======================== START TIME: Wed Jul 3 09:44:16 UTC 2024 python3 version = Python 3.10.14 ======================== The token has not been saved to the git credentials helper. Pass `add_to_git_credential=True` in this function directly or `--add-to-git-credential` if using via `huggingface-cli` if you want to set the git credential as well. Token is valid (permission: write). Your token has been saved to /admin/home/ferdinand_mom/.cache/huggingface/token Login successful Already on 'bench_cluster' M examples/config_tiny_llama.py M examples/config_tiny_llama.yaml M examples/train_tiny_llama.sh M src/nanotron/models/llama.py M src/nanotron/trainer.py Your branch is up to date with 'origin/bench_cluster'. Job status: RUNNING W0703 09:44:18.870000 140322518665024 torch/distributed/run.py:757] W0703 09:44:18.870000 140322518665024 torch/distributed/run.py:757] ***************************************** W0703 09:44:18.870000 140322518665024 torch/distributed/run.py:757] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. W0703 09:44:18.870000 140322518665024 torch/distributed/run.py:757] ***************************************** W0703 09:44:18.870000 140499797497664 torch/distributed/run.py:757] W0703 09:44:18.870000 140499797497664 torch/distributed/run.py:757] ***************************************** W0703 09:44:18.870000 140499797497664 torch/distributed/run.py:757] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. W0703 09:44:18.870000 140499797497664 torch/distributed/run.py:757] ***************************************** W0703 09:44:18.870000 139651575428928 torch/distributed/run.py:757] W0703 09:44:18.870000 139651575428928 torch/distributed/run.py:757] ***************************************** W0703 09:44:18.870000 139651575428928 torch/distributed/run.py:757] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. W0703 09:44:18.870000 139651575428928 torch/distributed/run.py:757] ***************************************** W0703 09:44:18.871000 140023290472256 torch/distributed/run.py:757] W0703 09:44:18.871000 140023290472256 torch/distributed/run.py:757] ***************************************** W0703 09:44:18.871000 140023290472256 torch/distributed/run.py:757] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. W0703 09:44:18.871000 140023290472256 torch/distributed/run.py:757] ***************************************** W0703 09:44:18.871000 139735752345408 torch/distributed/run.py:757] W0703 09:44:18.871000 139735752345408 torch/distributed/run.py:757] ***************************************** W0703 09:44:18.871000 139735752345408 torch/distributed/run.py:757] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. W0703 09:44:18.871000 139735752345408 torch/distributed/run.py:757] ***************************************** W0703 09:44:18.878000 140180047779648 torch/distributed/run.py:757] W0703 09:44:18.878000 140180047779648 torch/distributed/run.py:757] ***************************************** W0703 09:44:18.878000 140180047779648 torch/distributed/run.py:757] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. W0703 09:44:18.878000 140180047779648 torch/distributed/run.py:757] ***************************************** W0703 09:44:18.882000 140089905219392 torch/distributed/run.py:757] W0703 09:44:18.882000 140089905219392 torch/distributed/run.py:757] ***************************************** W0703 09:44:18.882000 140089905219392 torch/distributed/run.py:757] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. W0703 09:44:18.882000 140089905219392 torch/distributed/run.py:757] ***************************************** W0703 09:44:18.889000 140554625292096 torch/distributed/run.py:757] W0703 09:44:18.889000 140554625292096 torch/distributed/run.py:757] ***************************************** W0703 09:44:18.889000 140554625292096 torch/distributed/run.py:757] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. W0703 09:44:18.889000 140554625292096 torch/distributed/run.py:757] ***************************************** [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: Config: [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: Config(general=GeneralArgs(project='bench_cluster', [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: run='%date_%jobid', [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: seed=42, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: step=None, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: consumed_train_samples=None, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: benchmark_csv_path=None, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: ignore_sanity_checks=True), [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: parallelism=ParallelismArgs(dp=16, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: pp=4, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: tp=1, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: pp_engine=, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: tp_mode=, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: tp_linear_async_communication=False, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: expert_parallel_size=1), [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: model=ModelArgs(model_config=LlamaConfig(bos_token_id=1, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: eos_token_id=2, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: hidden_act='silu', [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: hidden_size=2048, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: initializer_range=0.02, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: intermediate_size=4096, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: is_llama_config=True, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: max_position_embeddings=4096, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: num_attention_heads=32, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: num_hidden_layers=24, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: num_key_value_heads=32, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: pad_token_id=None, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: pretraining_tp=1, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: rms_norm_eps=1e-05, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: rope_scaling=None, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: rope_theta=10000.0, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: tie_word_embeddings=True, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: use_cache=True, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: vocab_size=50257), [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: init_method=RandomInit(std=0.025), [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: dtype=torch.bfloat16, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: make_vocab_size_divisible_by=1, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: ddp_bucket_cap_mb=25), [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: tokenizer=TokenizerArgs(tokenizer_name_or_path='openai-community/gpt2', [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: tokenizer_revision=None, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: tokenizer_max_length=None), [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: checkpoints=CheckpointsArgs(checkpoints_path=Path('/dev/null'), [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: checkpoint_interval=100000, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: save_initial_state=False, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: resume_checkpoint_path=None, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: checkpoints_path_is_shared_file_system=False), [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: logging=LoggingArgs(log_level='info', [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: log_level_replica='info', [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: iteration_step_info_interval=1), [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: tokens=TokensArgs(sequence_length=4096, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: train_steps=20, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: micro_batch_size=1, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: batch_accumulation_per_replica=64, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: val_check_interval=-1, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: limit_val_batches=0, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: limit_test_batches=0), [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: optimizer=OptimizerArgs(optimizer_factory=AdamWOptimizerArgs(adam_eps=1e-08, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: adam_beta1=0.9, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: adam_beta2=0.95, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: torch_adam_is_fused=True, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: name='adamW'), [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: zero_stage=1, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: weight_decay=0.01, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: clip_grad=1.0, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: accumulate_grad_in_fp32=True, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: learning_rate_scheduler=LRSchedulerArgs(learning_rate=0.0001, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: lr_warmup_steps=1, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: lr_warmup_style='linear', [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: lr_decay_style='linear', [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: lr_decay_steps=19, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: lr_decay_starting_step=None, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: min_decay_lr=1e-05)), [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: data_stages=[DatasetStageArgs(name='Training Stage', [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: start_training_step=1, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: data=DataArgs(dataset=PretrainDatasetsArgs(hf_dataset_or_datasets='roneneldan/TinyStories', [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: hf_dataset_splits='train', [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: hf_dataset_config_name=None, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: dataset_processing_num_proc_per_process=64, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: dataset_overwrite_cache=False, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: text_column_name='text'), [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: seed=42, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: num_loading_workers=0))], [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: profiler=ProfilerArgs(profiler_export_path=Path('/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/64_GPUS/dp-16_tp-1_pp-4_mbz-1')), [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: lighteval=None) [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: Model Config: [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: LlamaConfig(bos_token_id=1, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: eos_token_id=2, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: hidden_act='silu', [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: hidden_size=2048, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: initializer_range=0.02, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: intermediate_size=4096, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: is_llama_config=True, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: max_position_embeddings=4096, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: num_attention_heads=32, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: num_hidden_layers=24, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: num_key_value_heads=32, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: pad_token_id=None, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: pretraining_tp=1, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: rms_norm_eps=1e-05, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: rope_scaling=None, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: rope_theta=10000.0, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: tie_word_embeddings=True, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: use_cache=True, [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: vocab_size=50257) [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: Building model.. [default0]:07/03/2024 09:44:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: Setting PP block ranks... [default0]:07/03/2024 09:44:51 [INFO|DP=8|PP=0|TP=0|ip-26-0-163-220]: No checkpoint path provided. [default1]:07/03/2024 09:44:51 [INFO|DP=9|PP=0|TP=0|ip-26-0-163-220]: No checkpoint path provided. [default2]:07/03/2024 09:44:51 [INFO|DP=10|PP=0|TP=0|ip-26-0-163-220]: No checkpoint path provided. [default5]:07/03/2024 09:44:51 [INFO|DP=13|PP=0|TP=0|ip-26-0-163-220]: No checkpoint path provided. [default3]:07/03/2024 09:44:51 [INFO|DP=11|PP=0|TP=0|ip-26-0-163-220]: No checkpoint path provided. [default2]:07/03/2024 09:44:51 [INFO|DP=10|PP=3|TP=0|ip-26-0-172-73]: No checkpoint path provided. [default1]:07/03/2024 09:44:51 [INFO|DP=9|PP=3|TP=0|ip-26-0-172-73]: No checkpoint path provided. [default0]:07/03/2024 09:44:51 [INFO|DP=8|PP=3|TP=0|ip-26-0-172-73]: No checkpoint path provided. [default0]:07/03/2024 09:44:51 [INFO|DP=8|PP=2|TP=0|ip-26-0-169-86]: No checkpoint path provided. [default3]:07/03/2024 09:44:51 [INFO|DP=11|PP=3|TP=0|ip-26-0-172-73]: No checkpoint path provided. [default4]:07/03/2024 09:44:51 [INFO|DP=12|PP=3|TP=0|ip-26-0-172-73]: No checkpoint path provided. [default7]:07/03/2024 09:44:51 [INFO|DP=15|PP=3|TP=0|ip-26-0-172-73]: No checkpoint path provided. [default5]:07/03/2024 09:44:51 [INFO|DP=13|PP=3|TP=0|ip-26-0-172-73]: No checkpoint path provided. [default1]:07/03/2024 09:44:51 [INFO|DP=9|PP=2|TP=0|ip-26-0-169-86]: No checkpoint path provided. [default7]:07/03/2024 09:44:51 [INFO|DP=15|PP=2|TP=0|ip-26-0-169-86]: No checkpoint path provided. [default2]:07/03/2024 09:44:51 [INFO|DP=10|PP=2|TP=0|ip-26-0-169-86]: No checkpoint path provided. [default4]:07/03/2024 09:44:51 [INFO|DP=12|PP=2|TP=0|ip-26-0-169-86]: No checkpoint path provided. [default5]:07/03/2024 09:44:51 [INFO|DP=13|PP=2|TP=0|ip-26-0-169-86]: No checkpoint path provided. [default3]:07/03/2024 09:44:51 [INFO|DP=11|PP=2|TP=0|ip-26-0-169-86]: No checkpoint path provided. [default6]:07/03/2024 09:44:51 [INFO|DP=14|PP=0|TP=0|ip-26-0-163-220]: No checkpoint path provided. [default4]:07/03/2024 09:44:51 [INFO|DP=12|PP=0|TP=0|ip-26-0-163-220]: No checkpoint path provided. [default2]:07/03/2024 09:44:51 [INFO|DP=10|PP=1|TP=0|ip-26-0-168-238]: No checkpoint path provided. [default0]:07/03/2024 09:44:51 [INFO|DP=8|PP=1|TP=0|ip-26-0-168-238]: No checkpoint path provided. [default4]:07/03/2024 09:44:51 [INFO|DP=12|PP=1|TP=0|ip-26-0-168-238]: No checkpoint path provided. [default5]:07/03/2024 09:44:51 [INFO|DP=13|PP=1|TP=0|ip-26-0-168-238]: No checkpoint path provided. [default7]:07/03/2024 09:44:51 [INFO|DP=15|PP=0|TP=0|ip-26-0-163-220]: No checkpoint path provided. [default1]:07/03/2024 09:44:51 [INFO|DP=9|PP=1|TP=0|ip-26-0-168-238]: No checkpoint path provided. [default6]:07/03/2024 09:44:51 [INFO|DP=14|PP=1|TP=0|ip-26-0-168-238]: No checkpoint path provided. [default3]:07/03/2024 09:44:51 [INFO|DP=11|PP=1|TP=0|ip-26-0-168-238]: No checkpoint path provided. [default7]:07/03/2024 09:44:51 [INFO|DP=15|PP=1|TP=0|ip-26-0-168-238]: No checkpoint path provided. [default6]:07/03/2024 09:44:51 [INFO|DP=14|PP=3|TP=0|ip-26-0-172-73]: No checkpoint path provided. [default6]:07/03/2024 09:44:51 [INFO|DP=14|PP=2|TP=0|ip-26-0-169-86]: No checkpoint path provided. [default5]:07/03/2024 09:44:51 [INFO|DP=5|PP=1|TP=0|ip-26-0-163-226]: No checkpoint path provided. [default4]:07/03/2024 09:44:51 [INFO|DP=4|PP=1|TP=0|ip-26-0-163-226]: No checkpoint path provided. [default6]:07/03/2024 09:44:51 [INFO|DP=6|PP=1|TP=0|ip-26-0-163-226]: No checkpoint path provided. [default0]:07/03/2024 09:44:51 [INFO|DP=0|PP=1|TP=0|ip-26-0-163-226]: Local number of parameters: 294M (560.05MiB) [default0]:07/03/2024 09:44:51 [INFO|DP=0|PP=1|TP=0|ip-26-0-163-226]: [After model building] Memory usage: 567.07MiB. Peak allocated: 569.10MiB Peak reserved: 594.00MiB [default0]:07/03/2024 09:44:51 [INFO|DP=0|PP=1|TP=0|ip-26-0-163-226]: No checkpoint path provided. [default2]:07/03/2024 09:44:51 [INFO|DP=2|PP=3|TP=0|ip-26-0-172-57]: No checkpoint path provided. [default2]:07/03/2024 09:44:51 [INFO|DP=2|PP=1|TP=0|ip-26-0-163-226]: No checkpoint path provided. [default7]:07/03/2024 09:44:51 [INFO|DP=7|PP=3|TP=0|ip-26-0-172-57]: No checkpoint path provided. [default0]:07/03/2024 09:44:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: Total number of parameters: 1.21G (2312.82MiB) [default0]:07/03/2024 09:44:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: Local number of parameters: 397M (756.37MiB) [default0]:07/03/2024 09:44:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [After model building] Memory usage: 763.38MiB. Peak allocated: 765.41MiB Peak reserved: 792.00MiB [default0]:07/03/2024 09:44:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: No checkpoint path provided. [default0]:07/03/2024 09:44:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: Parametrizing model parameters using StandardParametrizator [default1]:07/03/2024 09:44:51 [INFO|DP=1|PP=1|TP=0|ip-26-0-163-226]: No checkpoint path provided. [default7]:07/03/2024 09:44:51 [INFO|DP=7|PP=1|TP=0|ip-26-0-163-226]: No checkpoint path provided. [default0]:07/03/2024 09:44:51 [INFO|DP=0|PP=2|TP=0|ip-26-0-169-132]: Local number of parameters: 252M (480.05MiB) [default0]:07/03/2024 09:44:51 [INFO|DP=0|PP=2|TP=0|ip-26-0-169-132]: [After model building] Memory usage: 486.06MiB. Peak allocated: 488.09MiB Peak reserved: 502.00MiB [default0]:07/03/2024 09:44:51 [INFO|DP=0|PP=2|TP=0|ip-26-0-169-132]: No checkpoint path provided. [default1]:07/03/2024 09:44:51 [INFO|DP=1|PP=3|TP=0|ip-26-0-172-57]: No checkpoint path provided. [default6]:07/03/2024 09:44:51 [INFO|DP=6|PP=3|TP=0|ip-26-0-172-57]: No checkpoint path provided. [default0]:07/03/2024 09:44:51 [INFO|DP=0|PP=3|TP=0|ip-26-0-172-57]: Local number of parameters: 271M (516.35MiB) [default0]:07/03/2024 09:44:51 [INFO|DP=0|PP=3|TP=0|ip-26-0-172-57]: [After model building] Memory usage: 520.36MiB. Peak allocated: 522.39MiB Peak reserved: 534.00MiB [default0]:07/03/2024 09:44:51 [INFO|DP=0|PP=3|TP=0|ip-26-0-172-57]: No checkpoint path provided. [default1]:07/03/2024 09:44:51 [INFO|DP=1|PP=0|TP=0|ip-26-0-160-192]: No checkpoint path provided. [default5]:07/03/2024 09:44:51 [INFO|DP=5|PP=3|TP=0|ip-26-0-172-57]: No checkpoint path provided. [default4]:07/03/2024 09:44:51 [INFO|DP=4|PP=3|TP=0|ip-26-0-172-57]: No checkpoint path provided. [default4]:07/03/2024 09:44:51 [INFO|DP=4|PP=0|TP=0|ip-26-0-160-192]: No checkpoint path provided. [default5]:07/03/2024 09:44:51 [INFO|DP=5|PP=0|TP=0|ip-26-0-160-192]: No checkpoint path provided. [default1]:07/03/2024 09:44:51 [INFO|DP=1|PP=2|TP=0|ip-26-0-169-132]: No checkpoint path provided. [default6]:07/03/2024 09:44:51 [INFO|DP=6|PP=0|TP=0|ip-26-0-160-192]: No checkpoint path provided. [default2]:07/03/2024 09:44:51 [INFO|DP=2|PP=2|TP=0|ip-26-0-169-132]: No checkpoint path provided. [default4]:07/03/2024 09:44:51 [INFO|DP=4|PP=2|TP=0|ip-26-0-169-132]: No checkpoint path provided. [default5]:07/03/2024 09:44:51 [INFO|DP=5|PP=2|TP=0|ip-26-0-169-132]: No checkpoint path provided. [default6]:07/03/2024 09:44:51 [INFO|DP=6|PP=2|TP=0|ip-26-0-169-132]: No checkpoint path provided. [default2]:07/03/2024 09:44:51 [INFO|DP=2|PP=0|TP=0|ip-26-0-160-192]: No checkpoint path provided. [default7]:07/03/2024 09:44:51 [INFO|DP=7|PP=0|TP=0|ip-26-0-160-192]: No checkpoint path provided. [default7]:07/03/2024 09:44:51 [INFO|DP=7|PP=2|TP=0|ip-26-0-169-132]: No checkpoint path provided. [default3]:07/03/2024 09:44:51 [INFO|DP=3|PP=1|TP=0|ip-26-0-163-226]: No checkpoint path provided. [default3]:07/03/2024 09:44:51 [INFO|DP=3|PP=3|TP=0|ip-26-0-172-57]: No checkpoint path provided. [default3]:07/03/2024 09:44:51 [INFO|DP=3|PP=0|TP=0|ip-26-0-160-192]: No checkpoint path provided. [default3]:07/03/2024 09:44:51 [INFO|DP=3|PP=2|TP=0|ip-26-0-169-132]: No checkpoint path provided. [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [Optimizer Building] Using LearningRateForSP as learning rate [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] Size of optimizer params per rank: [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] DP Rank 0 has 24.8M out of 397M (6.25%) params' optimizer states [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] DP Rank 1 has 24.8M out of 397M (6.25%) params' optimizer states [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] DP Rank 2 has 24.8M out of 397M (6.25%) params' optimizer states [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] DP Rank 3 has 24.8M out of 397M (6.25%) params' optimizer states [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] DP Rank 4 has 24.8M out of 397M (6.25%) params' optimizer states [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] DP Rank 5 has 24.8M out of 397M (6.25%) params' optimizer states [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] DP Rank 6 has 24.8M out of 397M (6.25%) params' optimizer states [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] DP Rank 7 has 24.8M out of 397M (6.25%) params' optimizer states [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] DP Rank 8 has 24.8M out of 397M (6.25%) params' optimizer states [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] DP Rank 9 has 24.8M out of 397M (6.25%) params' optimizer states [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] DP Rank 10 has 24.8M out of 397M (6.25%) params' optimizer states [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] DP Rank 11 has 24.8M out of 397M (6.25%) params' optimizer states [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] DP Rank 12 has 24.8M out of 397M (6.25%) params' optimizer states [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] DP Rank 13 has 24.8M out of 397M (6.25%) params' optimizer states [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] DP Rank 14 has 24.8M out of 397M (6.25%) params' optimizer states [default0]:07/03/2024 09:45:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [ZeRO sharding] DP Rank 15 has 24.8M out of 397M (6.25%) params' optimizer states [default0]:07/03/2024 09:45:01 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [Training Plan] Stage Training Stage has 19 remaining training steps and has consumed 0 samples [default0]:07/03/2024 09:45:01 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: Using `datasets` library [default0]:07/03/2024 09:45:01 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: Loading tokenizer from openai-community/gpt2 and transformers/hf_hub versions ('4.41.2', '0.23.4') [default0]:07/03/2024 09:45:02 [WARNING|DP=0|PP=0|TP=0|ip-26-0-160-192]: Repo card metadata block was not found. Setting CardData to empty. [default0]:Repo card metadata block was not found. Setting CardData to empty. [default0]:07/03/2024 09:45:02 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [Training Plan] There are 1 training stages [default0]:07/03/2024 09:45:02 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [Stage Training Stage] start from step 1 [default0]:07/03/2024 09:45:02 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [default0]:07/03/2024 09:45:02 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: [Start training] datetime: 2024-07-03 09:45:02.858407 | mbs: 1 | grad_accum: 64 | global_batch_size: 1024 | sequence_length: 4096 | train_steps: 20 | start_iteration_step: 0 | consumed_train_samples: 0 [default0]:07/03/2024 09:45:02 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: Resuming training from stage Training Stage, it has trained for 0 samples and has 19 remaining train steps [default0]:07/03/2024 09:45:02 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-192]: Memory usage: 2370.67MiB. Peak allocated 2370.67MiB. Peak reserved: 2402.00MiB [default1]:Repo card metadata block was not found. Setting CardData to empty. [default7]:Repo card metadata block was not found. Setting CardData to empty. [default4]:07/03/2024 09:45:03 [WARNING|DP=4|PP=1|TP=0|ip-26-0-163-226]: Repo card metadata block was not found. Setting CardData to empty. [default4]:Repo card metadata block was not found. Setting CardData to empty. [default2]:07/03/2024 09:45:03 [WARNING|DP=2|PP=3|TP=0|ip-26-0-172-57]: Repo card metadata block was not found. Setting CardData to empty. [default5]:Repo card metadata block was not found. Setting CardData to empty. [default7]:07/03/2024 09:45:03 [WARNING|DP=7|PP=1|TP=0|ip-26-0-163-226]: Repo card metadata block was not found. Setting CardData to empty. [default1]:07/03/2024 09:45:02 [WARNING|DP=1|PP=1|TP=0|ip-26-0-163-226]: Repo card metadata block was not found. Setting CardData to empty. [default1]:07/03/2024 09:45:03 [WARNING|DP=9|PP=0|TP=0|ip-26-0-163-220]: Repo card metadata block was not found. Setting CardData to empty. [default5]:07/03/2024 09:45:03 [WARNING|DP=5|PP=3|TP=0|ip-26-0-172-57]: Repo card metadata block was not found. Setting CardData to empty. [default7]:Repo card metadata block was not found. Setting CardData to empty. [default4]:07/03/2024 09:45:03 [WARNING|DP=12|PP=3|TP=0|ip-26-0-172-73]: Repo card metadata block was not found. Setting CardData to empty. [default1]:07/03/2024 09:45:03 [WARNING|DP=1|PP=0|TP=0|ip-26-0-160-192]: Repo card metadata block was not found. Setting CardData to empty. [default7]:07/03/2024 09:45:02 [WARNING|DP=7|PP=0|TP=0|ip-26-0-160-192]: Repo card metadata block was not found. Setting CardData to empty. [default5]:07/03/2024 09:45:03 [WARNING|DP=13|PP=2|TP=0|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty. [default3]:07/03/2024 09:45:03 [WARNING|DP=11|PP=2|TP=0|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty. [default2]:Repo card metadata block was not found. Setting CardData to empty. [default1]:Repo card metadata block was not found. Setting CardData to empty. [default7]:Repo card metadata block was not found. Setting CardData to empty. [default1]:Repo card metadata block was not found. Setting CardData to empty. [default7]:07/03/2024 09:45:03 [WARNING|DP=15|PP=0|TP=0|ip-26-0-163-220]: Repo card metadata block was not found. Setting CardData to empty. [default4]:Repo card metadata block was not found. Setting CardData to empty. [default5]:Repo card metadata block was not found. Setting CardData to empty. [default3]:Repo card metadata block was not found. Setting CardData to empty. [default7]:07/03/2024 09:45:03 [WARNING|DP=15|PP=1|TP=0|ip-26-0-168-238]: Repo card metadata block was not found. Setting CardData to empty. [default6]:Repo card metadata block was not found. Setting CardData to empty. [default7]:Repo card metadata block was not found. Setting CardData to empty. [default5]:Repo card metadata block was not found. Setting CardData to empty. [default3]:Repo card metadata block was not found. Setting CardData to empty. [default6]:Repo card metadata block was not found. Setting CardData to empty. [default7]:07/03/2024 09:45:03 [WARNING|DP=7|PP=3|TP=0|ip-26-0-172-57]: Repo card metadata block was not found. Setting CardData to empty. [default5]:07/03/2024 09:45:03 [WARNING|DP=5|PP=1|TP=0|ip-26-0-163-226]: Repo card metadata block was not found. Setting CardData to empty. [default6]:07/03/2024 09:45:03 [WARNING|DP=6|PP=1|TP=0|ip-26-0-163-226]: Repo card metadata block was not found. Setting CardData to empty. [default2]:07/03/2024 09:45:03 [WARNING|DP=2|PP=1|TP=0|ip-26-0-163-226]: Repo card metadata block was not found. Setting CardData to empty. [default7]:Repo card metadata block was not found. Setting CardData to empty. [default0]:07/03/2024 09:45:03 [WARNING|DP=8|PP=0|TP=0|ip-26-0-163-220]: Repo card metadata block was not found. Setting CardData to empty. [default2]:07/03/2024 09:45:03 [WARNING|DP=10|PP=0|TP=0|ip-26-0-163-220]: Repo card metadata block was not found. Setting CardData to empty. [default5]:07/03/2024 09:45:03 [WARNING|DP=13|PP=0|TP=0|ip-26-0-163-220]: Repo card metadata block was not found. Setting CardData to empty. [default3]:07/03/2024 09:45:03 [WARNING|DP=11|PP=0|TP=0|ip-26-0-163-220]: Repo card metadata block was not found. Setting CardData to empty. [default0]:07/03/2024 09:45:03 [WARNING|DP=0|PP=3|TP=0|ip-26-0-172-57]: Repo card metadata block was not found. Setting CardData to empty. [default1]:07/03/2024 09:45:03 [WARNING|DP=1|PP=3|TP=0|ip-26-0-172-57]: Repo card metadata block was not found. Setting CardData to empty. [default6]:07/03/2024 09:45:03 [WARNING|DP=6|PP=3|TP=0|ip-26-0-172-57]: Repo card metadata block was not found. Setting CardData to empty. [default2]:Repo card metadata block was not found. Setting CardData to empty. [default0]:07/03/2024 09:45:03 [WARNING|DP=8|PP=3|TP=0|ip-26-0-172-73]: Repo card metadata block was not found. Setting CardData to empty. [default0]:07/03/2024 09:45:03 [WARNING|DP=0|PP=2|TP=0|ip-26-0-169-132]: Repo card metadata block was not found. Setting CardData to empty. [default6]:07/03/2024 09:45:03 [WARNING|DP=6|PP=2|TP=0|ip-26-0-169-132]: Repo card metadata block was not found. Setting CardData to empty. [default3]:07/03/2024 09:45:03 [WARNING|DP=11|PP=3|TP=0|ip-26-0-172-73]: Repo card metadata block was not found. Setting CardData to empty. [default4]:07/03/2024 09:45:03 [WARNING|DP=4|PP=0|TP=0|ip-26-0-160-192]: Repo card metadata block was not found. Setting CardData to empty. [default1]:07/03/2024 09:45:03 [WARNING|DP=1|PP=2|TP=0|ip-26-0-169-132]: Repo card metadata block was not found. Setting CardData to empty. [default4]:07/03/2024 09:45:03 [WARNING|DP=4|PP=3|TP=0|ip-26-0-172-57]: Repo card metadata block was not found. Setting CardData to empty. [default5]:07/03/2024 09:45:03 [WARNING|DP=5|PP=0|TP=0|ip-26-0-160-192]: Repo card metadata block was not found. Setting CardData to empty. [default3]:07/03/2024 09:45:03 [WARNING|DP=3|PP=3|TP=0|ip-26-0-172-57]: Repo card metadata block was not found. Setting CardData to empty. [default5]:07/03/2024 09:45:03 [WARNING|DP=5|PP=2|TP=0|ip-26-0-169-132]: Repo card metadata block was not found. Setting CardData to empty. [default4]:Repo card metadata block was not found. Setting CardData to empty. [default2]:07/03/2024 09:45:03 [WARNING|DP=2|PP=2|TP=0|ip-26-0-169-132]: Repo card metadata block was not found. Setting CardData to empty. [default7]:07/03/2024 09:45:03 [WARNING|DP=15|PP=3|TP=0|ip-26-0-172-73]: Repo card metadata block was not found. Setting CardData to empty. [default6]:07/03/2024 09:45:03 [WARNING|DP=14|PP=3|TP=0|ip-26-0-172-73]: Repo card metadata block was not found. Setting CardData to empty. [default3]:07/03/2024 09:45:03 [WARNING|DP=3|PP=2|TP=0|ip-26-0-169-132]: Repo card metadata block was not found. Setting CardData to empty. [default5]:Repo card metadata block was not found. Setting CardData to empty. [default7]:Repo card metadata block was not found. Setting CardData to empty. [default2]:07/03/2024 09:45:03 [WARNING|DP=2|PP=0|TP=0|ip-26-0-160-192]: Repo card metadata block was not found. Setting CardData to empty. [default3]:Repo card metadata block was not found. Setting CardData to empty. [default1]:Repo card metadata block was not found. Setting CardData to empty. [default6]:07/03/2024 09:45:03 [WARNING|DP=14|PP=0|TP=0|ip-26-0-163-220]: Repo card metadata block was not found. Setting CardData to empty. [default0]:Repo card metadata block was not found. Setting CardData to empty. [default0]:Repo card metadata block was not found. Setting CardData to empty. [default7]:Repo card metadata block was not found. Setting CardData to empty. [default2]:Repo card metadata block was not found. Setting CardData to empty. [default6]:Repo card metadata block was not found. Setting CardData to empty. [default6]:07/03/2024 09:45:03 [WARNING|DP=14|PP=2|TP=0|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty. [default4]:07/03/2024 09:45:03 [WARNING|DP=12|PP=0|TP=0|ip-26-0-163-220]: Repo card metadata block was not found. Setting CardData to empty. [default6]:Repo card metadata block was not found. Setting CardData to empty. [default1]:07/03/2024 09:45:03 [WARNING|DP=9|PP=2|TP=0|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty. [default7]:07/03/2024 09:45:03 [WARNING|DP=15|PP=2|TP=0|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty. [default2]:07/03/2024 09:45:03 [WARNING|DP=10|PP=2|TP=0|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty. [default4]:07/03/2024 09:45:03 [WARNING|DP=12|PP=2|TP=0|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty. [default6]:Repo card metadata block was not found. Setting CardData to empty. [default0]:Repo card metadata block was not found. Setting CardData to empty. [default6]:Repo card metadata block was not found. Setting CardData to empty. [default5]:Repo card metadata block was not found. Setting CardData to empty. [default5]:Repo card metadata block was not found. Setting CardData to empty. [default2]:Repo card metadata block was not found. Setting CardData to empty. [default6]:Repo card metadata block was not found. Setting CardData to empty. [default1]:Repo card metadata block was not found. Setting CardData to empty. [default7]:Repo card metadata block was not found. Setting CardData to empty. [default4]:Repo card metadata block was not found. Setting CardData to empty. [default7]:07/03/2024 09:45:03 [WARNING|DP=7|PP=2|TP=0|ip-26-0-169-132]: Repo card metadata block was not found. Setting CardData to empty. [default4]:07/03/2024 09:45:03 [WARNING|DP=12|PP=1|TP=0|ip-26-0-168-238]: Repo card metadata block was not found. Setting CardData to empty. [default2]:07/03/2024 09:45:03 [WARNING|DP=10|PP=1|TP=0|ip-26-0-168-238]: Repo card metadata block was not found. Setting CardData to empty. [default0]:07/03/2024 09:45:03 [WARNING|DP=8|PP=1|TP=0|ip-26-0-168-238]: Repo card metadata block was not found. Setting CardData to empty. [default3]:Repo card metadata block was not found. Setting CardData to empty. [default2]:Repo card metadata block was not found. Setting CardData to empty. [default0]:Repo card metadata block was not found. Setting CardData to empty. [default2]:Repo card metadata block was not found. Setting CardData to empty. [default4]:Repo card metadata block was not found. Setting CardData to empty. [default1]:07/03/2024 09:45:03 [WARNING|DP=9|PP=1|TP=0|ip-26-0-168-238]: Repo card metadata block was not found. Setting CardData to empty. [default6]:07/03/2024 09:45:03 [WARNING|DP=14|PP=1|TP=0|ip-26-0-168-238]: Repo card metadata block was not found. Setting CardData to empty. [default5]:07/03/2024 09:45:03 [WARNING|DP=13|PP=1|TP=0|ip-26-0-168-238]: Repo card metadata block was not found. Setting CardData to empty. [default1]:Repo card metadata block was not found. Setting CardData to empty. [default3]:Repo card metadata block was not found. Setting CardData to empty. [default3]:07/03/2024 09:45:03 [WARNING|DP=11|PP=1|TP=0|ip-26-0-168-238]: Repo card metadata block was not found. Setting CardData to empty. [default5]:Repo card metadata block was not found. Setting CardData to empty. [default0]:Repo card metadata block was not found. Setting CardData to empty. [default2]:Repo card metadata block was not found. Setting CardData to empty. [default1]:Repo card metadata block was not found. Setting CardData to empty. [default4]:Repo card metadata block was not found. Setting CardData to empty. [default3]:Repo card metadata block was not found. Setting CardData to empty. [default4]:Repo card metadata block was not found. Setting CardData to empty. [default0]:07/03/2024 09:45:03 [WARNING|DP=0|PP=1|TP=0|ip-26-0-163-226]: Repo card metadata block was not found. Setting CardData to empty. [default3]:07/03/2024 09:45:03 [WARNING|DP=3|PP=1|TP=0|ip-26-0-163-226]: Repo card metadata block was not found. Setting CardData to empty. [default0]:Repo card metadata block was not found. Setting CardData to empty. [default3]:Repo card metadata block was not found. Setting CardData to empty. [default1]:07/03/2024 09:45:03 [WARNING|DP=9|PP=3|TP=0|ip-26-0-172-73]: Repo card metadata block was not found. Setting CardData to empty. [default2]:07/03/2024 09:45:03 [WARNING|DP=10|PP=3|TP=0|ip-26-0-172-73]: Repo card metadata block was not found. Setting CardData to empty. [default4]:07/03/2024 09:45:03 [WARNING|DP=4|PP=2|TP=0|ip-26-0-169-132]: Repo card metadata block was not found. Setting CardData to empty. [default3]:07/03/2024 09:45:03 [WARNING|DP=3|PP=0|TP=0|ip-26-0-160-192]: Repo card metadata block was not found. Setting CardData to empty. [default3]:Repo card metadata block was not found. Setting CardData to empty. [default2]:Repo card metadata block was not found. Setting CardData to empty. [default1]:Repo card metadata block was not found. Setting CardData to empty. [default4]:Repo card metadata block was not found. Setting CardData to empty. [default0]:07/03/2024 09:45:03 [WARNING|DP=8|PP=2|TP=0|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty. [default0]:Repo card metadata block was not found. Setting CardData to empty. [default5]:07/03/2024 09:45:03 [WARNING|DP=13|PP=3|TP=0|ip-26-0-172-73]: Repo card metadata block was not found. Setting CardData to empty. [default6]:07/03/2024 09:45:03 [WARNING|DP=6|PP=0|TP=0|ip-26-0-160-192]: Repo card metadata block was not found. Setting CardData to empty. [default6]:Repo card metadata block was not found. Setting CardData to empty. [default5]:Repo card metadata block was not found. Setting CardData to empty. [default1]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.) [default1]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default7]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.) [default7]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default2]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.) [default2]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default4]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.) [default4]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default0]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.) [default0]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default1]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.) [default1]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default5]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.) [default5]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default3]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.) [default3]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default5]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.) [default5]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default6]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.) [default6]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default2]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.) [default2]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default6]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.) [default6]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default7]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.) [default7]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default4]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.) [default4]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default0]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.) [default0]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default3]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.) [default3]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default1]:[rank17]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600023 milliseconds before timing out. [default1]:[rank1]:[E ProcessGroupNCCL.cpp:563] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600035 milliseconds before timing out. [default1]:[rank17]: Traceback (most recent call last): [default1]:[rank17]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default1]:[rank17]: trainer.train(dataloader) [default1]:[rank17]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default1]:[rank17]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default1]:[rank17]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default1]:[rank17]: outputs = self.pipeline_engine.train_batch_iter( [default1]:[rank17]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default1]:[rank17]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default1]:[rank17]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default1]:[rank17]: grad_accumulator.backward(sum(activations)) [default1]:[rank17]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default1]:[rank17]: result = loss.backward() [default1]:[rank17]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default1]:[rank17]: torch.autograd.backward( [default1]:[rank17]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default1]:[rank17]: _engine_run_backward( [default1]:[rank17]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default1]:[rank17]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default1]:[rank17]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default1]:[rank17]: return user_fn(self, *args) [default1]:[rank17]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default1]:[rank17]: pipeline_state.run_communication() [default1]:[rank17]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default1]:[rank17]: send_activation() [default1]:[rank17]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default1]:[rank17]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default1]:[rank17]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default1]:[rank17]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default1]:[rank17]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default1]:[rank17]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default1]:[rank17]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default1]:[rank17]: dist.send( [default1]:[rank17]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default1]:[rank17]: return func(*args, **kwargs) [default1]:[rank17]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default1]:[rank17]: group.send([tensor], group_dst_rank, tag).wait() [default1]:[rank17]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default1]:[rank1]: Traceback (most recent call last): [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default1]:[rank1]: trainer.train(dataloader) [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default1]:[rank1]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default1]:[rank1]: outputs = self.pipeline_engine.train_batch_iter( [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default1]:[rank1]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default1]:[rank1]: grad_accumulator.backward(sum(activations)) [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default1]:[rank1]: result = loss.backward() [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default1]:[rank1]: torch.autograd.backward( [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default1]:[rank1]: _engine_run_backward( [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default1]:[rank1]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default1]:[rank1]: return user_fn(self, *args) [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default1]:[rank1]: pipeline_state.run_communication() [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default1]:[rank1]: send_activation() [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default1]:[rank1]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default1]:[rank1]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default1]:[rank1]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default1]:[rank1]: dist.send( [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default1]:[rank1]: return func(*args, **kwargs) [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default1]:[rank1]: group.send([tensor], group_dst_rank, tag).wait() [default1]:[rank1]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default1]:[rank1]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 0] Timeout at NCCL work: 22, last enqueued NCCL work: 23, last completed NCCL work: 21. [default1]:[rank1]:[E ProcessGroupNCCL.cpp:577] [Rank 0] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default1]:[rank1]:[E ProcessGroupNCCL.cpp:583] [Rank 0] To avoid data inconsistency, we are taking the entire process down. [default1]:[rank1]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600035 milliseconds before timing out. [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f8823eb8897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f8825191c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f8825196a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f8825197dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #4: + 0xd3e95 (0x7f8870c30e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #5: + 0x8609 (0x7f8875c77609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #6: clone + 0x43 (0x7f8875a42353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:terminate