27: ParallelTransformerLayerPipe
    28: ParallelTransformerLayerPipe
    29: ParallelTransformerLayerPipe
    30: ParallelTransformerLayerPipe
    31: ParallelTransformerLayerPipe
    32: ParallelTransformerLayerPipe
    33: ParallelTransformerLayerPipe
    34: _post_transformer_block
    35: NormPipe
    36: ParallelLinearPipe
  loss: partial
Configuring Optimizer type: adam with params: {'lr': 0.00016, 'betas': [0.9, 0.999], 'eps': 1e-08}
> learning rate decay style: cosine
DeepSpeed is enabled.
[2022-04-19 20:24:29,614] [INFO] [logging.py:60:log_dist] [Rank 0] DeepSpeed info: version=0.3.15+eb7f5cf, git-hash=eb7f5cf, git-branch=main
[2022-04-19 20:24:29,614] [WARNING] [config.py:77:_sanity_check] DeepSpeedConfig: cpu_offload is deprecated. Please use offload_optimizer.
[2022-04-19 20:24:29,683] [INFO] [engine.py:654:_configure_optimizer] Removing param_group that has no 'params' in the client Optimizer
[2022-04-19 20:24:29,683] [INFO] [engine.py:659:_configure_optimizer] Using client Optimizer as basic optimizer
[2022-04-19 20:24:29,683] [INFO] [engine.py:668:_configure_optimizer] DeepSpeed Basic Optimizer = FusedAdam
Checking ZeRO support for optimizer=FusedAdam type=<class 'apex.optimizers.fused_adam.FusedAdam'>
[2022-04-19 20:24:29,683] [INFO] [logging.py:60:log_dist] [Rank 0] Creating fp16 ZeRO stage 1 optimizer
Using /root/.cache/torch_extensions as PyTorch extensions root...
Emitting ninja build file /root/.cache/torch_extensions/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: 0.7416722774505615 seconds
[2022-04-19 20:24:30,427] [INFO] [stage1.py:160:__init__] ZeRO Elastic Checkpoint = True
[2022-04-19 20:24:30,427] [INFO] [logging.py:60:log_dist] [Rank 0] Using default max_elements_per_comm 500000000
[2022-04-19 20:24:30,428] [INFO] [logging.py:60:log_dist] [Rank 0] Total number of elements in model: 2775208960, max elements per com: 500000000
[2022-04-19 20:24:30,428] [INFO] [logging.py:60:log_dist] [Rank 0] sub_partition_count: 6, sub_partition_size: 500000000, padding: 224791040
[2022-04-19 20:24:30,428] [INFO] [logging.py:60:log_dist] [Rank 0] number of elements with padding: 2775208960 + 224791040 = 3000000000
[2022-04-19 20:24:30,445] [INFO] [stage1.py:375:get_data_parallel_sub_partitions] **** partition info:
[2022-04-19 20:24:30,445] [INFO] [stage1.py:376:get_data_parallel_sub_partitions] 	 total_num_elements=3000000000
[2022-04-19 20:24:30,445] [INFO] [stage1.py:377:get_data_parallel_sub_partitions] 	 world_size=1
[2022-04-19 20:24:30,446] [INFO] [stage1.py:378:get_data_parallel_sub_partitions] 	 max_elements_per_comm=500000000
[2022-04-19 20:24:30,446] [INFO] [stage1.py:379:get_data_parallel_sub_partitions] 	 sub_partition_size=500000000
[2022-04-19 20:24:30,446] [INFO] [stage1.py:380:get_data_parallel_sub_partitions] 	 num_sub_partitions=6
[2022-04-19 20:24:30,446] [INFO] [stage1.py:381:get_data_parallel_sub_partitions] 	 num_comm_intervals=6
[2022-04-19 20:24:30,446] [INFO] [stage1.py:382:get_data_parallel_sub_partitions] ****
Traceback (most recent call last):
  File "train.py", line 27, in <module>
    pretrain(neox_args=neox_args)
  File "/gpt-neox/megatron/training.py", line 86, in pretrain
    model, optimizer, lr_scheduler = setup_model_and_optimizer(
  File "/gpt-neox/megatron/training.py", line 423, in setup_model_and_optimizer
    model, optimizer, _, lr_scheduler = deepspeed.initialize(
  File "/usr/local/lib/python3.8/dist-packages/deepspeed/__init__.py", line 128, in initialize
    engine = PipelineEngine(args=args,
  File "/usr/local/lib/python3.8/dist-packages/deepspeed/runtime/pipe/engine.py", line 60, in __init__
    super().__init__(*super_args, **super_kwargs)
  File "/usr/local/lib/python3.8/dist-packages/deepspeed/runtime/engine.py", line 192, in __init__
    self._configure_optimizer(optimizer, model_parameters)
  File "/usr/local/lib/python3.8/dist-packages/deepspeed/runtime/engine.py", line 681, in _configure_optimizer
    self.optimizer = self._configure_zero_optimizer(basic_optimizer)
  File "/usr/local/lib/python3.8/dist-packages/deepspeed/runtime/engine.py", line 810, in _configure_zero_optimizer
