LibriSpeech Conformer Workload OOMs or NCCL Errors When Run With Multiple Trials

提供されている torchrun の再現手順から始め、次に trial の作成、モデルの初期化、パラメータの更新に関連する submission_runner.py、pytorch_nadamw_full_budget.py、input_pipeline.py を調査します。trial_1 が存在する状態で再開した場合と、trial 間でメモリおよびプロセスの状態を比較します。5 trial の LibriSpeech Conformer の実行が CUDA OOM なしで完了し、複数 trial の実行を再開しても NCCL broadcast でハングしなければ完了です。

We consistently observe an OOM error when running the one of the NAdamW baselines on LibriSpeech Conformer with multiple trials in PyTorch on 8 V100s with 16GB each. This is run for the external ruleset. The first trial will successfully run through, but any subsequent trial will OOM.

If we try to resume a multi-trial run, we will observe a NCCL error. This occurs even if we delete the trial_2 folder (but the trial_1 folder remains intact).

Description

As discussed above, we will observe OOM when running LibriSpeech Conformer with the NAdamW baseline with multiple trials on 8 V100s with 16GB each. This is an example of an OOM we observe on the subsequent trial:

I0229 11:25:52.179392 139670264227648 submission_runner.py:314] Starting training loop.                                                [1007/1887]
I0229 11:25:54.777722 139642405861120 logging_writer.py:48] [0] global_step=0, grad_norm=56.284924, loss=31.354658                                
I0229 11:25:54.783623 139670264227648 pytorch_nadamw_full_budget.py:296] 0) loss = 31.355, grad_norm = 56.285                                     
I0229 11:25:55.234757 139670264227648 spec.py:321] Evaluating on the training split.                                                              
I0229 11:26:08.470401 139670264227648 spec.py:333] Evaluating on the validation split.                                                            
I0229 11:26:20.220777 139670264227648 spec.py:349] Evaluating on the test split.                                                                  
I0229 11:26:26.181465 139670264227648 submission_runner.py:414] Time since start: 34.00s,       Step: 1,        {'train/ctc_loss': 30.998378480849
826, 'train/wer': 1.253173088944333, 'validation/ctc_loss': 29.436887863844042, 'validation/wer': 1.157427702409115, 'validation/num_examples': 53
48, 'test/ctc_loss': 29.538330745441247, 'test/wer': 1.1866837283935572, 'test/num_examples': 2472, 'score': 2.605729579925537, 'total_duration': 
34.00228404998779, 'accumulated_submission_time': 2.605729579925537, 'accumulated_eval_time': 30.946528434753418, 'accumulated_logging_time': 0}  
I0229 11:26:26.201718 139642405861120 logging_writer.py:48] [1] accumulated_eval_time=30.946528, accumulated_logging_time=0, accumulated_submissio
n_time=2.605730, global_step=1, preemption_count=0, score=2.605730, test/ctc_loss=29.538331, test/num_examples=2472, test/wer=1.186684, total_dura
tion=34.002284, train/ctc_loss=30.998378, train/wer=1.253173, validation/ctc_loss=29.436888, validation/num_examples=5348, validation/wer=1.157428
I0229 11:26:26.931042 139670264227648 checkpoint_utils.py:240] Saved checkpoint to /experiment_runs/nadamw_test/librispeech_conformer_pytorch/tria
l_4/checkpoint_1.                                                                                                                                 
I0229 11:26:28.782347 139642397468416 logging_writer.py:48] [1] global_step=1, grad_norm=60.090302, loss=31.114534                                
I0229 11:26:28.785351 139670264227648 pytorch_nadamw_full_budget.py:296] 1) loss = 31.115, grad_norm = 60.090                                     
I0229 11:26:30.031821 139642405861120 logging_writer.py:48] [2] global_step=2, grad_norm=72.279572, loss=30.439383                                
I0229 11:26:30.035584 139670264227648 pytorch_nadamw_full_budget.py:296] 2) loss = 30.439, grad_norm = 72.280                                     
I0229 11:26:30.954263 139642397468416 logging_writer.py:48] [3] global_step=3, grad_norm=108.894348, loss=29.346493                               
I0229 11:26:30.957295 139670264227648 pytorch_nadamw_full_budget.py:296] 3) loss = 29.346, grad_norm = 108.894                                    
Traceback (most recent call last):                                                                                                                
  File "submission_runner.py", line 697, in <module>                                                                                              
    app.run(main)                                                                                                                                 
  File "/usr/local/lib/python3.8/dist-packages/absl/app.py", line 308, in run
    _run_main(main, args)
  File "/usr/local/lib/python3.8/dist-packages/absl/app.py", line 254, in _run_main
    sys.exit(main(argv))
  File "submission_runner.py", line 665, in main
    score = score_submission_on_workload(
  File "submission_runner.py", line 576, in score_submission_on_workload
    timing, metrics = train_once(workload, workload_name,
  File "submission_runner.py", line 336, in train_once
    optimizer_state, model_params, model_state = update_params(
  File "/algorithmic-efficiency/submissions/baseline_submission/pytorch_nadamw_full_budget.py", line 276, in update_params
    loss.backward()
  File "/usr/local/lib/python3.8/dist-packages/torch/_tensor.py", line 492, in backward
    torch.autograd.backward(
  File "/usr/local/lib/python3.8/dist-packages/torch/autograd/__init__.py", line 251, in backward
    Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.51 GiB. GPU 4 has a total capacty of 15.77 GiB of which 865.31 MiB is free. P
rocess 443130 has 14.92 GiB memory in use. Of the allocated memory 6.25 GiB is allocated by PyTorch, and 7.69 GiB is reserved by PyTorch but unall
ocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management
 and PYTORCH_CUDA_ALLOC_CONF