called after throwing an instance of 'c10::DistBackendError' [default1]: what(): [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600035 milliseconds before timing out. [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f8823eb8897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f8825191c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f8825196a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f8825197dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #4: + 0xd3e95 (0x7f8870c30e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #5: + 0x8609 (0x7f8875c77609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #6: clone + 0x43 (0x7f8875a42353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f8823eb8897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: + 0xe32119 (0x7f8824e1b119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: + 0xd3e95 (0x7f8870c30e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #3: + 0x8609 (0x7f8875c77609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #4: clone + 0x43 (0x7f8875a42353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:[rank33]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600031 milliseconds before timing out. [default1]:[rank49]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600067 milliseconds before timing out. [default1]:[rank17]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 34, last enqueued NCCL work: 35, last completed NCCL work: 33. [default1]:[rank17]:[E ProcessGroupNCCL.cpp:577] [Rank 1] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default1]:[rank17]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default1]:[rank17]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600023 milliseconds before timing out. [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7efe0dad4897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7efe0edadc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7efe0edb2a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7efe0edb3dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #4: + 0xd3e95 (0x7efe5a84ce95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #5: + 0x8609 (0x7efe5f893609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #6: clone + 0x43 (0x7efe5f65e353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:terminate called after throwing an instance of 'c10::DistBackendError' [default1]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600023 milliseconds before timing out. [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7efe0dad4897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7efe0edadc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7efe0edb2a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7efe0edb3dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #4: + 0xd3e95 (0x7efe5a84ce95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #5: + 0x8609 (0x7efe5f893609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #6: clone + 0x43 (0x7efe5f65e353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7efe0dad4897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: + 0xe32119 (0x7efe0ea37119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: + 0xd3e95 (0x7efe5a84ce95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #3: + 0x8609 (0x7efe5f893609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #4: clone + 0x43 (0x7efe5f65e353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:[rank25]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600028 milliseconds before timing out. [default7]:[rank23]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600022 milliseconds before timing out. [default4]:[rank20]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600034 milliseconds before timing out. [default2]:[rank18]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600055 milliseconds before timing out. [default4]:[rank4]:[E ProcessGroupNCCL.cpp:563] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600009 milliseconds before timing out. [default2]:[rank2]:[E ProcessGroupNCCL.cpp:563] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600012 milliseconds before timing out. [default1]:[rank9]:[E ProcessGroupNCCL.cpp:563] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600058 milliseconds before timing out. [default5]:[rank21]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600090 milliseconds before timing out. [default0]:[rank16]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600094 milliseconds before timing out. [default7]:[rank7]:[E ProcessGroupNCCL.cpp:563] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600098 milliseconds before timing out. [default0]:[rank0]:[E ProcessGroupNCCL.cpp:563] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600064 milliseconds before timing out. [default5]:[rank5]:[E ProcessGroupNCCL.cpp:563] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600069 milliseconds before timing out. [default1]:[rank33]: Traceback (most recent call last): [default1]:[rank33]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default1]:[rank33]: trainer.train(dataloader) [default1]:[rank33]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default1]:[rank33]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default1]:[rank33]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default1]:[rank33]: outputs = self.pipeline_engine.train_batch_iter( [default1]:[rank33]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default1]:[rank33]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default1]:[rank33]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default1]:[rank33]: grad_accumulator.backward(sum(activations)) [default1]:[rank33]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default1]:[rank33]: result = loss.backward() [default1]:[rank33]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default1]:[rank33]: torch.autograd.backward( [default1]:[rank33]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default1]:[rank33]: _engine_run_backward( [default1]:[rank33]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default1]:[rank33]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default1]:[rank33]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default1]:[rank33]: return user_fn(self, *args) [default1]:[rank33]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default1]:[rank33]: pipeline_state.run_communication() [default1]:[rank33]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 172, in run_communication [default1]:[rank33]: self.grads_buffer.append(recv_grad()) [default1]:[rank33]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 50, in __call__ [default1]:[rank33]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default1]:[rank33]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default1]:[rank33]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default1]:[rank33]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default1]:[rank33]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default1]:[rank33]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default1]:[rank33]: dist.recv( [default1]:[rank33]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default1]:[rank33]: return func(*args, **kwargs) [default1]:[rank33]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default1]:[rank33]: pg.recv([tensor], group_src_rank, tag).wait() [default1]:[rank33]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default1]:[rank49]: Traceback (most recent call last): [default1]:[rank49]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default1]:[rank49]: trainer.train(dataloader) [default1]:[rank49]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default1]:[rank49]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default1]:[rank49]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default1]:[rank49]: outputs = self.pipeline_engine.train_batch_iter( [default1]:[rank49]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter [default1]:[rank49]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default1]:[rank49]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default1]:[rank49]: output = model(**micro_batch) [default1]:[rank49]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default1]:[rank49]: return self._call_impl(*args, **kwargs) [default1]:[rank49]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default1]:[rank49]: return forward_call(*args, **kwargs) [default1]:[rank49]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default1]:[rank49]: sharded_logits = self.model( [default1]:[rank49]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default1]:[rank49]: return self._call_impl(*args, **kwargs) [default1]:[rank49]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default1]:[rank49]: return forward_call(*args, **kwargs) [default1]:[rank49]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default1]:[rank49]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default1]:[rank49]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default1]:[rank49]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default1]:[rank49]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default1]:[rank49]: return self._call_impl(*args, **kwargs) [default1]:[rank49]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default1]:[rank49]: return forward_call(*args, **kwargs) [default1]:[rank49]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default1]:[rank49]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default1]:[rank49]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default1]:[rank49]: pipeline_state.run_communication() [default1]:[rank49]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default1]:[rank49]: recv_activation_tensor = recv_activation() [default1]:[rank49]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default1]:[rank49]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default1]:[rank49]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default1]:[rank49]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default1]:[rank49]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default1]:[rank49]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default1]:[rank49]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default1]:[rank49]: dist.recv( [default1]:[rank49]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default1]:[rank49]: return func(*args, **kwargs) [default1]:[rank49]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default1]:[rank49]: pg.recv([tensor], group_src_rank, tag).wait() [default1]:[rank49]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 1. [default3]:[rank3]:[E ProcessGroupNCCL.cpp:563] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600022 milliseconds before timing out. [default3]:[rank19]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600079 milliseconds before timing out. [default6]:[rank22]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600039 milliseconds before timing out. [default6]:[rank6]:[E ProcessGroupNCCL.cpp:563] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600030 milliseconds before timing out. [default1]:[rank49]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 14, last enqueued NCCL work: 15, last completed NCCL work: 13. [default1]:[rank49]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default1]:[rank49]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default1]:[rank49]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600067 milliseconds before timing out. [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f2a6e15f897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f2a6f438c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f2a6f43da80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f2a6f43edcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #4: + 0xd3e95 (0x7f2abaed7e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #5: + 0x8609 (0x7f2abff1e609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #6: clone + 0x43 (0x7f2abfce9353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:terminate called after throwing an instance of 'c10::DistBackendError' [default1]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600067 milliseconds before timing out. [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f2a6e15f897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f2a6f438c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f2a6f43da80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f2a6f43edcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #4: + 0xd3e95 (0x7f2abaed7e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #5: + 0x8609 (0x7f2abff1e609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #6: clone + 0x43 (0x7f2abfce9353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f2a6e15f897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: + 0xe32119 (0x7f2a6f0c2119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: + 0xd3e95 (0x7f2abaed7e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #3: + 0x8609 (0x7f2abff1e609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #4: clone + 0x43 (0x7f2abfce9353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default4]:[rank28]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600017 milliseconds before timing out. [default0]:[rank0]: Traceback (most recent call last): [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default0]:[rank0]: trainer.train(dataloader) [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default0]:[rank0]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default0]:[rank0]: outputs = self.pipeline_engine.train_batch_iter( [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default0]:[rank0]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default0]:[rank0]: grad_accumulator.backward(sum(activations)) [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default0]:[rank0]: result = loss.backward() [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default0]:[rank0]: torch.autograd.backward( [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default0]:[rank0]: _engine_run_backward( [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default0]:[rank0]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default0]:[rank0]: return user_fn(self, *args) [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default0]:[rank0]: pipeline_state.run_communication() [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default0]:[rank0]: send_activation() [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default0]:[rank0]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default0]:[rank0]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default0]:[rank0]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default0]:[rank0]: dist.send( [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default0]:[rank0]: return func(*args, **kwargs) [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default0]:[rank0]: group.send([tensor], group_dst_rank, tag).wait() [default0]:[rank0]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default0]:[rank24]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600009 milliseconds before timing out. [default5]:[rank29]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600062 milliseconds before timing out. [default2]:[rank18]: Traceback (most recent call last): [default2]:[rank18]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default2]:[rank18]: trainer.train(dataloader) [default2]:[rank18]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default2]:[rank18]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default2]:[rank18]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default2]:[rank18]: outputs = self.pipeline_engine.train_batch_iter( [default2]:[rank18]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default2]:[rank18]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default2]:[rank18]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default2]:[rank18]: grad_accumulator.backward(sum(activations)) [default2]:[rank18]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default2]:[rank18]: result = loss.backward() [default2]:[rank18]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default2]:[rank18]: torch.autograd.backward( [default2]:[rank18]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default2]:[rank18]: _engine_run_backward( [default2]:[rank18]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default2]:[rank18]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default2]:[rank18]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default2]:[rank18]: return user_fn(self, *args) [default2]:[rank18]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default2]:[rank18]: pipeline_state.run_communication() [default2]:[rank18]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default2]:[rank18]: send_activation() [default2]:[rank18]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default2]:[rank18]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default2]:[rank18]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default2]:[rank18]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default2]:[rank18]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default2]:[rank18]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default2]:[rank18]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default2]:[rank18]: dist.send( [default2]:[rank18]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default2]:[rank18]: return func(*args, **kwargs) [default2]:[rank18]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default2]:[rank18]: group.send([tensor], group_dst_rank, tag).wait() [default2]:[rank18]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default6]:[rank14]:[E ProcessGroupNCCL.cpp:563] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600088 milliseconds before timing out. [default4]:[rank12]:[E ProcessGroupNCCL.cpp:563] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600081 milliseconds before timing out. [default0]:[rank8]:[E ProcessGroupNCCL.cpp:563] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600045 milliseconds before timing out. [default7]:[rank31]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600007 milliseconds before timing out. [default6]:[rank30]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600036 milliseconds before timing out. [default1]:[rank33]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 26, last enqueued NCCL work: 27, last completed NCCL work: 25. [default1]:[rank33]:[E ProcessGroupNCCL.cpp:577] [Rank 2] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default1]:[rank33]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default1]:[rank33]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600031 milliseconds before timing out. [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fafce1f3897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fafcf4ccc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fafcf4d1a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fafcf4d2dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #4: + 0xd3e95 (0x7fb01af6be95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #5: + 0x8609 (0x7fb01ffb2609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #6: clone + 0x43 (0x7fb01fd7d353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:terminate called after throwing an instance of 'c10::DistBackendError' [default1]: what(): [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600031 milliseconds before timing out. [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fafce1f3897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fafcf4ccc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fafcf4d1a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fafcf4d2dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #4: + 0xd3e95 (0x7fb01af6be95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #5: + 0x8609 (0x7fb01ffb2609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #6: clone + 0x43 (0x7fb01fd7d353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fafce1f3897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: + 0xe32119 (0x7fafcf156119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: + 0xd3e95 (0x7fb01af6be95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #3: + 0x8609 (0x7fb01ffb2609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #4: clone + 0x43 (0x7fb01fd7d353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default3]:[rank11]:[E ProcessGroupNCCL.cpp:563] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600040 milliseconds before timing out. [default7]:[rank15]:[E ProcessGroupNCCL.cpp:563] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600069 milliseconds before timing out. [default5]:[rank13]:[E ProcessGroupNCCL.cpp:563] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600092 milliseconds before timing out. [default3]:[rank27]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600085 milliseconds before timing out. [default2]:[rank26]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600072 milliseconds before timing out. [default2]:[rank2]: Traceback (most recent call last): [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default2]:[rank2]: trainer.train(dataloader) [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default2]:[rank2]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default2]:[rank2]: outputs = self.pipeline_engine.train_batch_iter( [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default2]:[rank2]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default2]:[rank2]: grad_accumulator.backward(sum(activations)) [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default2]:[rank2]: result = loss.backward() [default2]:[rank2]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default2]:[rank2]: torch.autograd.backward( [default2]:[rank2]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default2]:[rank2]: _engine_run_backward( [default2]:[rank2]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default2]:[rank2]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default2]:[rank2]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default2]:[rank2]: return user_fn(self, *args) [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default2]:[rank2]: pipeline_state.run_communication() [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default2]:[rank2]: send_activation() [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default2]:[rank2]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default2]:[rank2]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default2]:[rank2]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default2]:[rank2]: dist.send( [default2]:[rank2]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default2]:[rank2]: return func(*args, **kwargs) [default2]:[rank2]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default2]:[rank2]: group.send([tensor], group_dst_rank, tag).wait() [default2]:[rank2]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default2]:[rank10]:[E ProcessGroupNCCL.cpp:563] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600001 milliseconds before timing out. [default7]:[rank23]: Traceback (most recent call last): [default7]:[rank23]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default7]:[rank23]: trainer.train(dataloader) [default7]:[rank23]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default7]:[rank23]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default7]:[rank23]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default7]:[rank23]: outputs = self.pipeline_engine.train_batch_iter( [default7]:[rank23]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default7]:[rank23]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default7]:[rank23]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default7]:[rank23]: grad_accumulator.backward(sum(activations)) [default7]:[rank23]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default7]:[rank23]: result = loss.backward() [default7]:[rank23]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default7]:[rank23]: torch.autograd.backward( [default7]:[rank23]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default7]:[rank23]: _engine_run_backward( [default7]:[rank23]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default7]:[rank23]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default7]:[rank23]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default7]:[rank23]: return user_fn(self, *args) [default7]:[rank23]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default7]:[rank23]: pipeline_state.run_communication() [default7]:[rank23]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default7]:[rank23]: send_activation() [default7]:[rank23]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default7]:[rank23]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default7]:[rank23]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default7]:[rank23]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default7]:[rank23]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default7]:[rank23]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default7]:[rank23]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default7]:[rank23]: dist.send( [default7]:[rank23]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default7]:[rank23]: return func(*args, **kwargs) [default7]:[rank23]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default7]:[rank23]: group.send([tensor], group_dst_rank, tag).wait() [default7]:[rank23]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default4]:[rank4]: Traceback (most recent call last): [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default4]:[rank4]: trainer.train(dataloader) [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default4]:[rank4]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default4]:[rank4]: outputs = self.pipeline_engine.train_batch_iter( [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default4]:[rank4]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default4]:[rank4]: grad_accumulator.backward(sum(activations)) [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default4]:[rank4]: result = loss.backward() [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default4]:[rank4]: torch.autograd.backward( [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default4]:[rank4]: _engine_run_backward( [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default4]:[rank4]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default4]:[rank4]: return user_fn(self, *args) [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default4]:[rank4]: pipeline_state.run_communication() [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default4]:[rank4]: send_activation() [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default4]:[rank4]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default4]:[rank4]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default4]:[rank4]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default4]:[rank4]: dist.send( [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default4]:[rank4]: return func(*args, **kwargs) [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default4]:[rank4]: group.send([tensor], group_dst_rank, tag).wait() [default4]:[rank4]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default5]:[rank5]: Traceback (most recent call last): [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default5]:[rank5]: trainer.train(dataloader) [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default5]:[rank5]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default5]:[rank5]: outputs = self.pipeline_engine.train_batch_iter( [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default5]:[rank5]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default5]:[rank5]: grad_accumulator.backward(sum(activations)) [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default5]:[rank5]: result = loss.backward() [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default5]:[rank5]: torch.autograd.backward( [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default5]:[rank5]: _engine_run_backward( [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default5]:[rank5]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default5]:[rank5]: return user_fn(self, *args) [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default5]:[rank5]: pipeline_state.run_communication() [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default5]:[rank5]: send_activation() [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default5]:[rank5]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default5]:[rank5]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default5]:[rank5]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default5]:[rank5]: dist.send( [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default5]:[rank5]: return func(*args, **kwargs) [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default5]:[rank5]: group.send([tensor], group_dst_rank, tag).wait() [default5]:[rank5]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default4]:[rank20]: Traceback (most recent call last): [default4]:[rank20]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default4]:[rank20]: trainer.train(dataloader) [default4]:[rank20]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default4]:[rank20]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default4]:[rank20]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default4]:[rank20]: outputs = self.pipeline_engine.train_batch_iter( [default4]:[rank20]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default4]:[rank20]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default4]:[rank20]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default4]:[rank20]: grad_accumulator.backward(sum(activations)) [default4]:[rank20]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default4]:[rank20]: result = loss.backward() [default4]:[rank20]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default4]:[rank20]: torch.autograd.backward( [default4]:[rank20]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default4]:[rank20]: _engine_run_backward( [default4]:[rank20]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default4]:[rank20]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default4]:[rank20]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default4]:[rank20]: return user_fn(self, *args) [default4]:[rank20]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default4]:[rank20]: pipeline_state.run_communication() [default4]:[rank20]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default4]:[rank20]: send_activation() [default4]:[rank20]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default4]:[rank20]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default4]:[rank20]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default4]:[rank20]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default4]:[rank20]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default4]:[rank20]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default4]:[rank20]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default4]:[rank20]: dist.send( [default4]:[rank20]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default4]:[rank20]: return func(*args, **kwargs) [default4]:[rank20]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default4]:[rank20]: group.send([tensor], group_dst_rank, tag).wait() [default4]:[rank20]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default7]:[rank55]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600005 milliseconds before timing out. [default0]:[rank0]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 0] Timeout at NCCL work: 22, last enqueued NCCL work: 23, last completed NCCL work: 21. [default0]:[rank0]:[E ProcessGroupNCCL.cpp:577] [Rank 0] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default0]:[rank0]:[E ProcessGroupNCCL.cpp:583] [Rank 0] To avoid data inconsistency, we are taking the entire process down. [default0]:[rank0]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600064 milliseconds before timing out. [default0]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fc1e085e897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fc1e1b37c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fc1e1b3ca80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fc1e1b3ddcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #4: + 0xd3e95 (0x7fc22d5d6e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #5: + 0x8609 (0x7fc23261d609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #6: clone + 0x43 (0x7fc2323e8353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default0]:terminate called after throwing an instance of 'c10::DistBackendError' [default0]: what(): [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600064 milliseconds before timing out. [default0]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fc1e085e897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fc1e1b37c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fc1e1b3ca80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fc1e1b3ddcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #4: + 0xd3e95 (0x7fc22d5d6e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #5: + 0x8609 (0x7fc23261d609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #6: clone + 0x43 (0x7fc2323e8353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default0]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fc1e085e897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/[default7]:[rank39]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600037 milliseconds before timing out. [default1]:[rank9]: Traceback (most recent call last): [default1]:[rank9]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default1]:[rank9]: trainer.train(dataloader) [default1]:[rank9]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default1]:[rank9]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default1]:[rank9]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default1]:[rank9]: outputs = self.pipeline_engine.train_batch_iter( [default1]:[rank9]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default1]:[rank9]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default1]:[rank9]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default1]:[rank9]: grad_accumulator.backward(sum(activations)) [default1]:[rank9]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default1]:[rank9]: result = loss.backward() [default1]:[rank9]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default1]:[rank9]: torch.autograd.backward( [default1]:[rank9]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default1]:[rank9]: _engine_run_backward( [default1]:[rank9]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default1]:[rank9]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default1]:[rank9]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default1]:[rank9]: return user_fn(self, *args) [default1]:[rank9]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default1]:[rank9]: pipeline_state.run_communication() [default1]:[rank9]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default1]:[rank9]: send_activation() [default1]:[rank9]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default1]:[rank9]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default1]:[rank9]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default1]:[rank9]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default1]:[rank9]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default1]:[rank9]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default1]:[rank9]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default1]:[rank9]: dist.send( [default1]:[rank9]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default1]:[rank9]: return func(*args, **kwargs) [default1]:[rank9]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default1]:[rank9]: group.send([tensor], group_dst_rank, tag).wait() [default1]:[rank9]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. libc10.so) [default0]:frame #1: + 0xe32119 (0x7fc1e17c1119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: + 0xd3e95 (0x7fc22d5d6e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #3: + 0x8609 (0x7fc23261d609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #4: clone + 0x43 (0x7fc2323e8353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default7]:[rank39]: Traceback (most recent call last): [default7]:[rank39]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default7]:[rank39]: trainer.train(dataloader) [default7]:[rank39]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default7]:[rank39]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default7]:[rank39]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default7]:[rank39]: outputs = self.pipeline_engine.train_batch_iter( [default7]:[rank39]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default7]:[rank39]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default7]:[rank39]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default7]:[rank39]: grad_accumulator.backward(sum(activations)) [default7]:[rank39]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default7]:[rank39]: result = loss.backward() [default7]:[rank39]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default7]:[rank39]: torch.autograd.backward( [default7]:[rank39]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default7]:[rank39]: _engine_run_backward( [default7]:[rank39]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default7]:[rank39]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default7]:[rank39]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default7]:[rank39]: return user_fn(self, *args) [default7]:[rank39]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default7]:[rank39]: pipeline_state.run_communication() [default7]:[rank39]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 172, in run_communication [default7]:[rank39]: self.grads_buffer.append(recv_grad()) [default7]:[rank39]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 50, in __call__ [default7]:[rank39]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default7]:[rank39]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default7]:[rank39]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default7]:[rank39]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default7]:[rank39]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default7]:[rank39]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default7]:[rank39]: dist.recv( [default7]:[rank39]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default7]:[rank39]: return func(*args, **kwargs) [default7]:[rank39]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default7]:[rank39]: pg.recv([tensor], group_src_rank, tag).wait() [default7]:[rank39]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default2]:[rank2]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 0] Timeout at NCCL work: 22, last enqueued NCCL work: 23, last completed NCCL work: 21. [default2]:[rank2]:[E ProcessGroupNCCL.cpp:577] [Rank 0] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default2]:[rank2]:[E ProcessGroupNCCL.cpp:583] [Rank 0] To avoid data inconsistency, we are taking the entire process down. [default2]:[rank2]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600012 milliseconds before timing out. [default2]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f9b2ac2d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f9b2bf06c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f9b2bf0ba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f9b2bf0cdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7f9b779a5e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #5: + 0x8609 (0x7f9b7c9ec609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7f9b7c7b7353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default2]:terminate called after throwing an instance of 'c10::DistBackendError' [default2]: what(): [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600012 milliseconds before timing out. [default2]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f9b2ac2d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f9b2bf06c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f9b2bf0ba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f9b2bf0cdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7f9b779a5e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #5: + 0x8609 (0x7f9b7c9ec609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7f9b7c7b7353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default2]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f9b2ac2d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: + 0xe32119 (0x7f9b2bb90119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: + 0xd3e95 (0x7f9b779a5e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #3: + 0x8609 (0x7f9b7c9ec609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #4: clone + 0x43 (0x7f9b7c7b7353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default2]:[rank50]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600052 milliseconds before timing out. [default2]:[rank34]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600071 milliseconds before timing out. [default4]:[rank36]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600005 milliseconds before timing out. [default4]:[rank52]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600072 milliseconds before timing out. [default5]:[rank21]: Traceback (most recent call last): [default5]:[rank21]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default5]:[rank21]: trainer.train(dataloader) [default5]:[rank21]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default5]:[rank21]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default5]:[rank21]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default5]:[rank21]: outputs = self.pipeline_engine.train_batch_iter( [default5]:[rank21]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default5]:[rank21]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default5]:[rank21]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default5]:[rank21]: grad_accumulator.backward(sum(activations)) [default5]:[rank21]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default5]:[rank21]: result = loss.backward() [default5]:[rank21]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default5]:[rank21]: torch.autograd.backward( [default5]:[rank21]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default5]:[rank21]: _engine_run_backward( [default5]:[rank21]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default5]:[rank21]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default5]:[rank21]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default5]:[rank21]: return user_fn(self, *args) [default5]:[rank21]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default5]:[rank21]: pipeline_state.run_communication() [default5]:[rank21]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default5]:[rank21]: send_activation() [default5]:[rank21]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default5]:[rank21]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default5]:[rank21]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default5]:[rank21]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default5]:[rank21]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default5]:[rank21]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default5]:[rank21]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default5]:[rank21]: dist.send( [default5]:[rank21]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default5]:[rank21]: return func(*args, **kwargs) [default5]:[rank21]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default5]:[rank21]: group.send([tensor], group_dst_rank, tag).wait() [default5]:[rank21]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default0]:[rank16]: Traceback (most recent call last): [default0]:[rank16]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default0]:[rank16]: trainer.train(dataloader) [default0]:[rank16]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default0]:[rank16]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default0]:[rank16]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default0]:[rank16]: outputs = self.pipeline_engine.train_batch_iter( [default0]:[rank16]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default0]:[rank16]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default0]:[rank16]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default0]:[rank16]: grad_accumulator.backward(sum(activations)) [default0]:[rank16]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default0]:[rank16]: result = loss.backward() [default0]:[rank16]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default0]:[rank16]: torch.autograd.backward( [default0]:[rank16]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default0]:[rank16]: _engine_run_backward( [default0]:[rank16]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default0]:[rank16]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default0]:[rank16]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default0]:[rank16]: return user_fn(self, *args) [default0]:[rank16]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default0]:[rank16]: pipeline_state.run_communication() [default0]:[rank16]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default0]:[rank16]: send_activation() [default0]:[rank16]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default0]:[rank16]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default0]:[rank16]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default0]:[rank16]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default0]:[rank16]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default0]:[rank16]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default0]:[rank16]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default0]:[rank16]: dist.send( [default0]:[rank16]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default0]:[rank16]: return func(*args, **kwargs) [default0]:[rank16]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default0]:[rank16]: group.send([tensor], group_dst_rank, tag).wait() [default0]:[rank16]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default7]:[rank7]: Traceback (most recent call last): [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default7]:[rank7]: trainer.train(dataloader) [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default7]:[rank7]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default7]:[rank7]: outputs = self.pipeline_engine.train_batch_iter( [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default7]:[rank7]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default7]:[rank7]: grad_accumulator.backward(sum(activations)) [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default7]:[rank7]: result = loss.backward() [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default7]:[rank7]: torch.autograd.backward( [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default7]:[rank7]: _engine_run_backward( [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default7]:[rank7]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default7]:[rank7]: return user_fn(self, *args) [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default7]:[rank7]: pipeline_state.run_communication() [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default7]:[rank7]: send_activation() [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default7]:[rank7]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default7]:[rank7]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default7]:[rank7]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default7]:[rank7]: dist.send( [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default7]:[rank7]: return func(*args, **kwargs) [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default7]:[rank7]: group.send([tensor], group_dst_rank, tag).wait() [default7]:[rank7]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default4]:[rank4]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 0] Timeout at NCCL work: 22, last enqueued NCCL work: 23, last completed NCCL work: 21. [default4]:[rank4]:[E ProcessGroupNCCL.cpp:577] [Rank 0] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default4]:[rank4]:[E ProcessGroupNCCL.cpp:583] [Rank 0] To avoid data inconsistency, we are taking the entire process down. [default4]:[rank4]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600009 milliseconds before timing out. [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f1b8017a897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f1b81453c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f1b81458a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f1b81459dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f1bccef2e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f1bd1f39609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f1bd1d04353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default4]:terminate called after throwing an instance of 'c10::DistBackendError' [default4]: what(): [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600009 milliseconds before timing out. [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f1b8017a897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f1b81453c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f1b81458a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f1b81459dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f1bccef2e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f1bd1f39609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f1bd1d04353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default4]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f1b8017a897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: + 0xe32119 (0x7f1b810dd119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: + 0xd3e95 (0x7f1bccef2e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #3: + 0x8609 (0x7f1bd1f39609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #4: clone + 0x43 (0x7f1bd1d04353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default5]:[rank5]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 0] Timeout at NCCL work: 22, last enqueued NCCL work: 23, last completed NCCL work: 21. [default5]:[rank5]:[E ProcessGroupNCCL.cpp:577] [Rank 0] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default5]:[rank5]:[E ProcessGroupNCCL.cpp:583] [Rank 0] To avoid data inconsistency, we are taking the entire process down. [default5]:[rank5]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600069 milliseconds before timing out. [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fc5b9ebe897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fc5bb197c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fc5bb19ca80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fc5bb19ddcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #4: + 0xd3e95 (0x7fc606c36e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #5: + 0x8609 (0x7fc60bc7d609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #6: clone + 0x43 (0x7fc60ba48353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default5]:terminate called after throwing an instance of 'c10::DistBackendError' [default5]: what(): [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600069 milliseconds before timing out. [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fc5b9ebe897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fc5bb197c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fc5bb19ca80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fc5bb19ddcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #4: + 0xd3e95 (0x7fc606c36e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #5: + 0x8609 (0x7fc60bc7d609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #6: clone + 0x43 (0x7fc60ba48353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default5]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fc5b9ebe897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: + 0xe32119 (0x7fc5bae21119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: + 0xd3e95 (0x7fc606c36e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #3: + 0x8609 (0x7fc60bc7d609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #4: clone + 0x43 (0x7fc60ba48353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default1]:[rank9]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 0] Timeout at NCCL work: 22, last enqueued NCCL work: 23, last completed NCCL work: 21. [default1]:[rank9]:[E ProcessGroupNCCL.cpp:577] [Rank 0] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default1]:[rank9]:[E ProcessGroupNCCL.cpp:583] [Rank 0] To avoid data inconsistency, we are taking the entire process down. [default1]:[rank9]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600058 milliseconds before timing out. [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7ff1fdc49897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7ff1fef22c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7ff1fef27a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7ff1fef28dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #4: + 0xd3e95 (0x7ff24a9c1e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #5: + 0x8609 (0x7ff24fa08609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #6: clone + 0x43 (0x7ff24f7d3353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:terminate called after throwing an instance of 'c10::DistBackendError' [default1]: what(): [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600058 milliseconds before timing out. [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7ff1fdc49897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7ff1fef22c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7ff1fef27a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7ff1fef28dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #4: + 0xd3e95 (0x7ff24a9c1e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #5: + 0x8609 (0x7ff24fa08609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #6: clone + 0x43 (0x7ff24f7d3353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7ff1fdc49897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: + 0xe32119 (0x7ff1febac119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: + 0xd3e95 (0x7ff24a9c1e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #3: + 0x8609 (0x7ff24fa08609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #4: clone + 0x43 (0x7ff24f7d3353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:[rank41]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600043 milliseconds before timing out. [default1]:[rank57]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600004 milliseconds before timing out. [default1]:[rank25]: Traceback (most recent call last): [default1]:[rank25]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default1]:[rank25]: trainer.train(dataloader) [default1]:[rank25]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default1]:[rank25]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default1]:[rank25]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default1]:[rank25]: outputs = self.pipeline_engine.train_batch_iter( [default1]:[rank25]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default1]:[rank25]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default1]:[rank25]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default1]:[rank25]: grad_accumulator.backward(sum(activations)) [default1]:[rank25]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default1]:[rank25]: result = loss.backward() [default1]:[rank25]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default1]:[rank25]: torch.autograd.backward( [default1]:[rank25]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default1]:[rank25]: _engine_run_backward( [default1]:[rank25]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default1]:[rank25]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default1]:[rank25]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default1]:[rank25]: return user_fn(self, *args) [default1]:[rank25]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default1]:[rank25]: pipeline_state.run_communication() [default1]:[rank25]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default1]:[rank25]: send_activation() [default1]:[rank25]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default1]:[rank25]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default1]:[rank25]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default1]:[rank25]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default1]:[rank25]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default1]:[rank25]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default1]:[rank25]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default1]:[rank25]: dist.send( [default1]:[rank25]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default1]:[rank25]: return func(*args, **kwargs) [default1]:[rank25]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default1]:[rank25]: group.send([tensor], group_dst_rank, tag).wait() [default1]:[rank25]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default3]:[rank3]: Traceback (most recent call last): [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default3]:[rank3]: trainer.train(dataloader) [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default3]:[rank3]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default3]:[rank3]: outputs = self.pipeline_engine.train_batch_iter( [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default3]:[rank3]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default3]:[rank3]: grad_accumulator.backward(sum(activations)) [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default3]:[rank3]: result = loss.backward() [default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default3]:[rank3]: torch.autograd.backward( [default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default3]:[rank3]: _engine_run_backward( [default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default3]:[rank3]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default3]:[rank3]: return user_fn(self, *args) [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default3]:[rank3]: pipeline_state.run_communication() [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default3]:[rank3]: send_activation() [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default3]:[rank3]: self.p2p.send[default0]:[rank32]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600012 milliseconds before timing out. _tensors([self.activation], to_rank=self.to_rank) [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default3]:[rank3]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default3]:[rank3]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default3]:[rank3]: dist.send( [default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default3]:[rank3]: return func(*args, **kwargs) [default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default3]:[rank3]: group.send([tensor], group_dst_rank, tag).wait() [default3]:[rank3]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default3]:[rank19]: Traceback (most recent call last): [default3]:[rank19]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default3]:[rank19]: trainer.train(dataloader) [default3]:[rank19]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default3]:[rank19]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default3]:[rank19]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default3]:[rank19]: outputs = self.pipeline_engine.train_batch_iter( [default3]:[rank19]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default3]:[rank19]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default3]:[rank19]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default3]:[rank19]: grad_accumulator.backward(sum(activations)) [default3]:[rank19]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default3]:[rank19]: result = loss.backward() [default3]:[rank19]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default3]:[rank19]: torch.autograd.backward( [default3]:[rank19]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default3]:[rank19]: _engine_run_backward( [default3]:[rank19]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default3]:[rank19]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default3]:[rank19]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default3]:[rank19]: return user_fn(self, *args) [default3]:[rank19]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default3]:[rank19]: pipeline_state.run_communication() [default3]:[rank19]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default3]:[rank19]: send_activation() [default3]:[rank19]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default3]:[rank19]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default3]:[rank19]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default3]:[rank19]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default3]:[rank19]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default3]:[rank19]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default3]:[rank19]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default3]:[rank19]: dist.send( [default3]:[rank19]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default3]:[rank19]: return func(*args, **kwargs) [default2]:[rank18]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 34, last enqueued NCCL work: 35, last completed NCCL work: 33. [default3]:[rank19]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default2]:[rank18]:[E ProcessGroupNCCL.cpp:577] [Rank 1] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default2]:[rank18]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default2]:[rank18]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600055 milliseconds before timing out. [default2]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f55b71ac897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f55b8485c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f55b848aa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f55b848bdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7f5603f24e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:[rank19]: group.send([tensor], group_dst_rank, tag).wait() [default2]:frame #5: + 0x8609 (0x7f5608f6b609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7f5608d36353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]:[rank19]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default2]: [default2]:terminate called after throwing an instance of 'c10::DistBackendError' [default2]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600055 milliseconds before timing out. [default2]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f55b71ac897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f55b8485c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f55b848aa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f55b848bdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7f5603f24e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #5: + 0x8609 (0x7f5608f6b609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7f5608d36353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default2]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f55b71ac897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: + 0xe32119 (0x7f55b810f119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: + 0xd3e95 (0x7f5603f24e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #3: + 0x8609 (0x7f5608f6b609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #4: clone + 0x43 (0x7f5608d36353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default0]:[rank48]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600081 milliseconds before timing out. [default5]:[rank37]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600006 milliseconds before timing out. [default2]:[rank34]: Traceback (most recent call last): [default2]:[rank34]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default2]:[rank34]: trainer.train(dataloader) [default2]:[rank34]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default2]:[rank34]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default2]:[rank34]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default2]:[rank34]: outputs = self.pipeline_engine.train_batch_iter( [default2]:[rank34]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default2]:[rank34]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default2]:[rank34]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default2]:[rank34]: grad_accumulator.backward(sum(activations)) [default2]:[rank34]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default2]:[rank34]: result = loss.backward() [default2]:[rank34]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default2]:[rank34]: torch.autograd.backward( [default2]:[rank34]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default2]:[rank34]: _engine_run_backward( [default2]:[rank34]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default2]:[rank34]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default2]:[rank34]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default2]:[rank34]: return user_fn(self, *args) [default2]:[rank34]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default2]:[rank34]: pipeline_state.run_communication() [default2]:[rank34]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 172, in run_communication [default2]:[rank34]: self.grads_buffer.append(recv_grad()) [default2]:[rank34]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 50, in __call__ [default2]:[rank34]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default2]:[rank34]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default2]:[rank34]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default2]:[rank34]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default2]:[rank34]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default2]:[rank34]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default2]:[rank34]: dist.recv( [default2]:[rank34]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default2]:[rank34]: return func(*args, **kwargs) [default2]:[rank34]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default2]:[rank34]: pg.recv([tensor], group_src_rank, tag).wait() [default2]:[rank34]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default5]:[rank53]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600071 milliseconds before timing out. [default7]:[rank55]: Traceback (most recent call last): [default7]:[rank55]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default7]:[rank55]: trainer.train(dataloader) [default7]:[rank55]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default7]:[rank55]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default7]:[rank55]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default7]:[rank55]: outputs = self.pipeline_engine.train_batch_iter( [default7]:[rank55]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter [default7]:[rank55]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default7]:[rank55]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default7]:[rank55]: output = model(**micro_batch) [default7]:[rank55]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default7]:[rank55]: return self._call_impl(*args, **kwargs) [default7]:[rank55]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default7]:[rank55]: return forward_call(*args, **kwargs) [default7]:[rank55]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default7]:[rank55]: sharded_logits = self.model( [default7]:[rank55]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default7]:[rank55]: return self._call_impl(*args, **kwargs) [default7]:[rank55]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default7]:[rank55]: return forward_call(*args, **kwargs) [default7]:[rank55]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default7]:[rank55]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default7]:[rank55]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default7]:[rank55]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default7]:[rank55]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default7]:[rank55]: return self._call_impl(*args, **kwargs) [default7]:[rank55]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default7]:[rank55]: return forward_call(*args, **kwargs) [default7]:[rank55]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default7]:[rank55]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default7]:[rank55]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default7]:[rank55]: pipeline_state.run_communication() [default7]:[rank55]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default7]:[rank55]: recv_activation_tensor = recv_activation() [default7]:[rank55]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default7]:[rank55]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default7]:[rank55]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default7]:[rank55]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default7]:[rank55]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default7]:[rank55]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default7]:[rank55]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default7]:[rank55]: dist.recv( [default7]:[rank55]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default7]:[rank55]: return func(*args, **kwargs) [default7]:[rank55]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default7]:[rank55]: pg.recv([tensor], group_src_rank, tag).wait() [default7]:[rank55]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 1. [default6]:[rank6]: Traceback (most recent call last): [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default6]:[rank6]: trainer.train(dataloader) [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default6]:[rank6]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default6]:[rank6]: outputs = self.pipeline_engine.train_batch_iter( [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default6]:[rank6]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default6]:[rank6]: grad_accumulator.backward(sum(activations)) [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default6]:[rank6]: result = loss.backward() [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default6]:[rank6]: torch.autograd.backward( [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default6]:[rank6]: _engine_run_backward( [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default6]:[rank6]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default6]:[rank6]: return user_fn(self, *args) [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default6]:[rank6]: pipeline_state.run_communication() [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default6]:[rank6]: send_activation() [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default6]:[rank6]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default6]:[rank6]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default6]:[rank6]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default6]:[rank6]: dist.send( [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default6]:[rank6]: return func(*args, **kwargs) [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default6]:[rank6]: group.send([tensor], group_dst_rank, tag).wait() [default6]:[rank6]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default6]:[rank22]: Traceback (most recent call last): [default6]:[rank22]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default6]:[rank22]: trainer.train(dataloader) [default6]:[rank22]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default6]:[rank22]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default6]:[rank22]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default6]:[rank22]: outputs = self.pipeline_engine.train_batch_iter( [default6]:[rank22]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default6]:[rank22]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default6]:[rank22]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default6]:[rank22]: grad_accumulator.backward(sum(activations)) [default6]:[rank22]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default6]:[rank22]: result = loss.backward() [default6]:[rank22]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default6]:[rank22]: torch.autograd.backward( [default6]:[rank22]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default6]:[rank22]: _engine_run_backward( [default6]:[rank22]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default6]:[rank22]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default6]:[rank22]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default6]:[rank22]: return user_fn(self, *args) [default6]:[rank22]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default6]:[rank22]: pipeline_state.run_communication() [default6]:[rank22]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default6]:[rank22]: send_activation() [default6]:[rank22]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default6]:[rank22]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default6]:[rank22]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default6]:[rank22]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default6]:[rank22]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default6]:[rank22]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default6]:[rank22]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default6]:[rank22]: dist.send( [default6]:[rank22]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default6]:[rank22]: return func(*args, **kwargs) [default6]:[rank22]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default6]:[rank22]: group.send([tensor], group_dst_rank, tag).wait() [default6]:[rank22]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default7]:[rank15]: Traceback (most recent call last): [default7]:[rank15]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default7]:[rank15]: trainer.train(dataloader) [default7]:[rank15]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default7]:[rank15]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default7]:[rank15]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default7]:[rank15]: outputs = self.pipeline_engine.train_batch_iter( [default7]:[rank15]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default7]:[rank15]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default7]:[rank15]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default7]:[rank15]: grad_accumulator.backward(sum(activations)) [default7]:[rank15]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default7]:[rank15]: result = loss.backward() [default7]:[rank15]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default7]:[rank15]: torch.autograd.backward( [default7]:[rank15]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default7]:[rank15]: _engine_run_backward( [default7]:[rank15]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default7]:[rank15]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default7]:[rank15]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default7]:[rank15]: return user_fn(self, *args) [default7]:[rank15]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default7]:[rank15]: pipeline_state.run_communication() [default7]:[rank15]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default7]:[rank15]: send_activation() [default7]:[rank15]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default7]:[rank15]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default7]:[rank15]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default7]:[rank15]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default7]:[rank15]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default7]:[rank15]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default7]:[rank15]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default7]:[rank15]: dist.send( [default7]:[rank15]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default7]:[rank15]: return func(*args, **kwargs) [default7]:[rank15]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default7]:[rank15]: group.send([tensor], group_dst_rank, tag).wait() [default7]:[rank15]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default0]:[rank24]: Traceback (most recent call last): [default0]:[rank24]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default0]:[rank24]: trainer.train(dataloader) [default0]:[rank24]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default0]:[rank24]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default0]:[rank24]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default0]:[rank24]: outputs = self.pipeline_engine.train_batch_iter( [default0]:[rank24]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default0]:[rank24]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default0]:[rank24]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default0]:[rank24]: grad_accumulator.backward(sum(activations)) [default0]:[rank24]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default0]:[rank24]: result = loss.backward() [default0]:[rank24]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default0]:[rank24]: torch.autograd.backward( [default0]:[rank24]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default0]:[rank24]: _engine_run_backward( [default0]:[rank24]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default0]:[rank24]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default0]:[rank24]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default0]:[rank24]: return user_fn(self, *args) [default0]:[rank24]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default0]:[rank24]: pipeline_state.run_communication() [default0]:[rank24]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default0]:[rank24]: send_activation() [default0]:[rank24]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default0]:[rank24]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default0]:[rank24]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default0]:[rank24]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default0]:[rank24]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default0]:[rank24]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default0]:[rank24]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default0]:[rank24]: dist.send( [default0]:[rank24]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default0]:[rank24]: return func(*args, **kwargs) [default0]:[rank24]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default0]:[rank24]: group.send([tensor], group_dst_rank, tag).wait() [default0]:[rank24]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default5]:[rank29]: Traceback (most recent call last): [default5]:[rank29]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default5]:[rank29]: trainer.train(dataloader) [default5]:[rank29]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default5]:[rank29]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default5]:[rank29]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default5]:[rank29]: outputs = self.pipeline_engine.train_batch_iter( [default5]:[rank29]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default5]:[rank29]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default5]:[rank29]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default5]:[rank29]: grad_accumulator.backward(sum(activations)) [default5]:[rank29]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default5]:[rank29]: result = loss.backward() [default5]:[rank29]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default5]:[rank29]: torch.autograd.backward( [default5]:[rank29]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default5]:[rank29]: _engine_run_backward( [default5]:[rank29]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default5]:[rank29]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default5]:[rank29]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default5]:[rank29]: return user_fn(self, *args) [default5]:[rank29]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default5]:[rank29]: pipeline_state.run_communication() [default5]:[rank29]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default5]:[rank29]: send_activation() [default5]:[rank29]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default5]:[rank29]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default5]:[rank29]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default5]:[rank29]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default5]:[rank29]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default5]:[rank29]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default5]:[rank29]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default5]:[rank29]: dist.send( [default5]:[rank29]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default5]:[rank29]: return func(*args, **kwargs) [default5]:[rank29]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default5]:[rank29]: group.send([tensor], group_dst_rank, tag).wait() [default5]:[rank29]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default1]:[rank41]: Traceback (most recent call last): [default1]:[rank41]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default1]:[rank41]: trainer.train(dataloader) [default1]:[rank41]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default1]:[rank41]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default1]:[rank41]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default1]:[rank41]: outputs = self.pipeline_engine.train_batch_iter( [default1]:[rank41]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default1]:[rank41]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default1]:[rank41]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default1]:[rank41]: grad_accumulator.backward(sum(activations)) [default1]:[rank41]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default1]:[rank41]: result = loss.backward() [default1]:[rank41]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default1]:[rank41]: torch.autograd.backward( [default1]:[rank41]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default1]:[rank41]: _engine_run_backward( [default1]:[rank41]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default1]:[rank41]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default1]:[rank41]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default1]:[rank41]: return user_fn(self, *args) [default1]:[rank41]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default1]:[rank41]: pipeline_state.run_communication() [default1]:[rank41]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 172, in run_communication [default1]:[rank41]: self.grads_buffer.append(recv_grad()) [default1]:[rank41]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 50, in __call__ [default1]:[rank41]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default1]:[rank41]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default1]:[rank41]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default1]:[rank41]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default1]:[rank41]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default1]:[rank41]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default1]:[rank41]: dist.recv( [default1]:[rank41]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default1]:[rank41]: return func(*args, **kwargs) [default1]:[rank41]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default1]:[rank41]: pg.recv([tensor], group_src_rank, tag).wait() [default1]:[rank41]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default4]:[rank28]: Traceback (most recent call last): [default4]:[rank28]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default4]:[rank28]: trainer.train(dataloader) [default4]:[rank28]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default4]:[rank28]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default4]:[rank28]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default4]:[rank28]: outputs = self.pipeline_engine.train_batch_iter( [default4]:[rank28]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default4]:[rank28]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default4]:[rank28]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default4]:[rank28]: grad_accumulator.backward(sum(activations)) [default4]:[rank28]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default4]:[rank28]: result = loss.backward() [default4]:[rank28]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default4]:[rank28]: torch.autograd.backward( [default4]:[rank28]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default4]:[rank28]: _engine_run_backward( [default4]:[rank28]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default4]:[rank28]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default4]:[rank28]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default4]:[rank28]: return user_fn(self, *args) [default4]:[rank28]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default4]:[rank28]: pipeline_state.run_communication() [default4]:[rank28]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default4]:[rank28]: send_activation() [default4]:[rank28]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default4]:[rank28]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default4]:[rank28]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default4]:[rank28]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default4]:[rank28]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default4]:[rank28]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default4]:[rank28]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default4]:[rank28]: dist.send( [default4]:[rank28]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default4]:[rank28]: return func(*args, **kwargs) [default4]:[rank28]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default4]:[rank28]: group.send([tensor], group_dst_rank, tag).wait() [default4]:[rank28]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default6]:[rank14]: Traceback (most recent call last): [default6]:[rank14]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default6]:[rank14]: trainer.train(dataloader) [default6]:[rank14]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default6]:[rank14]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default6]:[rank14]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default6]:[rank14]: outputs = self.pipeline_engine.train_batch_iter( [default6]:[rank14]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default6]:[rank14]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default6]:[rank14]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default6]:[rank14]: grad_accumulator.backward(sum(activations)) [default6]:[rank14]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default6]:[rank14]: result = loss.backward() [default6]:[rank14]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default6]:[rank14]: torch.autograd.backward( [default6]:[rank14]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default6]:[rank14]: _engine_run_backward( [default6]:[rank14]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default6]:[rank14]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default6]:[rank14]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default6]:[rank14]: return user_fn(self, *args) [default6]:[rank14]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default6]:[rank14]: pipeline_state.run_communication() [default6]:[rank14]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default6]:[rank14]: send_activation() [default6]:[rank14]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default6]:[rank14]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default6]:[rank14]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default6]:[rank14]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default6]:[rank14]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default6]:[rank14]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default6]:[rank14]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default6]:[rank14]: dist.send( [default6]:[rank14]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default6]:[rank14]: return func(*args, **kwargs) [default6]:[rank14]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default6]:[rank14]: group.send([tensor], group_dst_rank, tag).wait() [default6]:[rank14]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default5]:[rank13]: Traceback (most recent call last): [default5]:[rank13]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default5]:[rank13]: trainer.train(dataloader) [default5]:[rank13]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default5]:[rank13]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default5]:[rank13]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default5]:[rank13]: outputs = self.pipeline_engine.train_batch_iter( [default5]:[rank13]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default5]:[rank13]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default5]:[rank13]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default5]:[rank13]: grad_accumulator.backward(sum(activations)) [default5]:[rank13]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default5]:[rank13]: result = loss.backward() [default5]:[rank13]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default5]:[rank13]: torch.autograd.backward( [default5]:[rank13]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default5]:[rank13]: _engine_run_backward( [default5]:[rank13]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default5]:[rank13]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default5]:[rank13]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default5]:[rank13]: return user_fn(self, *args) [default5]:[rank13]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default5]:[rank13]: pipeline_state.run_communication() [default5]:[rank13]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default5]:[rank13]: send_activation() [default5]:[rank13]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default5]:[rank13]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default5]:[rank13]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default5]:[rank13]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default5]:[rank13]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default5]:[rank13]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default5]:[rank13]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default5]:[rank13]: dist.send( [default5]:[rank13]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default5]:[rank13]: return func(*args, **kwargs) [default5]:[rank13]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default5]:[rank13]: group.send([tensor], group_dst_rank, tag).wait() [default5]:[rank13]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default3]:[rank27]: Traceback (most recent call last): [default3]:[rank27]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default3]:[rank27]: trainer.train(dataloader) [default3]:[rank27]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default3]:[rank27]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default3]:[rank27]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default3]:[rank27]: outputs = self.pipeline_engine.train_batch_iter( [default3]:[rank27]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default3]:[rank27]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default3]:[rank27]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default3]:[rank27]: grad_accumulator.backward(sum(activations)) [default3]:[rank27]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default3]:[rank27]: result = loss.backward() [default3]:[rank27]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default3]:[rank27]: torch.autograd.backward( [default3]:[rank27]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default3]:[rank27]: _engine_run_backward( [default3]:[rank27]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default3]:[rank27]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default3]:[rank27]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default3]:[rank27]: return user_fn(self, *args) [default3]:[rank27]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default3]:[rank27]: pipeline_state.run_communication() [default3]:[rank27]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default3]:[rank27]: send_activation() [default3]:[rank27]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default3]:[rank27]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default3]:[rank27]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default3]:[rank27]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default3]:[rank27]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default3]:[rank27]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default3]:[rank27]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default3]:[rank27]: dist.send( [default3]:[rank27]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default3]:[rank27]: return func(*args, **kwargs) [default3]:[rank27]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default3]:[rank27]: group.send([tensor], group_dst_rank, tag).wait() [default3]:[rank27]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default0]:[rank48]: Traceback (most recent call last): [default0]:[rank48]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default0]:[rank48]: trainer.train(dataloader) [default0]:[rank48]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default0]:[rank48]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default0]:[rank48]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default0]:[rank48]: outputs = self.pipeline_engine.train_batch_iter( [default0]:[rank48]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter [default0]:[rank48]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default0]:[rank48]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default0]:[rank48]: output = model(**micro_batch) [default0]:[rank48]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default0]:[rank48]: return self._call_impl(*args, **kwargs) [default0]:[rank48]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default0]:[rank48]: return forward_call(*args, **kwargs) [default0]:[rank48]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default0]:[rank48]: sharded_logits = self.model( [default0]:[rank48]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default0]:[rank48]: return self._call_impl(*args, **kwargs) [default0]:[rank48]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default0]:[rank48]: return forward_call(*args, **kwargs) [default0]:[rank48]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default0]:[rank48]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default0]:[rank48]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default0]:[rank48]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default0]:[rank48]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default0]:[rank48]: return self._call_impl(*args, **kwargs) [default0]:[rank48]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default0]:[rank48]: return forward_call(*args, **kwargs) [default0]:[rank48]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default0]:[rank48]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default0]:[rank48]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default0]:[rank48]: pipeline_state.run_communication() [default0]:[rank48]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default0]:[rank48]: recv_activation_tensor = recv_activation() [default0]:[rank48]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default0]:[rank48]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default0]:[rank48]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default0]:[rank48]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default0]:[rank48]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default0]:[rank48]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default0]:[rank48]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default0]:[rank48]: dist.recv( [default0]:[rank48]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default0]:[rank48]: return func(*args, **kwargs) [default0]:[rank48]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default0]:[rank48]: pg.recv([tensor], group_src_rank, tag).wait() [default0]:[rank48]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 1. [default4]:[rank36]: Traceback (most recent call last): [default4]:[rank36]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default4]:[rank36]: trainer.train(dataloader) [default4]:[rank36]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default4]:[rank36]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default4]:[rank36]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default4]:[rank36]: outputs = self.pipeline_engine.train_batch_iter( [default4]:[rank36]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default4]:[rank36]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default4]:[rank36]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default4]:[rank36]: grad_accumulator.backward(sum(activations)) [default4]:[rank36]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default4]:[rank36]: result = loss.backward() [default4]:[rank36]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default4]:[rank36]: torch.autograd.backward( [default4]:[rank36]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default4]:[rank36]: _engine_run_backward( [default4]:[rank36]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default4]:[rank36]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default4]:[rank36]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default4]:[rank36]: return user_fn(self, *args) [default4]:[rank36]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default4]:[rank36]: pipeline_state.run_communication() [default4]:[rank36]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 172, in run_communication [default4]:[rank36]: self.grads_buffer.append(recv_grad()) [default4]:[rank36]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 50, in __call__ [default4]:[rank36]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default4]:[rank36]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default4]:[rank36]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default4]:[rank36]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default4]:[rank36]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default4]:[rank36]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default4]:[rank36]: dist.recv( [default4]:[rank36]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default4]:[rank36]: return func(*args, **kwargs) [default4]:[rank36]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default4]:[rank36]: pg.recv([tensor], group_src_rank, tag).wait() [default4]:[rank36]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default1]:[rank57]: Traceback (most recent call last): [default1]:[rank57]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default1]:[rank57]: trainer.train(dataloader) [default1]:[rank57]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default1]:[rank57]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default1]:[rank57]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default1]:[rank57]: outputs = self.pipeline_engine.train_batch_iter( [default1]:[rank57]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter [default1]:[rank57]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default1]:[rank57]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default1]:[rank57]: output = model(**micro_batch) [default1]:[rank57]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default1]:[rank57]: return self._call_impl(*args, **kwargs) [default1]:[rank57]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default1]:[rank57]: return forward_call(*args, **kwargs) [default1]:[rank57]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default1]:[rank57]: sharded_logits = self.model( [default1]:[rank57]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default1]:[rank57]: return self._call_impl(*args, **kwargs) [default1]:[rank57]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default1]:[rank57]: return forward_call(*args, **kwargs) [default1]:[rank57]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default1]:[rank57]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default1]:[rank57]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default1]:[rank57]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default1]:[rank57]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default1]:[rank57]: return self._call_impl(*args, **kwargs) [default1]:[rank57]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default1]:[rank57]: return forward_call(*args, **kwargs) [default1]:[rank57]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default1]:[rank57]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default1]:[rank57]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default1]:[rank57]: pipeline_state.run_communication() [default1]:[rank57]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default1]:[rank57]: recv_activation_tensor = recv_activation() [default1]:[rank57]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default1]:[rank57]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default1]:[rank57]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default1]:[rank57]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default1]:[rank57]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default1]:[rank57]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default1]:[rank57]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default1]:[rank57]: dist.recv( [default1]:[rank57]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default1]:[rank57]: return func(*args, **kwargs) [default1]:[rank57]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default1]:[rank57]: pg.recv([tensor], group_src_rank, tag).wait() [default1]:[rank57]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 1. [default7]:[rank55]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 14, last enqueued NCCL work: 15, last completed NCCL work: 13. [default7]:[rank55]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default7]:[rank55]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default7]:[rank55]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600005 milliseconds before timing out. [default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f62be830897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f62bfb09c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f62bfb0ea80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f62bfb0fdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #4: + 0xd3e95 (0x7f630b5a8e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #5: + 0x8609 (0x7f63105ef609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #6: clone + 0x43 (0x7f63103ba353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default7]:terminate called after throwing an instance of 'c10::DistBackendError' [default7]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600005 milliseconds before timing out. [default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f62be830897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f62bfb09c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f62bfb0ea80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f62bfb0fdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #4: + 0xd3e95 (0x7f630b5a8e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #5: + 0x8609 (0x7f63105ef609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #6: clone + 0x43 (0x7f63103ba353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default7]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f62be830897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: + 0xe32119 (0x7f62bf793119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: + 0xd3e95 (0x7f630b5a8e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #3: + 0x8609 (0x7f63105ef609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #4: clone + 0x43 (0x7f63103ba353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default6]:[rank46]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600042 milliseconds before timing out. [default5]:[rank61]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600019 milliseconds before timing out. [default6]:[rank62]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600006 milliseconds before timing out. [default3]:[rank51]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600058 milliseconds before timing out. [default7]:[rank7]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 0] Timeout at NCCL work: 22, last enqueued NCCL work: 23, last completed NCCL work: 21. [default7]:[rank7]:[E ProcessGroupNCCL.cpp:577] [Rank 0] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default7]:[rank7]:[E ProcessGroupNCCL.cpp:583] [Rank 0] To avoid data inconsistency, we are taking the entire process down. [default7]:[rank7]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600098 milliseconds before timing out. [default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f359b47c897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f359c755c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f359c75aa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f359c75bdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #4: + 0xd3e95 (0x7f35e81f4e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #5: + 0x8609 (0x7f35ed23b609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #6: clone + 0x43 (0x7f35ed006353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default7]:terminate called after throwing an instance of 'c10::DistBackendError' [default7]: what(): [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600098 milliseconds before timing out. [default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f359b47c897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f359c755c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f359c75aa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f359c75bdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #4: + 0xd3e95 (0x7f35e81f4e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #5: + 0x8609 (0x7f35ed23b609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #6: clone + 0x43 (0x7f35ed006353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default7]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f359b47c897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: + 0xe32119 (0x7f359c3df119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: + 0xd3e95 (0x7f35e81f4e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #3: + 0x8609 (0x7f35ed23b609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #4: clone + 0x43 (0x7f35ed006353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default7]:[rank39]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 26, last enqueued NCCL work: 27, last completed NCCL work: 25. [default7]:[rank39]:[E ProcessGroupNCCL.cpp:577] [Rank 2] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default7]:[rank39]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default7]:[rank39]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600037 milliseconds before timing out. [default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f71e471b897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f71e59f4c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f71e59f9a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f71e59fadcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #4: + 0xd3e95 (0x7f7231493e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #5: + 0x8609 (0x7f72364da609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #6: clone + 0x43 (0x7f72362a5353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default7]:terminate called after throwing an instance of 'c10::DistBackendError' [default7]: what(): [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600037 milliseconds before timing out. [default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f71e471b897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f71e59f4c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f71e59f9a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f71e59fadcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #4: + 0xd3e95 (0x7f7231493e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #5: + 0x8609 (0x7f72364da609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #6: clone + 0x43 (0x7f72362a5353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default7]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f71e471b897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: + 0xe32119 (0x7f71e567e119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: + 0xd3e95 (0x7f7231493e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #3: + 0x8609 (0x7f72364da609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #4: clone + 0x43 (0x7f72362a5353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default5]:[rank37]: Traceback (most recent call last): [default5]:[rank37]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default3]:[rank35]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600077 milliseconds before timing out. [default5]:[rank37]: trainer.train(dataloader) [default5]:[rank37]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default5]:[rank37]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default5]:[rank37]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default5]:[rank37]: outputs = self.pipeline_engine.train_batch_iter( [default5]:[rank37]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default5]:[rank37]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default5]:[rank37]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default5]:[rank37]: grad_accumulator.backward(sum(activations)) [default5]:[rank37]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default5]:[rank37]: result = loss.backward() [default5]:[rank37]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default5]:[rank37]: torch.autograd.backward( [default5]:[rank37]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default5]:[rank37]: _engine_run_backward( [default5]:[rank37]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default5]:[rank37]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default5]:[rank37]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default5]:[rank37]: return user_fn(self, *args) [default5]:[rank37]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default5]:[rank37]: pipeline_state.run_communication() [default5]:[rank37]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 172, in run_communication [default5]:[rank37]: self.grads_buffer.append(recv_grad()) [default5]:[rank37]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 50, in __call__ [default5]:[rank37]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default5]:[rank37]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default5]:[rank37]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default5]:[rank37]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default5]:[rank37]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default5]:[rank37]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default5]:[rank37]: dist.recv( [default5]:[rank37]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default5]:[rank37]: return func(*args, **kwargs) [default5]:[rank37]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default5]:[rank37]: pg.recv([tensor], group_src_rank, tag).wait() [default5]:[rank37]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default5]:[rank45]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600074 milliseconds before timing out. [default3]:[rank3]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 0] Timeout at NCCL work: 22, last enqueued NCCL work: 23, last completed NCCL work: 21. [default3]:[rank3]:[E ProcessGroupNCCL.cpp:577] [Rank 0] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default3]:[rank3]:[E ProcessGroupNCCL.cpp:583] [Rank 0] To avoid data inconsistency, we are taking the entire process down. [default3]:[rank3]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600022 milliseconds before timing out. [default3]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f2ab7f72897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f2ab924bc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f2ab9250a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f2ab9251dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #4: + 0xd3e95 (0x7f2b04ceae95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #5: + 0x8609 (0x7f2b09d31609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #6: clone + 0x43 (0x7f2b09afc353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default3]:terminate called after throwing an instance of 'c10::DistBackendError' [default3]: what(): [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600022 milliseconds before timing out. [default3]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f2ab7f72897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f2ab924bc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f2ab9250a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f2ab9251dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #4: + 0xd3e95 (0x7f2b04ceae95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #5: + 0x8609 (0x7f2b09d31609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #6: clone + 0x43 (0x7f2b09afc353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default3]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f2ab7f72897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: + 0xe32119 (0x7f2ab8ed5119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: + 0xd3e95 (0x7f2b04ceae95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #3: + 0x8609 (0x7f2b09d31609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #4: clone + 0x43 (0x7f2b09afc353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default2]:[rank58]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600019 milliseconds before timing out. [default2]:[rank42]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600049 milliseconds before timing out. [default4]:[rank52]: Traceback (most recent call last): [default4]:[rank52]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default4]:[rank52]: trainer.train(dataloader) [default4]:[rank52]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default4]:[rank52]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default4]:[rank52]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default4]:[rank52]: outputs = self.pipeline_engine.train_batch_iter( [default4]:[rank52]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter [default4]:[rank52]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default4]:[rank52]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default4]:[rank52]: output = model(**micro_batch) [default4]:[rank52]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default4]:[rank52]: return self._call_impl(*args, **kwargs) [default4]:[rank52]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default4]:[rank52]: return forward_call(*args, **kwargs) [default4]:[rank52]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default4]:[rank52]: sharded_logits = self.model( [default4]:[rank52]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default4]:[rank52]: return self._call_impl(*args, **kwargs) [default4]:[rank52]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default4]:[rank52]: return forward_call(*args, **kwargs) [default4]:[rank52]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default4]:[rank52]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default4]:[rank52]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default4]:[rank52]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default4]:[rank52]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default4]:[rank52]: return self._call_impl(*args, **kwargs) [default4]:[rank52]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default4]:[rank52]: return forward_call(*args, **kwargs) [default4]:[rank52]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default4]:[rank52]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default4]:[rank52]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default4]:[rank52]: pipeline_state.run_communication() [default4]:[rank52]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default4]:[rank52]: recv_activation_tensor = recv_activation() [default4]:[rank52]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default4]:[rank52]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default4]:[rank52]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default4]:[rank52]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default4]:[rank52]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default4]:[rank52]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default4]:[rank52]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default4]:[rank52]: dist.recv( [default4]:[rank52]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default4]:[rank52]: return func(*args, **kwargs) [default4]:[rank52]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default4]:[rank52]: pg.recv([tensor], group_src_rank, tag).wait() [default4]:[rank52]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 1. [default2]:[rank50]: Traceback (most recent call last): [default2]:[rank50]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default2]:[rank50]: trainer.train(dataloader) [default2]:[rank50]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default2]:[rank50]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default2]:[rank50]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default2]:[rank50]: outputs = self.pipeline_engine.train_batch_iter( [default2]:[rank50]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter [default2]:[rank50]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default2]:[rank50]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default2]:[rank50]: output = model(**micro_batch) [default2]:[rank50]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default2]:[rank50]: return self._call_impl(*args, **kwargs) [default2]:[rank50]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default2]:[rank50]: return forward_call(*args, **kwargs) [default2]:[rank50]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default2]:[rank50]: sharded_logits = self.model( [default2]:[rank50]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default2]:[rank50]: return self._call_impl(*args, **kwargs) [default2]:[rank50]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default2]:[rank50]: return forward_call(*args, **kwargs) [default2]:[rank50]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default2]:[rank50]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default2]:[rank50]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default2]:[rank50]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default2]:[rank50]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default2]:[rank50]: return self._call_impl(*args, **kwargs) [default2]:[rank50]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default2]:[rank50]: return forward_call(*args, **kwargs) [default2]:[rank50]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default2]:[rank50]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default2]:[rank50]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default2]:[rank50]: pipeline_state.run_communication() [default2]:[rank50]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default2]:[rank50]: recv_activation_tensor = recv_activation() [default2]:[rank50]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default2]:[rank50]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default2]:[rank50]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default2]:[rank50]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default2]:[rank50]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default2]:[rank50]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default2]:[rank50]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default2]:[rank50]: dist.recv( [default2]:[rank50]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default2]:[rank50]: return func(*args, **kwargs) [default2]:[rank50]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default2]:[rank50]: pg.recv([tensor], group_src_rank, tag).wait() [default2]:[rank50]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 1. [default6]:[rank6]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 0] Timeout at NCCL work: 22, last enqueued NCCL work: 23, last completed NCCL work: 21. [default6]:[rank6]:[E ProcessGroupNCCL.cpp:577] [Rank 0] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default6]:[rank6]:[E ProcessGroupNCCL.cpp:583] [Rank 0] To avoid data inconsistency, we are taking the entire process down. [default6]:[rank6]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600030 milliseconds before timing out. [default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f3eb2dab897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f3eb4084c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f3eb4089a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f3eb408adcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7f3effb23e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7f3f04b6a609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7f3f04935353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:[rank38]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600004 milliseconds before timing out. [default0]:[rank8]: Traceback (most recent call last): [default0]:[rank8]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default0]:[rank8]: trainer.train(dataloader) [default0]:[rank8]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default0]:[rank8]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default0]:[rank8]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default0]:[rank8]: outputs = self.pipeline_engine.train_batch_iter( [default0]:[rank8]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default0]:[rank8]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default0]:[rank8]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default0]:[rank8]: grad_accumulator.backward(sum(activations)) [default0]:[rank8]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default0]:[rank8]: result = loss.backward() [default0]:[rank8]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default0]:[rank8]: torch.autograd.backward( [default0]:[rank8]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default0]:[rank8]: _engine_run_backward( [default0]:[rank8]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default0]:[rank8]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default0]:[rank8]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default0]:[rank8]: return user_fn(self, *args) [default0]:[rank8]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default0]:[rank8]: pipeline_state.run_communication() [default0]:[rank8]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default0]:[rank8]: send_activation() [default0]:[rank8]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default0]:[rank8]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default0]:[rank8]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default0]:[rank8]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default0]:[rank8]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default0]:[rank8]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default0]:[rank8]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default0]:[rank8]: dist.send( [default0]:[rank8]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default0]:[rank8]: return func(*args, **kwargs) [default0]:[rank8]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default0]:[rank8]: group.send([tensor], group_dst_rank, tag).wait() [default0]:[rank8]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default6]:terminate called after throwing an instance of 'c10::DistBackendError' [default6]: what(): [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600030 milliseconds before timing out. [default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f3eb2dab897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f3eb4084c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f3eb4089a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f3eb408adcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7f3effb23e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7f3f04b6a609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7f3f04935353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f3eb2dab897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: + 0xe32119 (0x7f3eb3d0e119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: + 0xd3e95 (0x7f3effb23e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #3: + 0x8609 (0x7f3f04b6a609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #4: clone + 0x43 (0x7f3f04935353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default2]:[rank26]: Traceback (most recent call last): [default2]:[rank26]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default2]:[rank26]: trainer.train(dataloader) [default2]:[rank26]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default2]:[rank26]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default2]:[rank26]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default2]:[rank26]: outputs = self.pipeline_engine.train_batch_iter( [default2]:[rank26]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default2]:[rank26]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default2]:[rank26]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default2]:[rank26]: grad_accumulator.backward(sum(activations)) [default2]:[rank26]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default2]:[rank26]: result = loss.backward() [default2]:[rank26]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default2]:[rank26]: torch.autograd.backward( [default2]:[rank26]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default2]:[rank26]: _engine_run_backward( [default2]:[rank26]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default2]:[rank26]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default2]:[rank26]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default2]:[rank26]: return user_fn(self, *args) [default2]:[rank26]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default2]:[rank26]: pipeline_state.run_communication() [default2]:[rank26]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default2]:[rank26]: send_activation() [default2]:[rank26]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default2]:[rank26]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default2]:[rank26]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default2]:[rank26]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default2]:[rank26]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default2]:[rank26]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default2]:[rank26]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default2]:[rank26]: dist.send( [default2]:[rank26]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default2]:[rank26]: return func(*args, **kwargs) [default2]:[rank26]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default2]:[rank26]: group.send([tensor], group_dst_rank, tag).wait() [default2]:[rank26]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default6]:[rank54]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600075 milliseconds before timing out. [default5]:[rank45]: Traceback (most recent call last): [default5]:[rank45]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default5]:[rank45]: trainer.train(dataloader) [default5]:[rank45]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default5]:[rank45]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default5]:[rank45]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default5]:[rank45]: outputs = self.pipeline_engine.train_batch_iter( [default5]:[rank45]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default5]:[rank45]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default5]:[rank45]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default5]:[rank45]: grad_accumulator.backward(sum(activations)) [default5]:[rank45]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default5]:[rank45]: result = loss.backward() [default5]:[rank45]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default5]:[rank45]: torch.autograd.backward( [default5]:[rank45]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default5]:[rank45]: _engine_run_backward( [default5]:[rank45]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default5]:[rank45]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default5]:[rank45]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default5]:[rank45]: return user_fn(self, *args) [default5]:[rank45]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default5]:[rank45]: pipeline_state.run_communication() [default5]:[rank45]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 172, in run_communication [default5]:[rank45]: self.grads_buffer.append(recv_grad()) [default5]:[rank45]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 50, in __call__ [default5]:[rank45]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default5]:[rank45]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default5]:[rank45]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default5]:[rank45]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default5]:[rank45]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default5]:[rank45]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default5]:[rank45]: dist.recv( [default5]:[rank45]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default5]:[rank45]: return func(*args, **kwargs) [default5]:[rank45]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default5]:[rank45]: pg.recv([tensor], group_src_rank, tag).wait() [default5]:[rank45]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default6]:[rank30]: Traceback (most recent call last): [default6]:[rank30]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default6]:[rank30]: trainer.train(dataloader) [default6]:[rank30]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default6]:[rank30]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default6]:[rank30]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default6]:[rank30]: outputs = self.pipeline_engine.train_batch_iter( [default6]:[rank30]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default6]:[rank30]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default6]:[rank30]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default6]:[rank30]: grad_accumulator.backward(sum(activations)) [default6]:[rank30]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default6]:[rank30]: result = loss.backward() [default6]:[rank30]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default6]:[rank30]: torch.autograd.backward( [default6]:[rank30]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default6]:[rank30]: _engine_run_backward( [default6]:[rank30]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default6]:[rank30]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default6]:[rank30]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default6]:[rank30]: return user_fn(self, *args) [default6]:[rank30]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default6]:[rank30]: pipeline_state.run_communication() [default6]:[rank30]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default6]:[rank30]: send_activation() [default6]:[rank30]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default6]:[rank30]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default6]:[rank30]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default6]:[rank30]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default6]:[rank30]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default6]:[rank30]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default6]:[rank30]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default6]:[rank30]: dist.send( [default6]:[rank30]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default6]:[rank30]: return func(*args, **kwargs) [default6]:[rank30]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default6]:[rank30]: group.send([tensor], group_dst_rank, tag).wait() [default6]:[rank30]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default7]:[rank23]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 34, last enqueued NCCL work: 35, last completed NCCL work: 33. [default7]:[rank23]:[E ProcessGroupNCCL.cpp:577] [Rank 1] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default7]:[rank23]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default7]:[rank23]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600022 milliseconds before timing out. [default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f878368d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f8784966c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f878496ba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f878496cdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #4: + 0xd3e95 (0x7f87d0405e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #5: + 0x8609 (0x7f87d544c609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #6: clone + 0x43 (0x7f87d5217353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default7]:terminate called after throwing an instance of 'c10::DistBackendError' [default7]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600022 milliseconds before timing out. [default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f878368d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f8784966c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f878496ba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f878496cdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #4: + 0xd3e95 (0x7f87d0405e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #5: + 0x8609 (0x7f87d544c609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #6: clone + 0x43 (0x7f87d5217353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default7]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f878368d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: + 0xe32119 (0x7f87845f0119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: + 0xd3e95 (0x7f87d0405e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #3: + 0x8609 (0x7f87d544c609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #4: clone + 0x43 (0x7f87d5217353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default3]:[rank11]: Traceback (most recent call last): [default3]:[rank11]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default3]:[rank11]: trainer.train(dataloader) [default3]:[rank11]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default3]:[rank11]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default3]:[rank11]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default3]:[rank11]: outputs = self.pipeline_engine.train_batch_iter( [default3]:[rank11]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default3]:[rank11]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default3]:[rank11]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default3]:[rank11]: grad_accumulator.backward(sum(activations)) [default3]:[rank11]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default3]:[rank11]: result = loss.backward() [default3]:[rank11]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default3]:[rank11]: torch.autograd.backward( [default3]:[rank11]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default3]:[rank11]: _engine_run_backward( [default3]:[rank11]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default3]:[rank11]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default3]:[rank11]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default3]:[rank11]: return user_fn(self, *args) [default3]:[rank11]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default3]:[rank11]: pipeline_state.run_communication() [default3]:[rank11]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default3]:[rank11]: send_activation() [default3]:[rank11]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default3]:[rank11]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default3]:[rank11]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default3]:[rank11]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default3]:[rank11]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default3]:[rank11]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default3]:[rank11]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default3]:[rank11]: dist.send( [default3]:[rank11]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default3]:[rank11]: return func(*args, **kwargs) [default3]:[rank11]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default3]:[rank11]: group.send([tensor], group_dst_rank, tag).wait() [default3]:[rank11]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default4]:[rank12]: Traceback (most recent call last): [default4]:[rank12]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default4]:[rank12]: trainer.train(dataloader) [default4]:[rank12]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default4]:[rank12]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default4]:[rank12]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default4]:[rank12]: outputs = self.pipeline_engine.train_batch_iter( [default4]:[rank12]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default4]:[rank12]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default4]:[rank12]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default2]:[rank10]: Traceback (most recent call last): [default2]:[rank10]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default2]:[rank10]: trainer.train(dataloader) [default4]:[rank12]: grad_accumulator.backward(sum(activations)) [default4]:[rank12]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default2]:[rank10]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default2]:[rank10]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default4]:[rank12]: result = loss.backward() [default4]:[rank12]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default4]:[rank12]: torch.autograd.backward( [default4]:[rank12]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default2]:[rank10]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default4]:[rank12]: _engine_run_backward( [default4]:[rank12]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default2]:[rank10]: outputs = self.pipeline_engine.train_batch_iter( [default2]:[rank10]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default4]:[rank12]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default4]:[rank12]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default4]:[rank12]: return user_fn(self, *args) [default2]:[rank10]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default4]:[rank12]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default4]:[rank12]: pipeline_state.run_communication() [default4]:[rank12]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default4]:[rank12]: send_activation() [default4]:[rank12]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default2]:[rank10]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default2]:[rank10]: grad_accumulator.backward(sum(activations)) [default2]:[rank10]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default4]:[rank12]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default2]:[rank10]: result = loss.backward() [default4]:[rank12]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default2]:[rank10]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default4]:[rank12]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default2]:[rank10]: torch.autograd.backward( [default2]:[rank10]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default4]:[rank12]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default4]:[rank12]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default4]:[rank12]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default2]:[rank10]: _engine_run_backward( [default4]:[rank12]: dist.send( [default4]:[rank12]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default4]:[rank12]: return func(*args, **kwargs) [default4]:[rank12]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default4]:[rank12]: group.send([tensor], group_dst_rank, tag).wait() [default2]:[rank10]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default2]:[rank10]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default4]:[rank12]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default2]:[rank10]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default2]:[rank10]: return user_fn(self, *args) [default2]:[rank10]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default2]:[rank10]: pipeline_state.run_communication() [default2]:[rank10]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default2]:[rank10]: send_activation() [default2]:[rank10]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default2]:[rank10]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default2]:[rank10]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default2]:[rank10]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default2]:[rank10]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default2]:[rank10]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default2]:[rank10]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default2]:[rank10]: dist.send( [default2]:[rank10]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default2]:[rank10]: return func(*args, **kwargs) [default2]:[rank10]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default2]:[rank10]: group.send([tensor], group_dst_rank, tag).wait() [default2]:[rank10]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default1]:[rank57]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 14, last enqueued NCCL work: 15, last completed NCCL work: 13. [default1]:[rank57]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default1]:[rank57]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default1]:[rank57]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600004 milliseconds before timing out. [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f8aed3a5897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f8aee67ec62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f8aee683a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f8aee684dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #4: + 0xd3e95 (0x7f8b3a11de95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #5: + 0x8609 (0x7f8b3f164609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #6: clone + 0x43 (0x7f8b3ef2f353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:terminate called after throwing an instance of 'c10::DistBackendError' [default1]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600004 milliseconds before timing out. [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f8aed3a5897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f8aee67ec62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f8aee683a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f8aee684dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #4: + 0xd3e95 (0x7f8b3a11de95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #5: + 0x8609 (0x7f8b3f164609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #6: clone + 0x43 (0x7f8b3ef2f353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f8aed3a5897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: + 0xe32119 (0x7f8aee308119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: + 0xd3e95 (0x7f8b3a11de95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #3: + 0x8609 (0x7f8b3f164609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #4: clone + 0x43 (0x7f8b3ef2f353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default7]:[rank31]: Traceback (most recent call last): [default7]:[rank31]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default7]:[rank31]: trainer.train(dataloader) [default7]:[rank31]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default7]:[rank31]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default7]:[rank31]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default7]:[rank31]: outputs = self.pipeline_engine.train_batch_iter( [default7]:[rank31]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default7]:[rank31]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default7]:[rank31]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default7]:[rank31]: grad_accumulator.backward(sum(activations)) [default7]:[rank31]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default7]:[rank31]: result = loss.backward() [default7]:[rank31]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default7]:[rank31]: torch.autograd.backward( [default7]:[rank31]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default7]:[rank31]: _engine_run_backward( [default7]:[rank31]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default7]:[rank31]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default7]:[rank31]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default7]:[rank31]: return user_fn(self, *args) [default7]:[rank31]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default7]:[rank31]: pipeline_state.run_communication() [default7]:[rank31]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 142, in run_communication [default7]:[rank31]: send_activation() [default7]:[rank31]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 22, in __call__ [default7]:[rank31]: self.p2p.send_tensors([self.activation], to_rank=self.to_rank) [default7]:[rank31]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 348, in send_tensors [default7]:[rank31]: futures = self.isend_tensors(tensors=tensors, to_rank=to_rank, tag=tag) [default7]:[rank31]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 295, in isend_tensors [default7]:[rank31]: self._send_meta(tensor, to_rank=to_rank, tag=tag) [default7]:[rank31]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 221, in _send_meta [default7]:[rank31]: dist.send( [default7]:[rank31]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default7]:[rank31]: return func(*args, **kwargs) [default7]:[rank31]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1886, in send [default7]:[rank31]: group.send([tensor], group_dst_rank, tag).wait() [default7]:[rank31]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default4]:[rank20]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 34, last enqueued NCCL work: 35, last completed NCCL work: 33. [default4]:[rank20]:[E ProcessGroupNCCL.cpp:577] [Rank 1] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default4]:[rank20]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default4]:[rank20]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600034 milliseconds before timing out. [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f8e6d56c897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f8e6e845c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f8e6e84aa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f8e6e84bdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f8eba2e4e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f8ebf32b609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f8ebf0f6353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default4]:terminate called after throwing an instance of 'c10::DistBackendError' [default4]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600034 milliseconds before timing out. [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f8e6d56c897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f8e6e845c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f8e6e84aa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f8e6e84bdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f8eba2e4e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f8ebf32b609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f8ebf0f6353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default4]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f8e6d56c897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: + 0xe32119 (0x7f8e6e4cf119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: + 0xd3e95 (0x7f8eba2e4e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #3: + 0x8609 (0x7f8ebf32b609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #4: clone + 0x43 (0x7f8ebf0f6353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default0]:[rank32]: Traceback (most recent call last): [default0]:[rank32]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default0]:[rank32]: trainer.train(dataloader) [default0]:[rank32]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default0]:[rank32]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default0]:[rank32]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default0]:[rank32]: outputs = self.pipeline_engine.train_batch_iter( [default0]:[rank32]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default0]:[rank32]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default0]:[rank32]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default0]:[rank32]: grad_accumulator.backward(sum(activations)) [default0]:[rank32]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default0]:[rank32]: result = loss.backward() [default0]:[rank32]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default0]:[rank32]: torch.autograd.backward( [default0]:[rank32]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default0]:[rank32]: _engine_run_backward( [default0]:[rank32]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default0]:[rank32]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default0]:[rank32]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default0]:[rank32]: return user_fn(self, *args) [default0]:[rank32]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default0]:[rank32]: pipeline_state.run_communication() [default0]:[rank32]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 172, in run_communication [default0]:[rank32]: self.grads_buffer.append(recv_grad()) [default0]:[rank32]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 50, in __call__ [default0]:[rank32]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default0]:[rank32]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default0]:[rank32]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default0]:[rank32]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default0]:[rank32]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default0]:[rank32]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default0]:[rank32]: dist.recv( [default0]:[rank32]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default0]:[rank32]: return func(*args, **kwargs) [default0]:[rank32]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default0]:[rank32]: pg.recv([tensor], group_src_rank, tag).wait() [default0]:[rank32]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default6]:[rank14]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 0] Timeout at NCCL work: 22, last enqueued NCCL work: 23, last completed NCCL work: 21. [default6]:[rank14]:[E ProcessGroupNCCL.cpp:577] [Rank 0] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default6]:[rank14]:[E ProcessGroupNCCL.cpp:583] [Rank 0] To avoid data inconsistency, we are taking the entire process down. [default6]:[rank14]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600088 milliseconds before timing out. [default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fb666469897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fb667742c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fb667747a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fb667748dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7fb6b31e1e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7fb6b8228609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7fb6b7ff3353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:terminate called after throwing an instance of 'c10::DistBackendError' [default6]: what(): [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600088 milliseconds before timing out. [default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fb666469897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fb667742c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fb667747a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fb667748dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7fb6b31e1e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7fb6b8228609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7fb6b7ff3353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fb666469897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: + 0xe32119 (0x7fb6673cc119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: + 0xd3e95 (0x7fb6b31e1e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #3: + 0x8609 (0x7fb6b8228609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #4: clone + 0x43 (0x7fb6b7ff3353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default5]:[rank13]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 0] Timeout at NCCL work: 22, last enqueued NCCL work: 23, last completed NCCL work: 21. [default7]:[rank15]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 0] Timeout at NCCL work: 22, last enqueued NCCL work: 23, last completed NCCL work: 21. [default5]:[rank13]:[E ProcessGroupNCCL.cpp:577] [Rank 0] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default7]:[rank15]:[E ProcessGroupNCCL.cpp:577] [Rank 0] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default5]:[rank13]:[E ProcessGroupNCCL.cpp:583] [Rank 0] To avoid data inconsistency, we are taking the entire process down. [default7]:[rank15]:[E ProcessGroupNCCL.cpp:583] [Rank 0] To avoid data inconsistency, we are taking the entire process down. [default5]:[rank13]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600092 milliseconds before timing out. [default7]:[rank15]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600069 milliseconds before timing out. [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f06bef7d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f06c0256c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f506a295897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f06c025ba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f506b56ec62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f06c025cdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f506b573a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #4: + 0xd3e95 (0x7f070bcf5e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #5: + 0x8609 (0x7f0710d3c609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f506b574dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #4: + 0xd3e95 (0x7f50b700de95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #6: clone + 0x43 (0x7f0710b07353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default5]:frame #5: + 0x8609 (0x7f50bc054609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #6: clone + 0x43 (0x7f50bbe1f353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]:terminate called after throwing an instance of 'c10::DistBackendError' [default5]: [default7]: what(): [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600069 milliseconds before timing out. [default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f06bef7d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:terminate called after throwing an instance of 'c10::DistBackendError' [default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f06c0256c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f06c025ba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f06c025cdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #4: + 0xd3e95 (0x7f070bcf5e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]: what(): [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600092 milliseconds before timing out. [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f506a295897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f506b56ec62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f506b573a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f506b574dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #5: + 0x8609 (0x7f0710d3c609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #4: + 0xd3e95 (0x7f50b700de95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #6: clone + 0x43 (0x7f0710b07353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]:frame #5: + 0x8609 (0x7f50bc054609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]: [default5]:frame #6: clone + 0x43 (0x7f50bbe1f353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default5]: [default7]:[rank47]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600005 milliseconds before timing out. [default0]:[rank40]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600020 milliseconds before timing out. [default4]:[rank44]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600006 milliseconds before timing out. [default4]:[rank60]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600056 milliseconds before timing out. [default7]:[rank63]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600064 milliseconds before timing out. [default5]:[rank53]: Traceback (most recent call last): [default5]:[rank53]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default5]:[rank53]: trainer.train(dataloader) [default5]:[rank53]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default5]:[rank53]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default5]:[rank53]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default5]:[rank53]: outputs = self.pipeline_engine.train_batch_iter( [default5]:[rank53]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter [default5]:[rank53]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default5]:[rank53]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default5]:[rank53]: output = model(**micro_batch) [default5]:[rank53]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default5]:[rank53]: return self._call_impl(*args, **kwargs) [default5]:[rank53]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default5]:[rank53]: return forward_call(*args, **kwargs) [default5]:[rank53]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default5]:[rank53]: sharded_logits = self.model( [default5]:[rank53]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default5]:[rank53]: return self._call_impl(*args, **kwargs) [default5]:[rank53]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default5]:[rank53]: return forward_call(*args, **kwargs) [default5]:[rank53]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default5]:[rank53]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default5]:[rank53]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default5]:[rank53]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default5]:[rank53]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default5]:[rank53]: return self._call_impl(*args, **kwargs) [default5]:[rank53]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default5]:[rank53]: return forward_call(*args, **kwargs) [default5]:[rank53]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default5]:[rank53]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default5]:[rank53]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default5]:[rank53]: pipeline_state.run_communication() [default5]:[rank53]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default5]:[rank53]: recv_activation_tensor = recv_activation() [default5]:[rank53]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default5]:[rank53]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default5]:[rank53]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default5]:[rank53]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default5]:[rank53]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default5]:[rank53]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default5]:[rank53]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default5]:[rank53]: dist.recv( [default5]:[rank53]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default5]:[rank53]: return func(*args, **kwargs) [default5]:[rank53]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default5]:[rank53]: pg.recv([tensor], group_src_rank, tag).wait() [default5]:[rank53]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 1. [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f06bef7d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default7]:frame #1: + 0xe32119 (0x7f06bfee0119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: + 0xd3e95 (0x7f070bcf5e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f506a295897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #3: + 0x8609 (0x7f0710d3c609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #4: clone + 0x43 (0x7f0710b07353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default5]:frame #1: + 0xe32119 (0x7f506b1f8119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: + 0xd3e95 (0x7f50b700de95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #3: + 0x8609 (0x7f50bc054609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #4: clone + 0x43 (0x7f50bbe1f353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default0]:[rank48]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 14, last enqueued NCCL work: 15, last completed NCCL work: 13. [default0]:[rank48]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default0]:[rank48]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default0]:[rank48]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600081 milliseconds before timing out. [default0]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f30b7c11897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f30b8eeac62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f30b8eefa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f30b8ef0dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #4: + 0xd3e95 (0x7f3104989e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #5: + 0x8609 (0x7f31099d0609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #6: clone + 0x43 (0x7f310979b353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default0]:terminate called after throwing an instance of 'c10::DistBackendError' [default0]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600081 milliseconds before timing out. [default0]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f30b7c11897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f30b8eeac62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f30b8eefa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f30b8ef0dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #4: + 0xd3e95 (0x7f3104989e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #5: + 0x8609 (0x7f31099d0609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #6: clone + 0x43 (0x7f310979b353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default0]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f30b7c11897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: + 0xe32119 (0x7f30b8b74119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: + 0xd3e95 (0x7f3104989e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #3: + 0x8609 (0x7f31099d0609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #4: clone + 0x43 (0x7f310979b353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default0]:[rank56]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600085 milliseconds before timing out. [default1]:[rank25]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 34, last enqueued NCCL work: 35, last completed NCCL work: 33. [default1]:[rank25]:[E ProcessGroupNCCL.cpp:577] [Rank 1] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default1]:[rank25]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default1]:[rank25]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600028 milliseconds before timing out. [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f9dd6fbf897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f9dd8298c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f9dd829da80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f9dd829edcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #4: + 0xd3e95 (0x7f9e23d37e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #5: + 0x8609 (0x7f9e28d7e609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #6: clone + 0x43 (0x7f9e28b49353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:terminate called after throwing an instance of 'c10::DistBackendError' [default1]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600028 milliseconds before timing out. [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f9dd6fbf897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f9dd8298c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f9dd829da80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f9dd829edcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #4: + 0xd3e95 (0x7f9e23d37e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #5: + 0x8609 (0x7f9e28d7e609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #6: clone + 0x43 (0x7f9e28b49353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f9dd6fbf897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: + 0xe32119 (0x7f9dd7f22119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: + 0xd3e95 (0x7f9e23d37e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #3: + 0x8609 (0x7f9e28d7e609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #4: clone + 0x43 (0x7f9e28b49353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default2]:[rank34]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 26, last enqueued NCCL work: 27, last completed NCCL work: 25. [default2]:[rank34]:[E ProcessGroupNCCL.cpp:577] [Rank 2] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default2]:[rank34]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default2]:[rank34]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600071 milliseconds before timing out. [default2]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f3290ae0897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f3291db9c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f3291dbea80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f3291dbfdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7f32dd858e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #5: + 0x8609 (0x7f32e289f609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7f32e266a353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default2]:terminate called after throwing an instance of 'c10::DistBackendError' [default2]: what(): [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600071 milliseconds before timing out. [default2]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f3290ae0897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f3291db9c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f3291dbea80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f3291dbfdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7f32dd858e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #5: + 0x8609 (0x7f32e289f609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7f32e266a353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default2]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f3290ae0897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: + 0xe32119 (0x7f3291a43119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: + 0xd3e95 (0x7f32dd858e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #3: + 0x8609 (0x7f32e289f609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #4: clone + 0x43 (0x7f32e266a353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default3]:[rank51]: Traceback (most recent call last): [default3]:[rank51]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default3]:[rank51]: trainer.train(dataloader) [default3]:[rank51]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default3]:[rank51]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default3]:[rank51]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default3]:[rank51]: outputs = self.pipeline_engine.train_batch_iter( [default3]:[rank51]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter [default3]:[rank51]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default3]:[rank51]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default3]:[rank51]: output = model(**micro_batch) [default3]:[rank51]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default3]:[rank51]: return self._call_impl(*args, **kwargs) [default3]:[rank51]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default3]:[rank51]: return forward_call(*args, **kwargs) [default3]:[rank51]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default3]:[rank51]: sharded_logits = self.model( [default3]:[rank51]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default3]:[rank51]: return self._call_impl(*args, **kwargs) [default3]:[rank51]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default3]:[rank51]: return forward_call(*args, **kwargs) [default3]:[rank51]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default3]:[rank51]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default3]:[rank51]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default3]:[rank51]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default3]:[rank51]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default3]:[rank51]: return self._call_impl(*args, **kwargs) [default3]:[rank51]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default3]:[rank51]: return forward_call(*args, **kwargs) [default3]:[rank51]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default3]:[rank51]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default3]:[rank51]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default3]:[rank51]: pipeline_state.run_communication() [default3]:[rank51]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default3]:[rank51]: recv_activation_tensor = recv_activation() [default3]:[rank51]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default3]:[rank51]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default3]:[rank51]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default3]:[rank51]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default3]:[rank51]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default3]:[rank51]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default3]:[rank51]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default3]:[rank51]: dist.recv( [default3]:[rank51]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default3]:[rank51]: return func(*args, **kwargs) [default3]:[rank51]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default3]:[rank51]: pg.recv([tensor], group_src_rank, tag).wait() [default3]:[rank51]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 1. [default6]:[rank54]: Traceback (most recent call last): [default6]:[rank54]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default6]:[rank54]: trainer.train(dataloader) [default6]:[rank54]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default6]:[rank54]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default6]:[rank54]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default6]:[rank54]: outputs = self.pipeline_engine.train_batch_iter( [default6]:[rank54]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter [default6]:[rank54]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default6]:[rank54]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default6]:[rank54]: output = model(**micro_batch) [default6]:[rank54]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default6]:[rank54]: return self._call_impl(*args, **kwargs) [default6]:[rank54]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default6]:[rank54]: return forward_call(*args, **kwargs) [default6]:[rank54]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default6]:[rank54]: sharded_logits = self.model( [default6]:[rank54]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default6]:[rank54]: return self._call_impl(*args, **kwargs) [default6]:[rank54]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default6]:[rank54]: return forward_call(*args, **kwargs) [default6]:[rank54]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default6]:[rank54]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default6]:[rank54]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default6]:[rank54]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default6]:[rank54]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default6]:[rank54]: return self._call_impl(*args, **kwargs) [default6]:[rank54]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default6]:[rank54]: return forward_call(*args, **kwargs) [default6]:[rank54]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default6]:[rank54]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default6]:[rank54]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default6]:[rank54]: pipeline_state.run_communication() [default6]:[rank54]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default6]:[rank54]: recv_activation_tensor = recv_activation() [default6]:[rank54]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default6]:[rank54]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default6]:[rank54]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default6]:[rank54]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default6]:[rank54]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default6]:[rank54]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default6]:[rank54]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default6]:[rank54]: dist.recv( [default6]:[rank54]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default6]:[rank54]: return func(*args, **kwargs) [default6]:[rank54]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default6]:[rank54]: pg.recv([tensor], group_src_rank, tag).wait() [default6]:[rank54]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 1. [default3]:[rank43]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600081 milliseconds before timing out. [default3]:[rank59]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600073 milliseconds before timing out. [default6]:[rank62]: Traceback (most recent call last): [default6]:[rank62]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default6]:[rank62]: trainer.train(dataloader) [default6]:[rank62]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default6]:[rank62]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default6]:[rank62]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default6]:[rank62]: outputs = self.pipeline_engine.train_batch_iter( [default6]:[rank62]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter [default6]:[rank62]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default6]:[rank62]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default6]:[rank62]: output = model(**micro_batch) [default6]:[rank62]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default6]:[rank62]: return self._call_impl(*args, **kwargs) [default6]:[rank62]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default6]:[rank62]: return forward_call(*args, **kwargs) [default6]:[rank62]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default6]:[rank62]: sharded_logits = self.model( [default6]:[rank62]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default6]:[rank62]: return self._call_impl(*args, **kwargs) [default6]:[rank62]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default6]:[rank62]: return forward_call(*args, **kwargs) [default6]:[rank62]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default6]:[rank62]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default6]:[rank62]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default6]:[rank62]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default6]:[rank62]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default6]:[rank62]: return self._call_impl(*args, **kwargs) [default6]:[rank62]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default6]:[rank62]: return forward_call(*args, **kwargs) [default6]:[rank62]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default6]:[rank62]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default6]:[rank62]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default6]:[rank62]: pipeline_state.run_communication() [default6]:[rank62]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default6]:[rank62]: recv_activation_tensor = recv_activation() [default6]:[rank62]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default6]:[rank62]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default6]:[rank62]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default6]:[rank62]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default6]:[rank62]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default6]:[rank62]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default6]:[rank62]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default6]:[rank62]: dist.recv( [default6]:[rank62]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default6]:[rank62]: return func(*args, **kwargs) [default6]:[rank62]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default6]:[rank62]: pg.recv([tensor], group_src_rank, tag).wait() [default6]:[rank62]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 1. [default4]:[rank44]: Traceback (most recent call last): [default4]:[rank44]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default4]:[rank44]: trainer.train(dataloader) [default4]:[rank44]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default4]:[rank44]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default4]:[rank44]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default4]:[rank44]: outputs = self.pipeline_engine.train_batch_iter( [default4]:[rank44]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default4]:[rank44]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default4]:[rank44]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default4]:[rank44]: grad_accumulator.backward(sum(activations)) [default4]:[rank44]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default4]:[rank44]: result = loss.backward() [default4]:[rank44]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default4]:[rank44]: torch.autograd.backward( [default4]:[rank44]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default4]:[rank44]: _engine_run_backward( [default4]:[rank44]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default4]:[rank44]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default4]:[rank44]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default4]:[rank44]: return user_fn(self, *args) [default4]:[rank44]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default4]:[rank44]: pipeline_state.run_communication() [default4]:[rank44]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 172, in run_communication [default4]:[rank44]: self.grads_buffer.append(recv_grad()) [default4]:[rank44]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 50, in __call__ [default4]:[rank44]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default4]:[rank44]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default4]:[rank44]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default4]:[rank44]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default4]:[rank44]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default4]:[rank44]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default4]:[rank44]: dist.recv( [default4]:[rank44]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default4]:[rank44]: return func(*args, **kwargs) [default4]:[rank44]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default4]:[rank44]: pg.recv([tensor], group_src_rank, tag).wait() [default4]:[rank44]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default2]:[rank58]: Traceback (most recent call last): [default2]:[rank58]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default2]:[rank58]: trainer.train(dataloader) [default2]:[rank58]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default2]:[rank58]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default2]:[rank58]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default2]:[rank58]: outputs = self.pipeline_engine.train_batch_iter( [default2]:[rank58]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter [default2]:[rank58]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default2]:[rank58]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default2]:[rank58]: output = model(**micro_batch) [default2]:[rank58]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default2]:[rank58]: return self._call_impl(*args, **kwargs) [default2]:[rank58]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default2]:[rank58]: return forward_call(*args, **kwargs) [default2]:[rank58]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default2]:[rank58]: sharded_logits = self.model( [default2]:[rank58]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default2]:[rank58]: return self._call_impl(*args, **kwargs) [default2]:[rank58]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default2]:[rank58]: return forward_call(*args, **kwargs) [default2]:[rank58]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default2]:[rank58]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default2]:[rank58]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default2]:[rank58]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default2]:[rank58]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default2]:[rank58]: return self._call_impl(*args, **kwargs) [default2]:[rank58]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default2]:[rank58]: return forward_call(*args, **kwargs) [default2]:[rank58]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default2]:[rank58]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default2]:[rank58]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default2]:[rank58]: pipeline_state.run_communication() [default2]:[rank58]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default2]:[rank58]: recv_activation_tensor = recv_activation() [default2]:[rank58]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default2]:[rank58]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default2]:[rank58]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default2]:[rank58]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default2]:[rank58]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default2]:[rank58]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default2]:[rank58]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default2]:[rank58]: dist.recv( [default2]:[rank58]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default2]:[rank58]: return func(*args, **kwargs) [default2]:[rank58]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default2]:[rank58]: pg.recv([tensor], group_src_rank, tag).wait() [default2]:[rank58]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 1. [default2]:[rank42]: Traceback (most recent call last): [default2]:[rank42]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default2]:[rank42]: trainer.train(dataloader) [default2]:[rank42]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default2]:[rank42]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default2]:[rank42]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default2]:[rank42]: outputs = self.pipeline_engine.train_batch_iter( [default2]:[rank42]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default2]:[rank42]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default2]:[rank42]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default2]:[rank42]: grad_accumulator.backward(sum(activations)) [default2]:[rank42]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default2]:[rank42]: result = loss.backward() [default2]:[rank42]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default2]:[rank42]: torch.autograd.backward( [default2]:[rank42]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default2]:[rank42]: _engine_run_backward( [default2]:[rank42]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default2]:[rank42]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default2]:[rank42]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default2]:[rank42]: return user_fn(self, *args) [default2]:[rank42]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default2]:[rank42]: pipeline_state.run_communication() [default2]:[rank42]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 172, in run_communication [default2]:[rank42]: self.grads_buffer.append(recv_grad()) [default2]:[rank42]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 50, in __call__ [default2]:[rank42]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default2]:[rank42]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default2]:[rank42]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default2]:[rank42]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default2]:[rank42]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default2]:[rank42]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default2]:[rank42]: dist.recv( [default2]:[rank42]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default2]:[rank42]: return func(*args, **kwargs) [default2]:[rank42]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default2]:[rank42]: pg.recv([tensor], group_src_rank, tag).wait() [default2]:[rank42]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default6]:[rank38]: Traceback (most recent call last): [default6]:[rank38]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default6]:[rank38]: trainer.train(dataloader) [default6]:[rank38]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default6]:[rank38]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default6]:[rank38]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default6]:[rank38]: outputs = self.pipeline_engine.train_batch_iter( [default6]:[rank38]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default6]:[rank38]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default6]:[rank38]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default6]:[rank38]: grad_accumulator.backward(sum(activations)) [default6]:[rank38]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default6]:[rank38]: result = loss.backward() [default6]:[rank38]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default6]:[rank38]: torch.autograd.backward( [default6]:[rank38]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default6]:[rank38]: _engine_run_backward( [default6]:[rank38]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default6]:[rank38]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default6]:[rank38]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default6]:[rank38]: return user_fn(self, *args) [default6]:[rank38]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default6]:[rank38]: pipeline_state.run_communication() [default6]:[rank38]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 172, in run_communication [default6]:[rank38]: self.grads_buffer.append(recv_grad()) [default6]:[rank38]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 50, in __call__ [default6]:[rank38]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default6]:[rank38]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default6]:[rank38]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default6]:[rank38]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default6]:[rank38]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default6]:[rank38]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default6]:[rank38]: dist.recv( [default6]:[rank38]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default6]:[rank38]: return func(*args, **kwargs) [default6]:[rank38]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default6]:[rank38]: pg.recv([tensor], group_src_rank, tag).wait() [default6]:[rank38]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default1]:[rank41]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 26, last enqueued NCCL work: 27, last completed NCCL work: 25. [default1]:[rank41]:[E ProcessGroupNCCL.cpp:577] [Rank 2] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default1]:[rank41]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default1]:[rank41]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600043 milliseconds before timing out. [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fb7bdae4897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fb7bedbdc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fb7bedc2a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fb7bedc3dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #4: + 0xd3e95 (0x7fb80a85ce95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #5: + 0x8609 (0x7fb80f8a3609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #6: clone + 0x43 (0x7fb80f66e353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:terminate called after throwing an instance of 'c10::DistBackendError' [default1]: what(): [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600043 milliseconds before timing out. [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fb7bdae4897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fb7bedbdc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fb7bedc2a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fb7bedc3dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #4: + 0xd3e95 (0x7fb80a85ce95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #5: + 0x8609 (0x7fb80f8a3609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #6: clone + 0x43 (0x7fb80f66e353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default1]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fb7bdae4897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default1]:frame #1: + 0xe32119 (0x7fb7bea47119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default1]:frame #2: + 0xd3e95 (0x7fb80a85ce95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default1]:frame #3: + 0x8609 (0x7fb80f8a3609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default1]:frame #4: clone + 0x43 (0x7fb80f66e353 in /lib/x86_64-linux-gnu/libc.so.6) [default1]: [default4]:[rank52]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 14, last enqueued NCCL work: 15, last completed NCCL work: 13. [default4]:[rank52]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default4]:[rank52]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default4]:[rank52]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600072 milliseconds before timing out. [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f493e719897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f493f9f2c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f493f9f7a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f493f9f8dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f498b491e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f49904d8609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f49902a3353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default4]:terminate called after throwing an instance of 'c10::DistBackendError' [default4]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600072 milliseconds before timing out. [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f493e719897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f493f9f2c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f493f9f7a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f493f9f8dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f498b491e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f49904d8609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f49902a3353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default4]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f493e719897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: + 0xe32119 (0x7f493f67c119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: + 0xd3e95 (0x7f498b491e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #3: + 0x8609 (0x7f49904d8609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #4: clone + 0x43 (0x7f49902a3353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default2]:[rank50]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 14, last enqueued NCCL work: 15, last completed NCCL work: 13. [default2]:[rank50]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default2]:[rank50]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default2]:[rank50]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600052 milliseconds before timing out. [default2]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fcfb3cc0897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fcfb4f99c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fcfb4f9ea80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fcfb4f9fdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7fd000a38e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #5: + 0x8609 (0x7fd005a7f609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7fd00584a353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default2]:terminate called after throwing an instance of 'c10::DistBackendError' [default2]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600052 milliseconds before timing out. [default2]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fcfb3cc0897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fcfb4f99c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fcfb4f9ea80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fcfb4f9fdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7fd000a38e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #5: + 0x8609 (0x7fd005a7f609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7fd00584a353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default2]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fcfb3cc0897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: + 0xe32119 (0x7fcfb4c23119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: + 0xd3e95 (0x7fd000a38e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #3: + 0x8609 (0x7fd005a7f609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #4: clone + 0x43 (0x7fd00584a353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default3]:[rank35]: Traceback (most recent call last): [default3]:[rank35]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default3]:[rank35]: trainer.train(dataloader) [default3]:[rank35]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default3]:[rank35]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default3]:[rank35]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default3]:[rank35]: outputs = self.pipeline_engine.train_batch_iter( [default3]:[rank35]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default3]:[rank35]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default3]:[rank35]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default3]:[rank35]: grad_accumulator.backward(sum(activations)) [default3]:[rank35]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default3]:[rank35]: result = loss.backward() [default3]:[rank35]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default3]:[rank35]: torch.autograd.backward( [default3]:[rank35]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default3]:[rank35]: _engine_run_backward( [default3]:[rank35]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default3]:[rank35]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default3]:[rank35]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default3]:[rank35]: return user_fn(self, *args) [default3]:[rank35]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default3]:[rank35]: pipeline_state.run_communication() [default3]:[rank35]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 172, in run_communication [default3]:[rank35]: self.grads_buffer.append(recv_grad()) [default3]:[rank35]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 50, in __call__ [default3]:[rank35]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default3]:[rank35]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default3]:[rank35]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default3]:[rank35]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default3]:[rank35]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default3]:[rank35]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default3]:[rank35]: dist.recv( [default3]:[rank35]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default3]:[rank35]: return func(*args, **kwargs) [default3]:[rank35]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default3]:[rank35]: pg.recv([tensor], group_src_rank, tag).wait() [default3]:[rank35]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default0]:[rank40]: Traceback (most recent call last): [default0]:[rank40]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default0]:[rank40]: trainer.train(dataloader) [default0]:[rank40]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default0]:[rank40]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default0]:[rank40]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default0]:[rank40]: outputs = self.pipeline_engine.train_batch_iter( [default0]:[rank40]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default0]:[rank40]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default0]:[rank40]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default0]:[rank40]: grad_accumulator.backward(sum(activations)) [default0]:[rank40]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default0]:[rank40]: result = loss.backward() [default0]:[rank40]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default0]:[rank40]: torch.autograd.backward( [default0]:[rank40]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default0]:[rank40]: _engine_run_backward( [default0]:[rank40]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default0]:[rank40]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default0]:[rank40]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default0]:[rank40]: return user_fn(self, *args) [default0]:[rank40]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default0]:[rank40]: pipeline_state.run_communication() [default0]:[rank40]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 172, in run_communication [default0]:[rank40]: self.grads_buffer.append(recv_grad()) [default0]:[rank40]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 50, in __call__ [default0]:[rank40]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default0]:[rank40]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default0]:[rank40]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default0]:[rank40]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default0]:[rank40]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default0]:[rank40]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default0]:[rank40]: dist.recv( [default0]:[rank40]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default0]:[rank40]: return func(*args, **kwargs) [default0]:[rank40]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default0]:[rank40]: pg.recv([tensor], group_src_rank, tag).wait() [default0]:[rank40]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default0]:[rank8]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 0] Timeout at NCCL work: 22, last enqueued NCCL work: 23, last completed NCCL work: 21. [default0]:[rank8]:[E ProcessGroupNCCL.cpp:577] [Rank 0] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default0]:[rank8]:[E ProcessGroupNCCL.cpp:583] [Rank 0] To avoid data inconsistency, we are taking the entire process down. [default0]:[rank8]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600045 milliseconds before timing out. [default0]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f34d4726897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f34d59ffc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f34d5a04a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f34d5a05dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #4: + 0xd3e95 (0x7f352149ee95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #5: + 0x8609 (0x7f35264e5609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #6: clone + 0x43 (0x7f35262b0353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default0]:terminate called after throwing an instance of 'c10::DistBackendError' [default0]: what(): [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600045 milliseconds before timing out. [default0]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f34d4726897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f34d59ffc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f34d5a04a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f34d5a05dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #4: + 0xd3e95 (0x7f352149ee95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #5: + 0x8609 (0x7f35264e5609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #6: clone + 0x43 (0x7f35262b0353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default0]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f34d4726897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: + 0xe32119 (0x7f34d5689119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: + 0xd3e95 (0x7f352149ee95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #3: + 0x8609 (0x7f35264e5609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #4: clone + 0x43 (0x7f35262b0353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default6]:[rank46]: Traceback (most recent call last): [default6]:[rank46]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default6]:[rank46]: trainer.train(dataloader) [default6]:[rank46]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default6]:[rank46]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default6]:[rank46]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default6]:[rank46]: outputs = self.pipeline_engine.train_batch_iter( [default6]:[rank46]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default6]:[rank46]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default6]:[rank46]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default6]:[rank46]: grad_accumulator.backward(sum(activations)) [default6]:[rank46]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default6]:[rank46]: result = loss.backward() [default6]:[rank46]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default6]:[rank46]: torch.autograd.backward( [default6]:[rank46]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default6]:[rank46]: _engine_run_backward( [default6]:[rank46]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default6]:[rank46]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default6]:[rank46]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default6]:[rank46]: return user_fn(self, *args) [default6]:[rank46]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default6]:[rank46]: pipeline_state.run_communication() [default6]:[rank46]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 172, in run_communication [default6]:[rank46]: self.grads_buffer.append(recv_grad()) [default6]:[rank46]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 50, in __call__ [default6]:[rank46]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default6]:[rank46]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default6]:[rank46]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default6]:[rank46]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default6]:[rank46]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default6]:[rank46]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default6]:[rank46]: dist.recv( [default6]:[rank46]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default6]:[rank46]: return func(*args, **kwargs) [default6]:[rank46]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default6]:[rank46]: pg.recv([tensor], group_src_rank, tag).wait() [default6]:[rank46]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default5]:[rank61]: Traceback (most recent call last): [default5]:[rank61]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default5]:[rank61]: trainer.train(dataloader) [default5]:[rank61]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default5]:[rank61]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default5]:[rank61]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default5]:[rank61]: outputs = self.pipeline_engine.train_batch_iter( [default5]:[rank61]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter [default5]:[rank61]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default5]:[rank61]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default5]:[rank61]: output = model(**micro_batch) [default5]:[rank61]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default5]:[rank61]: return self._call_impl(*args, **kwargs) [default5]:[rank61]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default5]:[rank61]: return forward_call(*args, **kwargs) [default5]:[rank61]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default5]:[rank61]: sharded_logits = self.model( [default5]:[rank61]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default5]:[rank61]: return self._call_impl(*args, **kwargs) [default5]:[rank61]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default5]:[rank61]: return forward_call(*args, **kwargs) [default5]:[rank61]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default5]:[rank61]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default5]:[rank61]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default5]:[rank61]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default5]:[rank61]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default5]:[rank61]: return self._call_impl(*args, **kwargs) [default5]:[rank61]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default5]:[rank61]: return forward_call(*args, **kwargs) [default5]:[rank61]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default5]:[rank61]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default5]:[rank61]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default5]:[rank61]: pipeline_state.run_communication() [default5]:[rank61]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default5]:[rank61]: recv_activation_tensor = recv_activation() [default5]:[rank61]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default5]:[rank61]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default5]:[rank61]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default5]:[rank61]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default5]:[rank61]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default5]:[rank61]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default5]:[rank61]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default5]:[rank61]: dist.recv( [default5]:[rank61]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default5]:[rank61]: return func(*args, **kwargs) [default5]:[rank61]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default5]:[rank61]: pg.recv([tensor], group_src_rank, tag).wait() [default5]:[rank61]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 1. [default0]:[rank56]: Traceback (most recent call last): [default0]:[rank56]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default0]:[rank56]: trainer.train(dataloader) [default0]:[rank56]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default0]:[rank56]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default0]:[rank56]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default0]:[rank56]: outputs = self.pipeline_engine.train_batch_iter( [default0]:[rank56]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter [default0]:[rank56]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default0]:[rank56]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default0]:[rank56]: output = model(**micro_batch) [default0]:[rank56]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default0]:[rank56]: return self._call_impl(*args, **kwargs) [default0]:[rank56]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default0]:[rank56]: return forward_call(*args, **kwargs) [default0]:[rank56]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default0]:[rank56]: sharded_logits = self.model( [default0]:[rank56]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default0]:[rank56]: return self._call_impl(*args, **kwargs) [default0]:[rank56]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default0]:[rank56]: return forward_call(*args, **kwargs) [default0]:[rank56]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default0]:[rank56]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default0]:[rank56]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default0]:[rank56]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default0]:[rank56]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default0]:[rank56]: return self._call_impl(*args, **kwargs) [default0]:[rank56]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default0]:[rank56]: return forward_call(*args, **kwargs) [default0]:[rank56]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default0]:[rank56]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default0]:[rank56]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default0]:[rank56]: pipeline_state.run_communication() [default0]:[rank56]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default0]:[rank56]: recv_activation_tensor = recv_activation() [default0]:[rank56]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default0]:[rank56]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default0]:[rank56]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default0]:[rank56]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default0]:[rank56]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default0]:[rank56]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default0]:[rank56]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default0]:[rank56]: dist.recv( [default0]:[rank56]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default0]:[rank56]: return func(*args, **kwargs) [default0]:[rank56]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default0]:[rank56]: pg.recv([tensor], group_src_rank, tag).wait() [default0]:[rank56]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 1. [default3]:[rank11]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 0] Timeout at NCCL work: 22, last enqueued NCCL work: 23, last completed NCCL work: 21. [default3]:[rank11]:[E ProcessGroupNCCL.cpp:577] [Rank 0] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default3]:[rank11]:[E ProcessGroupNCCL.cpp:583] [Rank 0] To avoid data inconsistency, we are taking the entire process down. [default3]:[rank11]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600040 milliseconds before timing out. [default3]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7efc6211a897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7efc633f3c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7efc633f8a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7efc633f9dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #4: + 0xd3e95 (0x7efcaee92e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #5: + 0x8609 (0x7efcb3ed9609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #6: clone + 0x43 (0x7efcb3ca4353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default3]:terminate called after throwing an instance of 'c10::DistBackendError' [default3]: what(): [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600040 milliseconds before timing out. [default3]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7efc6211a897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7efc633f3c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7efc633f8a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7efc633f9dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #4: + 0xd3e95 (0x7efcaee92e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #5: + 0x8609 (0x7efcb3ed9609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #6: clone + 0x43 (0x7efcb3ca4353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default3]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7efc6211a897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: + 0xe32119 (0x7efc6307d119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: + 0xd3e95 (0x7efcaee92e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #3: + 0x8609 (0x7efcb3ed9609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #4: clone + 0x43 (0x7efcb3ca4353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default4]:[rank12]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 0] Timeout at NCCL work: 22, last enqueued NCCL work: 23, last completed NCCL work: 21. [default4]:[rank12]:[E ProcessGroupNCCL.cpp:577] [Rank 0] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default4]:[rank12]:[E ProcessGroupNCCL.cpp:583] [Rank 0] To avoid data inconsistency, we are taking the entire process down. [default4]:[rank12]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600081 milliseconds before timing out. [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f928b784897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:[rank10]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 0] Timeout at NCCL work: 22, last enqueued NCCL work: 23, last completed NCCL work: 21. [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f928ca5dc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f928ca62a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f928ca63dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f92d84fce95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f92dd543609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:[rank10]:[E ProcessGroupNCCL.cpp:577] [Rank 0] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default2]:[rank10]:[E ProcessGroupNCCL.cpp:583] [Rank 0] To avoid data inconsistency, we are taking the entire process down. [default4]:frame #6: clone + 0x43 (0x7f92dd30e353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]:[rank10]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600001 milliseconds before timing out. [default2]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7ff3f0c31897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]: [default4]:terminate called after throwing an instance of 'c10::DistBackendError' [default2]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7ff3f1f0ac62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]: what(): [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600081 milliseconds before timing out. [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7ff3f1f0fa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7ff3f1f10dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7ff43d9a9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f928b784897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #5: + 0x8609 (0x7ff4429f0609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7ff4427bb353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f928ca5dc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f928ca62a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:terminate called after throwing an instance of 'c10::DistBackendError' [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f928ca63dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]: what(): [PG 4 Rank 0] Process group watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=22, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600001 milliseconds before timing out. [default2]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7ff3f0c31897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7ff3f1f0ac62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f92d84fce95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f92dd543609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f92dd30e353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default4]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default2]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7ff3f1f0fa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7ff3f1f10dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f928b784897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: + 0xe32119 (0x7f928c6e7119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: + 0xd3e95 (0x7f92d84fce95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #3: + 0x8609 (0x7f92dd543609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #4: clone + 0x43 (0x7f92dd30e353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default2]:frame #4: + 0xd3e95 (0x7ff43d9a9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #5: + 0x8609 (0x7ff4429f0609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7ff4427bb353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default2]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7ff3f0c31897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: + 0xe32119 (0x7ff3f1b94119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: + 0xd3e95 (0x7ff43d9a9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #3: + 0x8609 (0x7ff4429f0609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #4: clone + 0x43 (0x7ff4427bb353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default2]:[rank58]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 14, last enqueued NCCL work: 15, last completed NCCL work: 13. [default2]:[rank58]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default2]:[rank58]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default2]:[rank58]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600019 milliseconds before timing out. [default2]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f3eae292897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f3eaf56bc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f3eaf570a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f3eaf571dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7f3efb00ae95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #5: + 0x8609 (0x7f3f00051609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7f3effe1c353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default2]:terminate called after throwing an instance of 'c10::DistBackendError' [default2]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600019 milliseconds before timing out. [default2]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f3eae292897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f3eaf56bc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f3eaf570a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f3eaf571dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7f3efb00ae95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #5: + 0x8609 (0x7f3f00051609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7f3effe1c353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default2]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f3eae292897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: + 0xe32119 (0x7f3eaf1f5119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: + 0xd3e95 (0x7f3efb00ae95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #3: + 0x8609 (0x7f3f00051609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #4: clone + 0x43 (0x7f3effe1c353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default6]:[rank62]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 14, last enqueued NCCL work: 15, last completed NCCL work: 13. [default6]:[rank62]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default6]:[rank62]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default6]:[rank62]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600006 milliseconds before timing out. [default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f9a448fc897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f9a45bd5c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f9a45bdaa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f9a45bdbdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7f9a91674e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7f9a966bb609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7f9a96486353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:terminate called after throwing an instance of 'c10::DistBackendError' [default6]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600006 milliseconds before timing out. [default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f9a448fc897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f9a45bd5c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f9a45bdaa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f9a45bdbdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7f9a91674e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7f9a966bb609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7f9a96486353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f9a448fc897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: + 0xe32119 (0x7f9a4585f119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: + 0xd3e95 (0x7f9a91674e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #3: + 0x8609 (0x7f9a966bb609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #4: clone + 0x43 (0x7f9a96486353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default7]:[rank47]: Traceback (most recent call last): [default7]:[rank47]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default7]:[rank47]: trainer.train(dataloader) [default7]:[rank47]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default7]:[rank47]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default7]:[rank47]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default7]:[rank47]: outputs = self.pipeline_engine.train_batch_iter( [default7]:[rank47]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default7]:[rank47]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default7]:[rank47]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default7]:[rank47]: grad_accumulator.backward(sum(activations)) [default7]:[rank47]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default7]:[rank47]: result = loss.backward() [default7]:[rank47]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default7]:[rank47]: torch.autograd.backward( [default7]:[rank47]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default7]:[rank47]: _engine_run_backward( [default7]:[rank47]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default7]:[rank47]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default7]:[rank47]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default7]:[rank47]: return user_fn(self, *args) [default7]:[rank47]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default7]:[rank47]: pipeline_state.run_communication() [default7]:[rank47]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 172, in run_communication [default7]:[rank47]: self.grads_buffer.append(recv_grad()) [default7]:[rank47]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 50, in __call__ [default7]:[rank47]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default7]:[rank47]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default7]:[rank47]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default7]:[rank47]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default7]:[rank47]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default7]:[rank47]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default7]:[rank47]: dist.recv( [default7]:[rank47]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default7]:[rank47]: return func(*args, **kwargs) [default7]:[rank47]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default7]:[rank47]: pg.recv([tensor], group_src_rank, tag).wait() [default7]:[rank47]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default3]:[rank43]: Traceback (most recent call last): [default3]:[rank43]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default3]:[rank43]: trainer.train(dataloader) [default3]:[rank43]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default3]:[rank43]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default3]:[rank43]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default3]:[rank43]: outputs = self.pipeline_engine.train_batch_iter( [default3]:[rank43]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 295, in train_batch_iter [default3]:[rank43]: self.backward(context=context, state=state, grad_accumulator=grad_accumulator) [default3]:[rank43]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 86, in backward [default3]:[rank43]: grad_accumulator.backward(sum(activations)) [default3]:[rank43]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/optim/gradient_accumulator.py", line 205, in backward [default3]:[rank43]: result = loss.backward() [default3]:[rank43]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/_tensor.py", line 525, in backward [default3]:[rank43]: torch.autograd.backward( [default3]:[rank43]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/__init__.py", line 267, in backward [default3]:[rank43]: _engine_run_backward( [default3]:[rank43]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward [default3]:[rank43]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass [default3]:[rank43]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 301, in apply [default3]:[rank43]: return user_fn(self, *args) [default3]:[rank43]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 40, in backward [default3]:[rank43]: pipeline_state.run_communication() [default3]:[rank43]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 172, in run_communication [default3]:[rank43]: self.grads_buffer.append(recv_grad()) [default3]:[rank43]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 50, in __call__ [default3]:[rank43]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default3]:[rank43]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default3]:[rank43]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default3]:[rank43]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default3]:[rank43]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default3]:[rank43]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default3]:[rank43]: dist.recv( [default3]:[rank43]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default3]:[rank43]: return func(*args, **kwargs) [default3]:[rank43]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default3]:[rank43]: pg.recv([tensor], group_src_rank, tag).wait() [default3]:[rank43]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 0. [default0]:[rank16]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 34, last enqueued NCCL work: 35, last completed NCCL work: 33. [default0]:[rank16]:[E ProcessGroupNCCL.cpp:577] [Rank 1] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default0]:[rank16]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default0]:[rank16]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600094 milliseconds before timing out. [default0]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f14f5492897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f14f676bc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f14f6770a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f14f6771dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #4: + 0xd3e95 (0x7f154220ae95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #5: + 0x8609 (0x7f1547251609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #6: clone + 0x43 (0x7f154701c353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default0]:terminate called after throwing an instance of 'c10::DistBackendError' [default0]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600094 milliseconds before timing out. [default0]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f14f5492897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f14f676bc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f14f6770a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f14f6771dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #4: + 0xd3e95 (0x7f154220ae95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #5: + 0x8609 (0x7f1547251609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #6: clone + 0x43 (0x7f154701c353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default0]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f14f5492897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: + 0xe32119 (0x7f14f63f5119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: + 0xd3e95 (0x7f154220ae95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #3: + 0x8609 (0x7f1547251609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #4: clone + 0x43 (0x7f154701c353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default5]:[rank21]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 34, last enqueued NCCL work: 35, last completed NCCL work: 33. [default5]:[rank21]:[E ProcessGroupNCCL.cpp:577] [Rank 1] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default5]:[rank21]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default5]:[rank21]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600090 milliseconds before timing out. [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fca74141897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fca7541ac62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fca7541fa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fca75420dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #4: + 0xd3e95 (0x7fcac0eb9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #5: + 0x8609 (0x7fcac5f00609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #6: clone + 0x43 (0x7fcac5ccb353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default5]:terminate called after throwing an instance of 'c10::DistBackendError' [default5]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600090 milliseconds before timing out. [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fca74141897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fca7541ac62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fca7541fa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fca75420dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #4: + 0xd3e95 (0x7fcac0eb9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #5: + 0x8609 (0x7fcac5f00609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #6: clone + 0x43 (0x7fcac5ccb353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default5]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fca74141897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: + 0xe32119 (0x7fca750a4119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: + 0xd3e95 (0x7fcac0eb9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #3: + 0x8609 (0x7fcac5f00609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #4: clone + 0x43 (0x7fcac5ccb353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default3]:[rank19]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 34, last enqueued NCCL work: 35, last completed NCCL work: 33. [default3]:[rank19]:[E ProcessGroupNCCL.cpp:577] [Rank 1] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default3]:[rank19]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default3]:[rank19]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600079 milliseconds before timing out. [default3]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f052eb46897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f052fe1fc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f052fe24a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f052fe25dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #4: + 0xd3e95 (0x7f057b8bee95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #5: + 0x8609 (0x7f0580905609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #6: clone + 0x43 (0x7f05806d0353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default3]:terminate called after throwing an instance of 'c10::DistBackendError' [default3]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600079 milliseconds before timing out. [default3]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f052eb46897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f052fe1fc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f052fe24a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f052fe25dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #4: + 0xd3e95 (0x7f057b8bee95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #5: + 0x8609 (0x7f0580905609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #6: clone + 0x43 (0x7f05806d0353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default3]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f052eb46897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: + 0xe32119 (0x7f052faa9119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: + 0xd3e95 (0x7f057b8bee95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #3: + 0x8609 (0x7f0580905609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #4: clone + 0x43 (0x7f05806d0353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default5]:[rank53]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 14, last enqueued NCCL work: 15, last completed NCCL work: 13. [default5]:[rank53]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default5]:[rank53]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default5]:[rank53]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600071 milliseconds before timing out. [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f2cd4473897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f2cd574cc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f2cd5751a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f2cd5752dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #4: + 0xd3e95 (0x7f2d211ebe95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #5: + 0x8609 (0x7f2d26232609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #6: clone + 0x43 (0x7f2d25ffd353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default5]:terminate called after throwing an instance of 'c10::DistBackendError' [default5]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600071 milliseconds before timing out. [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f2cd4473897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f2cd574cc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f2cd5751a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f2cd5752dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #4: + 0xd3e95 (0x7f2d211ebe95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #5: + 0x8609 (0x7f2d26232609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #6: clone + 0x43 (0x7f2d25ffd353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default5]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f2cd4473897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: + 0xe32119 (0x7f2cd53d6119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: + 0xd3e95 (0x7f2d211ebe95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #3: + 0x8609 (0x7f2d26232609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #4: clone + 0x43 (0x7f2d25ffd353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default4]:[rank36]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 26, last enqueued NCCL work: 27, last completed NCCL work: 25. [default4]:[rank36]:[E ProcessGroupNCCL.cpp:577] [Rank 2] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default4]:[rank36]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default4]:[rank36]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600005 milliseconds before timing out. [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f5699daf897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f569b088c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f569b08da80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f569b08edcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f56e6b27e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f56ebb6e609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f56eb939353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default4]:terminate called after throwing an instance of 'c10::DistBackendError' [default4]: what(): [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600005 milliseconds before timing out. [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f5699daf897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f569b088c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f569b08da80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f569b08edcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f56e6b27e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f56ebb6e609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f56eb939353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default4]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f5699daf897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: + 0xe32119 (0x7f569ad12119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: + 0xd3e95 (0x7f56e6b27e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #3: + 0x8609 (0x7f56ebb6e609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #4: clone + 0x43 (0x7f56eb939353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default5]:[rank45]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 26, last enqueued NCCL work: 27, last completed NCCL work: 25. [default5]:[rank45]:[E ProcessGroupNCCL.cpp:577] [Rank 2] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default5]:[rank45]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default5]:[rank45]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600074 milliseconds before timing out. [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f4e00e25897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f4e020fec62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f4e02103a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f4e02104dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #4: + 0xd3e95 (0x7f4e4db9de95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #5: + 0x8609 (0x7f4e52be4609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #6: clone + 0x43 (0x7f4e529af353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default5]:terminate called after throwing an instance of 'c10::DistBackendError' [default5]: what(): [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600074 milliseconds before timing out. [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f4e00e25897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f4e020fec62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f4e02103a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f4e02104dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #4: + 0xd3e95 (0x7f4e4db9de95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #5: + 0x8609 (0x7f4e52be4609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #6: clone + 0x43 (0x7f4e529af353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default5]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f4e00e25897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: + 0xe32119 (0x7f4e01d88119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: + 0xd3e95 (0x7f4e4db9de95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #3: + 0x8609 (0x7f4e52be4609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #4: clone + 0x43 (0x7f4e529af353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default5]:[rank37]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 26, last enqueued NCCL work: 27, last completed NCCL work: 25. [default5]:[rank37]:[E ProcessGroupNCCL.cpp:577] [Rank 2] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default5]:[rank37]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default5]:[rank37]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600006 milliseconds before timing out. [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f98222db897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f98235b4c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f98235b9a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f98235badcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #4: + 0xd3e95 (0x7f986f053e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #5: + 0x8609 (0x7f987409a609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #6: clone + 0x43 (0x7f9873e65353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default5]:terminate called after throwing an instance of 'c10::DistBackendError' [default5]: what(): [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600006 milliseconds before timing out. [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f98222db897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f98235b4c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f98235b9a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f98235badcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #4: + 0xd3e95 (0x7f986f053e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #5: + 0x8609 (0x7f987409a609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #6: clone + 0x43 (0x7f9873e65353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default5]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f98222db897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: + 0xe32119 (0x7f982323e119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: + 0xd3e95 (0x7f986f053e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #3: + 0x8609 (0x7f987409a609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #4: clone + 0x43 (0x7f9873e65353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default3]:[rank59]: Traceback (most recent call last): [default3]:[rank59]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default3]:[rank59]: trainer.train(dataloader) [default3]:[rank59]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default3]:[rank59]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default3]:[rank59]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default3]:[rank59]: outputs = self.pipeline_engine.train_batch_iter( [default3]:[rank59]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter [default3]:[rank59]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default3]:[rank59]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default3]:[rank59]: output = model(**micro_batch) [default3]:[rank59]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default3]:[rank59]: return self._call_impl(*args, **kwargs) [default3]:[rank59]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default3]:[rank59]: return forward_call(*args, **kwargs) [default3]:[rank59]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default3]:[rank59]: sharded_logits = self.model( [default3]:[rank59]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default3]:[rank59]: return self._call_impl(*args, **kwargs) [default3]:[rank59]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default3]:[rank59]: return forward_call(*args, **kwargs) [default3]:[rank59]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default3]:[rank59]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default3]:[rank59]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default3]:[rank59]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default3]:[rank59]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default3]:[rank59]: return self._call_impl(*args, **kwargs) [default3]:[rank59]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default3]:[rank59]: return forward_call(*args, **kwargs) [default3]:[rank59]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default3]:[rank59]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default3]:[rank59]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default3]:[rank59]: pipeline_state.run_communication() [default3]:[rank59]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default3]:[rank59]: recv_activation_tensor = recv_activation() [default3]:[rank59]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default3]:[rank59]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default3]:[rank59]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default3]:[rank59]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default3]:[rank59]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default3]:[rank59]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default3]:[rank59]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default3]:[rank59]: dist.recv( [default3]:[rank59]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default3]:[rank59]: return func(*args, **kwargs) [default3]:[rank59]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default3]:[rank59]: pg.recv([tensor], group_src_rank, tag).wait() [default3]:[rank59]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 1. [default3]:[rank51]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 14, last enqueued NCCL work: 15, last completed NCCL work: 13. [default3]:[rank51]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default3]:[rank51]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default3]:[rank51]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600058 milliseconds before timing out. [default3]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f6c32b4e897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f6c33e27c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f6c33e2ca80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f6c33e2ddcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #4: + 0xd3e95 (0x7f6c7f8c6e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #5: + 0x8609 (0x7f6c8490d609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #6: clone + 0x43 (0x7f6c846d8353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default3]:terminate called after throwing an instance of 'c10::DistBackendError' [default3]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600058 milliseconds before timing out. [default3]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f6c32b4e897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f6c33e27c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f6c33e2ca80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f6c33e2ddcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #4: + 0xd3e95 (0x7f6c7f8c6e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #5: + 0x8609 (0x7f6c8490d609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #6: clone + 0x43 (0x7f6c846d8353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default3]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f6c32b4e897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: + 0xe32119 (0x7f6c33ab1119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: + 0xd3e95 (0x7f6c7f8c6e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #3: + 0x8609 (0x7f6c8490d609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #4: clone + 0x43 (0x7f6c846d8353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default6]:[rank22]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 34, last enqueued NCCL work: 35, last completed NCCL work: 33. [default6]:[rank22]:[E ProcessGroupNCCL.cpp:577] [Rank 1] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default6]:[rank22]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default6]:[rank22]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600039 milliseconds before timing out. [default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f3c5d180897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f3c5e459c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f3c5e45ea80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f3c5e45fdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7f3ca9ef8e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7f3caef3f609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7f3caed0a353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:terminate called after throwing an instance of 'c10::DistBackendError' [default6]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600039 milliseconds before timing out. [default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f3c5d180897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f3c5e459c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f3c5e45ea80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f3c5e45fdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7f3ca9ef8e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7f3caef3f609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7f3caed0a353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f3c5d180897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: + 0xe32119 (0x7f3c5e0e3119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: + 0xd3e95 (0x7f3ca9ef8e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #3: + 0x8609 (0x7f3caef3f609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #4: clone + 0x43 (0x7f3caed0a353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:[rank54]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 14, last enqueued NCCL work: 15, last completed NCCL work: 13. [default6]:[rank54]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default6]:[rank54]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default6]:[rank54]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600075 milliseconds before timing out. [default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fd213188897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fd214461c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fd214466a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fd214467dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7fd25ff00e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7fd264f47609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7fd264d12353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:terminate called after throwing an instance of 'c10::DistBackendError' [default6]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600075 milliseconds before timing out. [default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fd213188897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fd214461c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fd214466a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fd214467dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7fd25ff00e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7fd264f47609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7fd264d12353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fd213188897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: + 0xe32119 (0x7fd2140eb119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: + 0xd3e95 (0x7fd25ff00e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #3: + 0x8609 (0x7fd264f47609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #4: clone + 0x43 (0x7fd264d12353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default4]:[rank60]: Traceback (most recent call last): [default4]:[rank60]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default4]:[rank60]: trainer.train(dataloader) [default4]:[rank60]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default4]:[rank60]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default4]:[rank60]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default4]:[rank60]: outputs = self.pipeline_engine.train_batch_iter( [default4]:[rank60]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter [default4]:[rank60]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default4]:[rank60]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default4]:[rank60]: output = model(**micro_batch) [default4]:[rank60]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default4]:[rank60]: return self._call_impl(*args, **kwargs) [default4]:[rank60]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default4]:[rank60]: return forward_call(*args, **kwargs) [default4]:[rank60]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default4]:[rank60]: sharded_logits = self.model( [default4]:[rank60]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default4]:[rank60]: return self._call_impl(*args, **kwargs) [default4]:[rank60]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default4]:[rank60]: return forward_call(*args, **kwargs) [default4]:[rank60]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default4]:[rank60]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default4]:[rank60]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default4]:[rank60]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default4]:[rank60]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default4]:[rank60]: return self._call_impl(*args, **kwargs) [default4]:[rank60]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default4]:[rank60]: return forward_call(*args, **kwargs) [default4]:[rank60]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default4]:[rank60]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default4]:[rank60]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default4]:[rank60]: pipeline_state.run_communication() [default4]:[rank60]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default4]:[rank60]: recv_activation_tensor = recv_activation() [default4]:[rank60]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default4]:[rank60]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default4]:[rank60]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default4]:[rank60]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default4]:[rank60]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default4]:[rank60]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default4]:[rank60]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default4]:[rank60]: dist.recv( [default4]:[rank60]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default4]:[rank60]: return func(*args, **kwargs) [default4]:[rank60]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default4]:[rank60]: pg.recv([tensor], group_src_rank, tag).wait() [default4]:[rank60]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 1. [default5]:[rank61]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 14, last enqueued NCCL work: 15, last completed NCCL work: 13. [default5]:[rank61]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default5]:[rank61]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default5]:[rank61]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600019 milliseconds before timing out. [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fc77ee09897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fc7800e2c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fc7800e7a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fc7800e8dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #4: + 0xd3e95 (0x7fc7cbb81e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #5: + 0x8609 (0x7fc7d0bc8609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #6: clone + 0x43 (0x7fc7d0993353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default5]:terminate called after throwing an instance of 'c10::DistBackendError' [default5]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600019 milliseconds before timing out. [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fc77ee09897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fc7800e2c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fc7800e7a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fc7800e8dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #4: + 0xd3e95 (0x7fc7cbb81e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #5: + 0x8609 (0x7fc7d0bc8609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #6: clone + 0x43 (0x7fc7d0993353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default5]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fc77ee09897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: + 0xe32119 (0x7fc77fd6c119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: + 0xd3e95 (0x7fc7cbb81e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #3: + 0x8609 (0x7fc7d0bc8609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #4: clone + 0x43 (0x7fc7d0993353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default0]:[rank56]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 14, last enqueued NCCL work: 15, last completed NCCL work: 13. [default0]:[rank56]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default0]:[rank56]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default0]:[rank56]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600085 milliseconds before timing out. [default0]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fdf765ea897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:[rank63]: Traceback (most recent call last): [default0]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fdf778c3c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:[rank63]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in [default7]:[rank63]: trainer.train(dataloader) [default7]:[rank63]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train [default7]:[rank63]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader) [default7]:[rank63]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step [default7]:[rank63]: outputs = self.pipeline_engine.train_batch_iter( [default0]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fdf778c8a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fdf778c9dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #4: + 0xd3e95 (0x7fdfc3362e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #5: + 0x8609 (0x7fdfc83a9609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:[rank63]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter [default0]:frame #6: clone + 0x43 (0x7fdfc8174353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default7]:[rank63]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default7]:[rank63]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default7]:[rank63]: output = model(**micro_batch) [default0]:terminate called after throwing an instance of 'c10::DistBackendError' [default0]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600085 milliseconds before timing out. [default0]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fdf765ea897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fdf778c3c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fdf778c8a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fdf778c9dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #4: + 0xd3e95 (0x7fdfc3362e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #5: + 0x8609 (0x7fdfc83a9609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:[rank63]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default0]:frame #6: clone + 0x43 (0x7fdfc8174353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default0]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default7]:[rank63]: return self._call_impl(*args, **kwargs) [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fdf765ea897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:[rank63]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default7]:[rank63]: return forward_call(*args, **kwargs) [default7]:[rank63]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default0]:frame #1: + 0xe32119 (0x7fdf7754d119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:[rank63]: sharded_logits = self.model( [default0]:frame #2: + 0xd3e95 (0x7fdfc3362e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:[rank63]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default7]:[rank63]: return self._call_impl(*args, **kwargs) [default7]:[rank63]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default0]:frame #3: + 0x8609 (0x7fdfc83a9609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #4: clone + 0x43 (0x7fdfc8174353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default7]:[rank63]: return forward_call(*args, **kwargs) [default7]:[rank63]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default7]:[rank63]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default7]:[rank63]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default7]:[rank63]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default7]:[rank63]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl [default7]:[rank63]: return self._call_impl(*args, **kwargs) [default7]:[rank63]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default7]:[rank63]: return forward_call(*args, **kwargs) [default7]:[rank63]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default7]:[rank63]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default7]:[rank63]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default7]:[rank63]: pipeline_state.run_communication() [default7]:[rank63]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default7]:[rank63]: recv_activation_tensor = recv_activation() [default7]:[rank63]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default7]:[rank63]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default7]:[rank63]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default7]:[rank63]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default7]:[rank63]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default7]:[rank63]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default7]:[rank63]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 269, in _recv_meta [default7]:[rank63]: dist.recv( [default7]:[rank63]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 75, in wrapper [default7]:[rank63]: return func(*args, **kwargs) [default7]:[rank63]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1932, in recv [default7]:[rank63]: pg.recv([tensor], group_src_rank, tag).wait() [default7]:[rank63]: torch.distributed.DistBackendError: NCCL communicator was aborted on rank 1. [default0]:[rank24]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 34, last enqueued NCCL work: 35, last completed NCCL work: 33. [default0]:[rank24]:[E ProcessGroupNCCL.cpp:577] [Rank 1] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default0]:[rank24]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default0]:[rank24]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600009 milliseconds before timing out. [default0]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fa0854dd897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fa0867b6c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fa0867bba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fa0867bcdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #4: + 0xd3e95 (0x7fa0d2255e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #5: + 0x8609 (0x7fa0d729c609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #6: clone + 0x43 (0x7fa0d7067353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default0]:terminate called after throwing an instance of 'c10::DistBackendError' [default0]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600009 milliseconds before timing out. [default0]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fa0854dd897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fa0867b6c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fa0867bba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fa0867bcdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #4: + 0xd3e95 (0x7fa0d2255e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #5: + 0x8609 (0x7fa0d729c609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #6: clone + 0x43 (0x7fa0d7067353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default0]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fa0854dd897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: + 0xe32119 (0x7fa086440119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: + 0xd3e95 (0x7fa0d2255e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #3: + 0x8609 (0x7fa0d729c609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #4: clone + 0x43 (0x7fa0d7067353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default5]:[rank29]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 34, last enqueued NCCL work: 35, last completed NCCL work: 33. [default5]:[rank29]:[E ProcessGroupNCCL.cpp:577] [Rank 1] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default5]:[rank29]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default5]:[rank29]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600062 milliseconds before timing out. [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fa4d22d3897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fa4d35acc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fa4d35b1a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fa4d35b2dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #4: + 0xd3e95 (0x7fa51f04be95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #5: + 0x8609 (0x7fa524092609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #6: clone + 0x43 (0x7fa523e5d353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default5]:terminate called after throwing an instance of 'c10::DistBackendError' [default5]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600062 milliseconds before timing out. [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fa4d22d3897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fa4d35acc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fa4d35b1a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fa4d35b2dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #4: + 0xd3e95 (0x7fa51f04be95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #5: + 0x8609 (0x7fa524092609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #6: clone + 0x43 (0x7fa523e5d353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default5]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fa4d22d3897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default5]:frame #1: + 0xe32119 (0x7fa4d3236119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default5]:frame #2: + 0xd3e95 (0x7fa51f04be95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default5]:frame #3: + 0x8609 (0x7fa524092609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default5]:frame #4: clone + 0x43 (0x7fa523e5d353 in /lib/x86_64-linux-gnu/libc.so.6) [default5]: [default4]:[rank28]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 34, last enqueued NCCL work: 35, last completed NCCL work: 33. [default4]:[rank28]:[E ProcessGroupNCCL.cpp:577] [Rank 1] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default4]:[rank28]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default4]:[rank28]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600017 milliseconds before timing out. [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f06f5aa8897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f06f6d81c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f06f6d86a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f06f6d87dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f0742820e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f0747867609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f0747632353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default4]:terminate called after throwing an instance of 'c10::DistBackendError' [default4]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600017 milliseconds before timing out. [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f06f5aa8897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f06f6d81c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f06f6d86a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f06f6d87dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f0742820e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f0747867609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f0747632353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default4]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f06f5aa8897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: + 0xe32119 (0x7f06f6a0b119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: + 0xd3e95 (0x7f0742820e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #3: + 0x8609 (0x7f0747867609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #4: clone + 0x43 (0x7f0747632353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default3]:[rank27]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 34, last enqueued NCCL work: 35, last completed NCCL work: 33. [default3]:[rank27]:[E ProcessGroupNCCL.cpp:577] [Rank 1] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default3]:[rank27]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default3]:[rank27]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600085 milliseconds before timing out. [default3]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f69796c0897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f697a999c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f697a99ea80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f697a99fdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #4: + 0xd3e95 (0x7f69c6438e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #5: + 0x8609 (0x7f69cb47f609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #6: clone + 0x43 (0x7f69cb24a353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default3]:terminate called after throwing an instance of 'c10::DistBackendError' [default3]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600085 milliseconds before timing out. [default3]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f69796c0897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f697a999c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f697a99ea80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f697a99fdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #4: + 0xd3e95 (0x7f69c6438e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #5: + 0x8609 (0x7f69cb47f609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #6: clone + 0x43 (0x7f69cb24a353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default3]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f69796c0897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: + 0xe32119 (0x7f697a623119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: + 0xd3e95 (0x7f69c6438e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #3: + 0x8609 (0x7f69cb47f609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #4: clone + 0x43 (0x7f69cb24a353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default0]:[rank32]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 26, last enqueued NCCL work: 27, last completed NCCL work: 25. [default0]:[rank32]:[E ProcessGroupNCCL.cpp:577] [Rank 2] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default0]:[rank32]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default0]:[rank32]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600012 milliseconds before timing