    optimizer = FP16_DeepSpeedZeroOptimizer_Stage1(
  File "/usr/local/lib/python3.8/dist-packages/deepspeed/runtime/zero/stage1.py", line 311, in __init__
    self._initialize_optimizer_states()
  File "/usr/local/lib/python3.8/dist-packages/deepspeed/runtime/zero/stage1.py", line 316, in _initialize_optimizer_states
    sub_partition_grad = torch.zeros(int(
RuntimeError: CUDA out of memory. Tried to allocate 1.86 GiB (GPU 0; 23.70 GiB total capacity; 20.49 GiB already allocated; 1.74 GiB free; 20.50 GiB reserved in total by PyTorch)
Killing subprocess 8346
Traceback (most recent call last):
  File "/usr/lib/python3.8/runpy.py", line 194, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/usr/lib/python3.8/runpy.py", line 87, in _run_code
    exec(code, run_globals)
  File "/usr/local/lib/python3.8/dist-packages/deepspeed/launcher/launch.py", line 179, in <module>
    main()
  File "/usr/local/lib/python3.8/dist-packages/deepspeed/launcher/launch.py", line 169, in main
    sigkill_handler(signal.SIGTERM, None)  # not coming back
  File "/usr/local/lib/python3.8/dist-packages/deepspeed/launcher/launch.py", line 147, in sigkill_handler
    raise subprocess.CalledProcessError(returncode=last_return_code, cmd=cmd)
subprocess.CalledProcessError: Command '['/usr/bin/python', '-u', 'train.py', '--local_rank=0', '--deepspeed_config', '{"train_batch_size": 1, "train_micro_batch_size_per_gpu": 1, "optimizer": {"type": "adam", "params": {"lr": 0.00016, "betas": [0.9, 0.999], "eps": 1e-08}}, "fp16": {"fp16": true, "enabled": true, "loss_scale": 0, "initial_scale_power": 16, "loss_scale_window": 1000, "hysteresis": 2, "min_loss_scale": 1}, "gradient_clipping": 1.0, "zero_optimization": {"stage": 1, "allgather_partitions": true, "allgather_bucket_size": 500000000, "overlap_comm": true, "reduce_scatter": true, "reduce_bucket_size": 500000000, "contiguous_gradients": true, "cpu_offload": false}, "wall_clock_breakdown": true, "zero_allow_untested_optimizer": true}', '--megatron_config', '{"train_batch_size": 1, "train_micro_batch_size_per_gpu": 1, "optimizer": {"type": "adam", "params": {"lr": 0.00016, "betas": [0.9, 0.999], "eps": 1e-08}}, "fp16": {"fp16": true, "enabled": true, "loss_scale": 0, "initial_scale_power": 16, "loss_scale_window": 1000, "hysteresis": 2, "min_loss_scale": 1}, "gradient_clipping": 1.0, "zero_optimization": {"stage": 1, "allgather_partitions": true, "allgather_bucket_size": 500000000, "overlap_comm": true, "reduce_scatter": true, "reduce_bucket_size": 500000000, "contiguous_gradients": true, "cpu_offload": false}, "wall_clock_breakdown": true, "zero_allow_untested_optimizer": true, "precision": "fp16", "num_layers": 32, "hidden_size": 2560, "num_attention_heads": 32, "seq_length": 2048, "max_position_embeddings": 2048, "pos_emb": "rotary", "no_weight_tying": true, "attention_config": ["global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global", "global"], "sparsity_config": {}, "scaled_upper_triang_masked_softmax_fusion": true, "bias_gelu_fusion": true, "lr_decay_style": "cosine", "lr_decay_iters": 160000, "zero_stage": 1, "zero_reduce_scatter": true, "zero_contiguous_gradients": true, "zero_reduce_bucket_size": 500000000, "zero_allgather_bucket_size": 500000000, "lr": 0.00016, "data_path": "./data/enron/enron_text_document", "data_impl": "mmap", "save": "checkpoints", "config_files": {"2-7B.yml": "# GPT-2 pretraining setup\\n{\\n   # parallelism settings ( you will want to change these based on your cluster setup, ideally scheduling pipeline stages\\n   # across the node boundaries )\\n   \\"pipe-parallel-size\\": 1,\\n   \\"model-parallel-size\\": 1,\\n\\n   # model settings\\n   \\"num-layers\\": 32,\\n   \\"hidden-size\\": 2560,\\n   \\"num-attention-heads\\": 32,\\n   \\"seq-length\\": 2048,\\n   \\"max-position-embeddings\\": 2048,\\n   \\"norm\\": \\"layernorm\\",\\n   \\"pos-emb\\": \\"rotary\\",\\n   \\"no-weight-tying\\": true,\\n\\n   # these should provide some speedup but takes a while to build, set to true if desired\\n   \\"scaled-upper-triang-masked-softmax-fusion\\": true,\\n   \\"bias-gelu-fusion\\": true,\\n\\n   # optimizer settings\\n   \\"zero_allow_untested_optimizer\\": true,\\n   \\"optimizer\\": {\\n     \\"type\\": \\"adam\\",\\n     \\"params\\": {\\n       \\"lr\\": 0.00016,\\n       \\"betas\\": [0.9, 0.999],\\n       \\"eps\\": 1.0e-8,\\n     }\\n   },\\n   \\"zero_optimization\\": {\\n    \\"stage\\": 1,\\n    \\"allgather_partitions\\": True,\\n    \\"allgather_bucket_size\\": 500000000,\\n    \\"overlap_comm\\": True,\\n    \\"reduce_scatter\\": True,\\n    \\"reduce_bucket_size\\": 500000000,\\n    \\"contiguous_gradients\\": True,\\n    \\"cpu_offload\\": False\\n  },\\n\\n   # batch / data settings # original batch size per gpu was 8, and gradient accumulation steps was 4\\n   \\"train_micro_batch_size_per_gpu\\": 1,\\n   \\"gradient_accumulation_steps\\": 1,\\n   \\"data-impl\\": \\"mmap\\",\\n   \\"split\\": \\"989,10,1\\",\\n\\n   # activation checkpointing\\n   \\"checkpoint-activations\\": true,\\n   \\"checkpoint-num-layers\\": 1,\\n   \\"partition-activations\\": true,\\n   \\"synchronize-each-layer\\": true,\\n\\n   # regularization\\n   \\"gradient_clipping\\": 1.0,\\n   \\"weight-decay\\": 0,\\n   \\"hidden-dropout\\": 0,\\n   \\"attention-dropout\\": 0,\\n\\n   # precision settings\\n   \\"fp16\\": { \\n     \\"fp16\\": true,\\n     \\"enabled\\": true,\\n     \\"loss_scale\\": 0,\\n     \\"initial_scale_power\\": 16,\\n     \\"loss_scale_window\\": 1000,\\n     \\"hysteresis\\": 2,\\n     \\"min_loss_scale\\": 1\\n   },\\n\\n   # misc. training settings\\n   \\"train-iters\\": 160000,\\n   \\"lr-decay-iters\\": 160000,\\n   \\"distributed-backend\\": \\"nccl\\",\\n   \\"lr-decay-style\\": \\"cosine\\",\\n   \\"warmup\\": 0.01,\\n   \\"save-interval\\": 1000,\\n   \\"eval-interval\\": 1000,\\n   \\"eval-iters\\": 10,\\n\\n   # logging\\n   \\"log-interval\\": 100,\\n   \\"steps_per_print\\": 10,\\n   \\"keep-last-n-checkpoints\\": 1,\\n   \\"wall_clock_breakdown\\": true,\\n}\\n", "local_setup.yml": "# Suggested data paths when using GPT-NeoX locally\\n{\\n  \\"data-path\\": \\"./data/enron/enron_text_document\\",\\n  \\n  # or for weighted datasets: \\n  # \\"train-data-paths\\": [\\"data/enron/enron_text_document\\", \\"data/enron/enron_text_document\\"],\\n  # \\"test-data-paths\\": [\\"data/enron/enron_text_document\\", \\"data/enron/enron_text_document\\"],\\n  # \\"valid-data-paths\\": [\\"data/enron/enron_text_document\\", \\"data/enron/enron_text_document\\"],\\n  # \\"train-data-weights\\": [1., 2.],\\n  # \\"test-data-weights\\": [2., 1.],\\n  # \\"valid-data-weights\\": [0.5, 0.4],\\n\\n  # If weight_by_num_documents is True, Builds dataset weights from a multinomial distribution over groups of data according to the number of documents in each group. \\n  # WARNING: setting this to True will override any user provided weights\\n  # \\"weight_by_num_documents\\": false,\\n  # \\"weighted_sampler_alpha\\": 0.3,\\n\\n  \\"vocab-file\\": \\"data/code-vocab.json\\",\\n  \\"merge-file\\": \\"data/code-merges.txt\\",\\n\\n  \\"save\\": \\"checkpoints\\",\\n  \\"load\\": \\"checkpoints\\",\\n  \\"checkpoint_validation_with_forward_pass\\": False,\\n  \\n  \\"tensorboard-dir\\": \\"tensorboard\\",\\n  \\"log-dir\\": \\"logs\\",\\n  \\"use_wandb\\": True,\\n  \\"wandb_host\\": \\"https://api.wandb.ai\\",\\n  \\"wandb_project\\": \\"neox\\"\\n}\\n"}, "load": "checkpoints", "save_interval": 1000, "batch_size": 1, "train_iters": 160000, "eval_iters": 10, "keep_last_n_checkpoints": 1, "split": "989,10,1", "vocab_file": "data/code-vocab.json", "merge_file": "data/code-merges.txt", "attention_dropout": 0, "hidden_dropout": 0, "weight_decay": 0, "checkpoint_activations": true, "synchronize_each_layer": true, "partition_activations": true, "gas": 1, "clip_grad": 1.0, "dynamic_loss_scale": true, "pipe_parallel_size": 1, "is_pipe_parallel": true, "use_wandb": true, "wandb_group": "DeMkeMCDWQ8LA6TGSz7jtQ_2b3bycop", "log_dir": "logs", "tensorboard_dir": "tensorboard", "log_interval": 100, "user_script": "train.py", "global_num_gpus": 1}']' returned non-zero exit status 1.