Alternatively, if we try to resume the multi-trial run, we will observe the NCCL error:

I0225 07:05:33.506379 140182089410368 submission_runner.py:589] Timing: 61068.13836145401                        
I0225 07:05:33.506450 140182089410368 submission_runner.py:591] Total number of evals: 47                        
I0225 07:05:33.506514 140182089410368 submission_runner.py:592] ====================                             
I0225 07:05:33.506577 140182089410368 submission_runner.py:545] Using RNG seed 1817859550                        
I0225 07:05:33.507572 140182089410368 submission_runner.py:554] --- Tuning run 2/5 ---                           
I0225 07:05:33.507675 140182089410368 submission_runner.py:559] Creating tuning directory at /experiment_runs/nad
amw_baseline/librispeech_conformer_pytorch/trial_2.                                                              
I0225 07:05:33.507940 140182089410368 logger_utils.py:92] Saving hparams to /experiment_runs/nadamw_baseline/libr
ispeech_conformer_pytorch/trial_2/hparams.json.                                                                  
I0225 07:05:33.508602 140182089410368 submission_runner.py:206] Initializing dataset.                            
I0225 07:05:33.508736 140182089410368 input_pipeline.py:20] Loading split = train-clean-100                      
I0225 07:05:33.534054 140182089410368 input_pipeline.py:20] Loading split = train-clean-360                      
I0225 07:05:33.866440 140182089410368 input_pipeline.py:20] Loading split = train-other-500
I0225 07:05:34.242223 140182089410368 submission_runner.py:213] Initializing model.
[E ProcessGroupNCCL.cpp:474] [Rank 7] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=5, OpType=BRO
ADCAST, NumelIn=4918, NumelOut=4918, Timeout(ms)=1800000) ran for 1800075 milliseconds before timing out.
[E ProcessGroupNCCL.cpp:474] [Rank 4] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=5, OpType=BRO
ADCAST, NumelIn=4918, NumelOut=4918, Timeout(ms)=1800000) ran for 1800070 milliseconds before timing out.
[E ProcessGroupNCCL.cpp:474] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=5, OpType=BRO
ADCAST, NumelIn=4918, NumelOut=4918, Timeout(ms)=1800000) ran for 1800128 milliseconds before timing out.
[E ProcessGroupNCCL.cpp:474] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=5, OpType=BRO
ADCAST, NumelIn=4918, NumelOut=4918, Timeout(ms)=1800000) ran for 1800597 milliseconds before timing out.
[E ProcessGroupNCCL.cpp:474] [Rank 6] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=5, OpType=BRO
ADCAST, NumelIn=4918, NumelOut=4918, Timeout(ms)=1800000) ran for 1800573 milliseconds before timing out.
[E ProcessGroupNCCL.cpp:474] [Rank 5] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=5, OpType=BRO
ADCAST, NumelIn=4918, NumelOut=4918, Timeout(ms)=1800000) ran for 1800614 milliseconds before timing out.
[E ProcessGroupNCCL.cpp:488] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUD
A kernels, subsequent GPU operations might run on corrupted/incomplete data.
[E ProcessGroupNCCL.cpp:494] To avoid data inconsistency, we are taking the entire process down.
[E ProcessGroupNCCL.cpp:915] [Rank 2] NCCL watchdog thread terminated with exception: [Rank 2] Watchdog caught co
llective operation timeout: WorkNCCL(SeqNum=5, OpType=BROADCAST, NumelIn=4918, NumelOut=4918, Timeout(ms)=1800000
) ran for 1800128 milliseconds before timing out.       
terminate called after throwing an instance of 'std::runtime_error'                                              
  what():  [Rank 2] NCCL watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation
 timeout: WorkNCCL(SeqNum=5, OpType=BROADCAST, NumelIn=4918, NumelOut=4918, Timeout(ms)=1800000) ran for 1800128 
milliseconds before timing out.                         
Fatal Python error: Aborted 

cc @anana10c @mikerabbat @tsunghsienlee @yuchenhao @shintaro-iwasaki

Steps to Reproduce

In the Docker container, run:

torchrun --redirects 1:0,2:0,3:0,4:0,5:0,6:0,7:0 \
    --standalone \
    --nnodes=1 \
    --nproc_per_node=8 \
    submission_runner.py \
    --framework=pytorch \
    --data_dir=/data/librispeech/ \
    --workload=librispeech_conformer \
    --experiment_dir=/experiment_runs \
    --experiment_name=librispeech_conformer_baseline \
  --submission_path=reference_algorithms/paper_baselines/nadamw/pytorch/submission.py \
    --tuning_search_space=reference_algorithms/paper_baselines/nadamw/tuning_search_space.json \
  --librispeech_tokenizer_vocab_path=/data/librispeech/spm_model.vocab \
  --num_tuning_trials=5 

Source or Possible Fix

We are not aware of a possible fix for this issue. We suspect there may be a memory leak in the PyTorch LibriSpeech Conformer workload. Please let us know how to proceed. Thanks in advance!