@VHellendoorn
Copy link
Owner

Hi Aftab,

Thanks for submitting this issue. It took a while to debug; the main thing I have found so far is that I can run this just fine on 4 GPUs (with 48GB of memory, instead of 24GB), but not on 2 and 1. And to be clear, the batch size scales to the number of GPUs, so when I run on 4, I am also running with 4x the data compared to a single GPU. I have never seen this kind of behavior before; normally, when you scale the batch size by the number of GPUs like that, you can run just fine on any number of GPUs.

My main guess is that something about the model or optimizer state is getting distributed across the GPUs. That way, when there are more GPUs, the memory cost per GPU is lower. I have tried disabling every setting I could think of in the config, but could not make this behavior go away, so I am not sure if this is something I can remedy. It may be worth posting to the original GPT-NeoX repository directly. My only concern is that this could be a Docker issue, somehow, in which case it isn't on their end either; you may want to try this in a virtualenv instead, to test it without using Docker. I will see what else I can try, but for now this is the state of affairs: training the largest model requires an excessive amount of memory on a single GPU, presumably because the optimizer states alone take up more than 48GB.

-Vincent

@AftabHussain
Copy link
Author

Hi Vincent,

Thank you so much for your time. Indeed it is very interesting why Pytorch always reserves such a huge chunk of memory, such that there is always less free space than the amount the script tries to allocate. I saw changing the hidden_size parameter changes the amount of memory the script tries to allocate, however, the free space availability is always lesser.

My Best Regards,
Aftab

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
2 participants
@VHellendoorn @AftabHussain