out. [default0]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fbf38361897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fbf3963ac62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fbf3963fa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fbf39640dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #4: + 0xd3e95 (0x7fbf850d9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #5: + 0x8609 (0x7fbf8a120609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #6: clone + 0x43 (0x7fbf89eeb353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default0]:terminate called after throwing an instance of 'c10::DistBackendError' [default0]: what(): [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600012 milliseconds before timing out. [default0]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fbf38361897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fbf3963ac62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fbf3963fa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fbf39640dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #4: + 0xd3e95 (0x7fbf850d9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #5: + 0x8609 (0x7fbf8a120609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #6: clone + 0x43 (0x7fbf89eeb353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default0]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fbf38361897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: + 0xe32119 (0x7fbf392c4119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: + 0xd3e95 (0x7fbf850d9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #3: + 0x8609 (0x7fbf8a120609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #4: clone + 0x43 (0x7fbf89eeb353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default2]:[rank26]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 34, last enqueued NCCL work: 35, last completed NCCL work: 33. [default2]:[rank26]:[E ProcessGroupNCCL.cpp:577] [Rank 1] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default2]:[rank26]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default2]:[rank26]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600072 milliseconds before timing out. [default2]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fdd8e20f897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fdd8f4e8c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fdd8f4eda80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fdd8f4eedcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7fdddaf87e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #5: + 0x8609 (0x7fdddffce609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7fdddfd99353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default2]:terminate called after throwing an instance of 'c10::DistBackendError' [default2]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600072 milliseconds before timing out. [default2]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fdd8e20f897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fdd8f4e8c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fdd8f4eda80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fdd8f4eedcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7fdddaf87e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #5: + 0x8609 (0x7fdddffce609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7fdddfd99353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default2]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fdd8e20f897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: + 0xe32119 (0x7fdd8f172119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: + 0xd3e95 (0x7fdddaf87e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #3: + 0x8609 (0x7fdddffce609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #4: clone + 0x43 (0x7fdddfd99353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default3]:[rank59]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 14, last enqueued NCCL work: 15, last completed NCCL work: 13. [default3]:[rank59]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default3]:[rank59]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default3]:[rank59]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600073 milliseconds before timing out. [default3]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f6a2ce61897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f6a2e13ac62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f6a2e13fa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f6a2e140dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #4: + 0xd3e95 (0x7f6a79bd9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #5: + 0x8609 (0x7f6a7ec20609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #6: clone + 0x43 (0x7f6a7e9eb353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default3]:terminate called after throwing an instance of 'c10::DistBackendError' [default3]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600073 milliseconds before timing out. [default3]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f6a2ce61897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f6a2e13ac62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f6a2e13fa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f6a2e140dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #4: + 0xd3e95 (0x7f6a79bd9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #5: + 0x8609 (0x7f6a7ec20609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #6: clone + 0x43 (0x7f6a7e9eb353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default3]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f6a2ce61897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: + 0xe32119 (0x7f6a2ddc4119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: + 0xd3e95 (0x7f6a79bd9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #3: + 0x8609 (0x7f6a7ec20609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #4: clone + 0x43 (0x7f6a7e9eb353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default7]:[rank31]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 34, last enqueued NCCL work: 35, last completed NCCL work: 33. [default7]:[rank31]:[E ProcessGroupNCCL.cpp:577] [Rank 1] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default7]:[rank31]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default7]:[rank31]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600007 milliseconds before timing out. [default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7ff09997c897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7ff09ac55c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7ff09ac5aa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7ff09ac5bdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #4: + 0xd3e95 (0x7ff0e66f4e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #5: + 0x8609 (0x7ff0eb73b609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #6: clone + 0x43 (0x7ff0eb506353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default7]:terminate called after throwing an instance of 'c10::DistBackendError' [default7]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600007 milliseconds before timing out. [default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7ff09997c897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7ff09ac55c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7ff09ac5aa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7ff09ac5bdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #4: + 0xd3e95 (0x7ff0e66f4e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #5: + 0x8609 (0x7ff0eb73b609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #6: clone + 0x43 (0x7ff0eb506353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default7]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7ff09997c897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: + 0xe32119 (0x7ff09a8df119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: + 0xd3e95 (0x7ff0e66f4e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #3: + 0x8609 (0x7ff0eb73b609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #4: clone + 0x43 (0x7ff0eb506353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default6]:[rank30]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 34, last enqueued NCCL work: 35, last completed NCCL work: 33. [default6]:[rank30]:[E ProcessGroupNCCL.cpp:577] [Rank 1] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default6]:[rank30]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default6]:[rank30]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600036 milliseconds before timing out. [default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f180b020897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f180c2f9c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f180c2fea80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f180c2ffdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7f1857d98e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7f185cddf609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7f185cbaa353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:terminate called after throwing an instance of 'c10::DistBackendError' [default6]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=34, OpType=SEND, NumelIn=6, NumelOut=6, Timeout(ms)=600000) ran for 600036 milliseconds before timing out. [default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f180b020897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f180c2f9c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f180c2fea80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f180c2ffdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7f1857d98e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7f185cddf609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7f185cbaa353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f180b020897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: + 0xe32119 (0x7f180bf83119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: + 0xd3e95 (0x7f1857d98e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #3: + 0x8609 (0x7f185cddf609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #4: clone + 0x43 (0x7f185cbaa353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default4]:[rank60]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 14, last enqueued NCCL work: 15, last completed NCCL work: 13. [default4]:[rank60]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default4]:[rank60]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default4]:[rank60]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600056 milliseconds before timing out. [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f1fe6066897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f1fe733fc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f1fe7344a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f1fe7345dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f2032ddee95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f2037e25609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f2037bf0353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default4]:terminate called after throwing an instance of 'c10::DistBackendError' [default4]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600056 milliseconds before timing out. [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f1fe6066897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f1fe733fc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f1fe7344a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f1fe7345dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f2032ddee95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f2037e25609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f2037bf0353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default4]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f1fe6066897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: + 0xe32119 (0x7f1fe6fc9119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: + 0xd3e95 (0x7f2032ddee95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #3: + 0x8609 (0x7f2037e25609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #4: clone + 0x43 (0x7f2037bf0353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: W0703 09:55:21.580000 140322518665024 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1133780 closing signal SIGTERM W0703 09:55:21.580000 140322518665024 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1133782 closing signal SIGTERM W0703 09:55:21.580000 140322518665024 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1133783 closing signal SIGTERM W0703 09:55:21.581000 140322518665024 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1133784 closing signal SIGTERM W0703 09:55:21.581000 140322518665024 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1133785 closing signal SIGTERM W0703 09:55:21.581000 140322518665024 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1133786 closing signal SIGTERM W0703 09:55:21.581000 140322518665024 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1133787 closing signal SIGTERM W0703 09:55:21.610000 139651575428928 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 3289913 closing signal SIGTERM W0703 09:55:21.610000 139651575428928 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 3289915 closing signal SIGTERM W0703 09:55:21.610000 139651575428928 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 3289916 closing signal SIGTERM W0703 09:55:21.610000 139651575428928 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 3289917 closing signal SIGTERM W0703 09:55:21.610000 139651575428928 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 3289918 closing signal SIGTERM W0703 09:55:21.610000 139651575428928 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 3289919 closing signal SIGTERM W0703 09:55:21.611000 139651575428928 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 3289920 closing signal SIGTERM W0703 09:55:21.618000 140180047779648 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1212113 closing signal SIGTERM W0703 09:55:21.618000 140180047779648 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1212115 closing signal SIGTERM W0703 09:55:21.618000 140180047779648 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1212116 closing signal SIGTERM W0703 09:55:21.618000 140180047779648 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1212117 closing signal SIGTERM W0703 09:55:21.618000 140180047779648 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1212118 closing signal SIGTERM W0703 09:55:21.618000 140180047779648 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1212119 closing signal SIGTERM W0703 09:55:21.618000 140180047779648 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1212120 closing signal SIGTERM [default4]:[rank44]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 26, last enqueued NCCL work: 27, last completed NCCL work: 25. [default4]:[rank44]:[E ProcessGroupNCCL.cpp:577] [Rank 2] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default4]:[rank44]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default4]:[rank44]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600006 milliseconds before timing out. [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f5fa037e897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f5fa1657c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f5fa165ca80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f5fa165ddcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f5fed0f6e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f5ff213d609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f5ff1f08353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default4]:terminate called after throwing an instance of 'c10::DistBackendError' [default4]: what(): [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600006 milliseconds before timing out. [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f5fa037e897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f5fa1657c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f5fa165ca80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f5fa165ddcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f5fed0f6e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f5ff213d609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f5ff1f08353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default4]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f5fa037e897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: + 0xe32119 (0x7f5fa12e1119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: + 0xd3e95 (0x7f5fed0f6e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #3: + 0x8609 (0x7f5ff213d609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #4: clone + 0x43 (0x7f5ff1f08353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default7]:[rank63]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 14, last enqueued NCCL work: 15, last completed NCCL work: 13. [default7]:[rank63]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default7]:[rank63]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default7]:[rank63]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600064 milliseconds before timing out. [default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f08bcd41897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f08be01ac62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f08be01fa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f08be020dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #4: + 0xd3e95 (0x7f0909ab9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #5: + 0x8609 (0x7f090eb00609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #6: clone + 0x43 (0x7f090e8cb353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default7]:terminate called after throwing an instance of 'c10::DistBackendError' [default7]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=14, OpType=SEND, NumelIn=8388608, NumelOut=8388608, Timeout(ms)=600000) ran for 600064 milliseconds before timing out. [default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f08bcd41897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f08be01ac62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f08be01fa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f08be020dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #4: + 0xd3e95 (0x7f0909ab9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #5: + 0x8609 (0x7f090eb00609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #6: clone + 0x43 (0x7f090e8cb353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default7]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f08bcd41897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: + 0xe32119 (0x7f08bdca4119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: + 0xd3e95 (0x7f0909ab9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #3: + 0x8609 (0x7f090eb00609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #4: clone + 0x43 (0x7f090e8cb353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default6]:[rank38]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 26, last enqueued NCCL work: 27, last completed NCCL work: 25. [default6]:[rank38]:[E ProcessGroupNCCL.cpp:577] [Rank 2] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default6]:[rank38]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default6]:[rank38]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600004 milliseconds before timing out. [default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f06416c3897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f064299cc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f06429a1a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f06429a2dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7f068e43be95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7f0693482609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7f069324d353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:terminate called after throwing an instance of 'c10::DistBackendError' [default6]: what(): [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600004 milliseconds before timing out. [default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f06416c3897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f064299cc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f06429a1a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f06429a2dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7f068e43be95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7f0693482609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7f069324d353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f06416c3897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: + 0xe32119 (0x7f0642626119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: + 0xd3e95 (0x7f068e43be95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #3: + 0x8609 (0x7f0693482609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #4: clone + 0x43 (0x7f069324d353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default2]:[rank42]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 26, last enqueued NCCL work: 27, last completed NCCL work: 25. [default2]:[rank42]:[E ProcessGroupNCCL.cpp:577] [Rank 2] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default2]:[rank42]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default2]:[rank42]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600049 milliseconds before timing out. [default2]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fd2c03dd897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fd2c16b6c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fd2c16bba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fd2c16bcdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7fd30d155e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #5: + 0x8609 (0x7fd31219c609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7fd311f67353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default2]:terminate called after throwing an instance of 'c10::DistBackendError' [default2]: what(): [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600049 milliseconds before timing out. [default2]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fd2c03dd897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fd2c16b6c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fd2c16bba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fd2c16bcdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7fd30d155e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #5: + 0x8609 (0x7fd31219c609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7fd311f67353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default2]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default2]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fd2c03dd897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: + 0xe32119 (0x7fd2c1340119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: + 0xd3e95 (0x7fd30d155e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #3: + 0x8609 (0x7fd31219c609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #4: clone + 0x43 (0x7fd311f67353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: [default3]:[rank35]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 26, last enqueued NCCL work: 27, last completed NCCL work: 25. [default3]:[rank35]:[E ProcessGroupNCCL.cpp:577] [Rank 2] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default3]:[rank35]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default3]:[rank35]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600077 milliseconds before timing out. [default3]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f8424002897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f84252dbc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f84252e0a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f84252e1dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #4: + 0xd3e95 (0x7f8470d7ae95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #5: + 0x8609 (0x7f8475dc1609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #6: clone + 0x43 (0x7f8475b8c353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default3]:terminate called after throwing an instance of 'c10::DistBackendError' [default3]: what(): [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600077 milliseconds before timing out. [default3]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f8424002897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f84252dbc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f84252e0a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f84252e1dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #4: + 0xd3e95 (0x7f8470d7ae95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #5: + 0x8609 (0x7f8475dc1609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #6: clone + 0x43 (0x7f8475b8c353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default3]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f8424002897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: + 0xe32119 (0x7f8424f65119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: + 0xd3e95 (0x7f8470d7ae95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #3: + 0x8609 (0x7f8475dc1609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #4: clone + 0x43 (0x7f8475b8c353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default0]:[rank40]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 26, last enqueued NCCL work: 27, last completed NCCL work: 25. [default0]:[rank40]:[E ProcessGroupNCCL.cpp:577] [Rank 2] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default0]:[rank40]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default0]:[rank40]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600020 milliseconds before timing out. [default0]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f1c8af7d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f1c8c256c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f1c8c25ba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f1c8c25cdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #4: + 0xd3e95 (0x7f1cd7cf5e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #5: + 0x8609 (0x7f1cdcd3c609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #6: clone + 0x43 (0x7f1cdcb07353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default0]:terminate called after throwing an instance of 'c10::DistBackendError' [default0]: what(): [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600020 milliseconds before timing out. [default0]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f1c8af7d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f1c8c256c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f1c8c25ba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f1c8c25cdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #4: + 0xd3e95 (0x7f1cd7cf5e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #5: + 0x8609 (0x7f1cdcd3c609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #6: clone + 0x43 (0x7f1cdcb07353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default0]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default0]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f1c8af7d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default0]:frame #1: + 0xe32119 (0x7f1c8bee0119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default0]:frame #2: + 0xd3e95 (0x7f1cd7cf5e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default0]:frame #3: + 0x8609 (0x7f1cdcd3c609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default0]:frame #4: clone + 0x43 (0x7f1cdcb07353 in /lib/x86_64-linux-gnu/libc.so.6) [default0]: [default6]:[rank46]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 26, last enqueued NCCL work: 27, last completed NCCL work: 25. [default6]:[rank46]:[E ProcessGroupNCCL.cpp:577] [Rank 2] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default6]:[rank46]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default6]:[rank46]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600042 milliseconds before timing out. [default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f7fd5c2a897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f7fd6f03c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f7fd6f08a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f7fd6f09dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7f80229a2e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7f80279e9609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7f80277b4353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:terminate called after throwing an instance of 'c10::DistBackendError' [default6]: what(): [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600042 milliseconds before timing out. [default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f7fd5c2a897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f7fd6f03c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f7fd6f08a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f7fd6f09dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7f80229a2e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7f80279e9609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7f80277b4353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default6]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f7fd5c2a897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: + 0xe32119 (0x7f7fd6b8d119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: + 0xd3e95 (0x7f80229a2e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #3: + 0x8609 (0x7f80279e9609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #4: clone + 0x43 (0x7f80277b4353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default7]:[rank47]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 26, last enqueued NCCL work: 27, last completed NCCL work: 25. [default7]:[rank47]:[E ProcessGroupNCCL.cpp:577] [Rank 2] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default7]:[rank47]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default7]:[rank47]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600005 milliseconds before timing out. [default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fccd4c2d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fccd5f06c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fccd5f0ba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fccd5f0cdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #4: + 0xd3e95 (0x7fcd219a5e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #5: + 0x8609 (0x7fcd269ec609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #6: clone + 0x43 (0x7fcd267b7353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default7]:terminate called after throwing an instance of 'c10::DistBackendError' [default7]: what(): [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600005 milliseconds before timing out. [default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fccd4c2d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7fccd5f06c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fccd5f0ba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fccd5f0cdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #4: + 0xd3e95 (0x7fcd219a5e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #5: + 0x8609 (0x7fcd269ec609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #6: clone + 0x43 (0x7fcd267b7353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default7]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fccd4c2d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default7]:frame #1: + 0xe32119 (0x7fccd5b90119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default7]:frame #2: + 0xd3e95 (0x7fcd219a5e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default7]:frame #3: + 0x8609 (0x7fcd269ec609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default7]:frame #4: clone + 0x43 (0x7fcd267b7353 in /lib/x86_64-linux-gnu/libc.so.6) [default7]: [default3]:[rank43]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 26, last enqueued NCCL work: 27, last completed NCCL work: 25. [default3]:[rank43]:[E ProcessGroupNCCL.cpp:577] [Rank 2] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data. [default3]:[rank43]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default3]:[rank43]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600081 milliseconds before timing out. [default3]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f1652b54897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f1653e2dc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f1653e32a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f1653e33dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #4: + 0xd3e95 (0x7f169f8cce95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #5: + 0x8609 (0x7f16a4913609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #6: clone + 0x43 (0x7f16a46de353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default3]:terminate called after throwing an instance of 'c10::DistBackendError' [default3]: what(): [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=26, OpType=SEND, NumelIn=4096, NumelOut=4096, Timeout(ms)=600000) ran for 600081 milliseconds before timing out. [default3]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f1652b54897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional > >) + 0x1d2 (0x7f1653e2dc62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f1653e32a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f1653e33dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #4: + 0xd3e95 (0x7f169f8cce95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #5: + 0x8609 (0x7f16a4913609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #6: clone + 0x43 (0x7f16a46de353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: [default3]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first): [default3]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f1652b54897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default3]:frame #1: + 0xe32119 (0x7f1653ab7119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default3]:frame #2: + 0xd3e95 (0x7f169f8cce95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default3]:frame #3: + 0x8609 (0x7f16a4913609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default3]:frame #4: clone + 0x43 (0x7f16a46de353 in /lib/x86_64-linux-gnu/libc.so.6) [default3]: E0703 09:55:24.224000 140180047779648 torch/distributed/elastic/multiprocessing/api.py:826] failed (exitcode: -6) local_rank: 1 (pid: 1212114) of binary: /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/python3.10 Traceback (most recent call last): File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/torchrun", line 8, in sys.exit(main()) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper return f(*args, **kwargs) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 879, in main run(args) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 870, in run elastic_launch( File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 132, in __call__ return launch_agent(self._config, self._entrypoint, list(args)) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 263, in launch_agent raise ChildFailedError( torch.distributed.elastic.multiprocessing.errors.ChildFailedError: ============================================================ /fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py FAILED ------------------------------------------------------------ Failures: ------------------------------------------------------------ Root Cause (first observed failure): [0]: time : 2024-07-03_09:55:21 host : ip-26-0-160-192.ec2.internal rank : 1 (local_rank: 1) exitcode : -6 (pid: 1212114) error_file: traceback : Signal 6 (SIGABRT) received by PID 1212114 ============================================================ srun: error: ip-26-0-160-192: task 0: Exited with exit code 1 W0703 09:55:25.395000 140494136764160 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1252] The node 'ip-26-0-169-132.ec2.internal_2476171_0' has failed to send a keep-alive heartbeat to the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:25.525000 139730091611904 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1252] The node 'ip-26-0-169-86.ec2.internal_1905143_0' has failed to send a keep-alive heartbeat to the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:25.527000 140548964558592 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1252] The node 'ip-26-0-168-238.ec2.internal_1932497_0' has failed to send a keep-alive heartbeat to the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:25.605000 140084244485888 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1252] The node 'ip-26-0-172-73.ec2.internal_975153_0' has failed to send a keep-alive heartbeat to the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:26.345000 139645914695424 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1252] The node 'ip-26-0-163-226.ec2.internal_3289843_0' has failed to send a keep-alive heartbeat to the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:26.376000 140017629738752 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1252] The node 'ip-26-0-163-220.ec2.internal_843202_0' has failed to send a keep-alive heartbeat to the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:26.380000 140316857931520 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1252] The node 'ip-26-0-172-57.ec2.internal_1133710_0' has failed to send a keep-alive heartbeat to the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:26.640000 140023290472256 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 843272 closing signal SIGTERM W0703 09:55:26.640000 140023290472256 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 843274 closing signal SIGTERM W0703 09:55:26.640000 140023290472256 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 843275 closing signal SIGTERM W0703 09:55:26.640000 140023290472256 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 843276 closing signal SIGTERM W0703 09:55:26.640000 140023290472256 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 843277 closing signal SIGTERM W0703 09:55:26.640000 140023290472256 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 843278 closing signal SIGTERM W0703 09:55:26.640000 140023290472256 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 843279 closing signal SIGTERM W0703 09:55:26.650000 139735752345408 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1905214 closing signal SIGTERM W0703 09:55:26.650000 139735752345408 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1905215 closing signal SIGTERM W0703 09:55:26.650000 139735752345408 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1905216 closing signal SIGTERM W0703 09:55:26.650000 139735752345408 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1905217 closing signal SIGTERM W0703 09:55:26.650000 139735752345408 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1905218 closing signal SIGTERM W0703 09:55:26.650000 139735752345408 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1905219 closing signal SIGTERM W0703 09:55:26.651000 139735752345408 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1905220 closing signal SIGTERM W0703 09:55:26.651000 139735752345408 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1905221 closing signal SIGTERM W0703 09:55:26.656000 140089905219392 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 975227 closing signal SIGTERM W0703 09:55:26.656000 140089905219392 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 975228 closing signal SIGTERM W0703 09:55:26.656000 140089905219392 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 975230 closing signal SIGTERM W0703 09:55:26.688000 140499797497664 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2476241 closing signal SIGTERM W0703 09:55:26.688000 140499797497664 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2476243 closing signal SIGTERM W0703 09:55:26.688000 140499797497664 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2476244 closing signal SIGTERM W0703 09:55:26.688000 140499797497664 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2476245 closing signal SIGTERM W0703 09:55:26.688000 140499797497664 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2476246 closing signal SIGTERM W0703 09:55:26.688000 140499797497664 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2476247 closing signal SIGTERM W0703 09:55:26.688000 140499797497664 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2476248 closing signal SIGTERM W0703 09:55:26.774000 140554625292096 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1932568 closing signal SIGTERM W0703 09:55:26.774000 140554625292096 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1932569 closing signal SIGTERM W0703 09:55:26.774000 140554625292096 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1932570 closing signal SIGTERM W0703 09:55:26.774000 140554625292096 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1932571 closing signal SIGTERM W0703 09:55:26.774000 140554625292096 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1932572 closing signal SIGTERM W0703 09:55:26.774000 140554625292096 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1932573 closing signal SIGTERM W0703 09:55:26.774000 140554625292096 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1932574 closing signal SIGTERM W0703 09:55:26.774000 140554625292096 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1932575 closing signal SIGTERM E0703 09:55:27.220000 139651575428928 torch/distributed/elastic/multiprocessing/api.py:826] failed (exitcode: -6) local_rank: 1 (pid: 3289914) of binary: /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/python3.10 W0703 09:55:27.232000 139651575428928 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-163-226.ec2.internal_3289843_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:27.271000 139651575428928 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-163-226.ec2.internal_3289843_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:27.279000 139651575428928 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-163-226.ec2.internal_3289843_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. Traceback (most recent call last): File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/torchrun", line 8, in sys.exit(main()) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper return f(*args, **kwargs) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 879, in main run(args) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 870, in run elastic_launch( File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 132, in __call__ return launch_agent(self._config, self._entrypoint, list(args)) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 263, in launch_agent raise ChildFailedError( torch.distributed.elastic.multiprocessing.errors.ChildFailedError: ============================================================ /fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py FAILED ------------------------------------------------------------ Failures: ------------------------------------------------------------ Root Cause (first observed failure): [0]: time : 2024-07-03_09:55:21 host : ip-26-0-163-226.ec2.internal rank : 17 (local_rank: 1) exitcode : -6 (pid: 3289914) error_file: traceback : Signal 6 (SIGABRT) received by PID 3289914 ============================================================ E0703 09:55:27.607000 140089905219392 torch/distributed/elastic/multiprocessing/api.py:826] failed (exitcode: -6) local_rank: 0 (pid: 975223) of binary: /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/python3.10 W0703 09:55:27.619000 140089905219392 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-172-73.ec2.internal_975153_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:27.647000 140089905219392 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-172-73.ec2.internal_975153_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:27.667000 140089905219392 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-172-73.ec2.internal_975153_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. Traceback (most recent call last): File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/torchrun", line 8, in sys.exit(main()) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper return f(*args, **kwargs) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 879, in main run(args) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 870, in run elastic_launch( File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 132, in __call__ return launch_agent(self._config, self._entrypoint, list(args)) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 263, in launch_agent raise ChildFailedError( torch.distributed.elastic.multiprocessing.errors.ChildFailedError: ============================================================ /fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py FAILED ------------------------------------------------------------ Failures: [1]: time : 2024-07-03_09:55:26 host : ip-26-0-172-73.ec2.internal rank : 57 (local_rank: 1) exitcode : -6 (pid: 975224) error_file: traceback : Signal 6 (SIGABRT) received by PID 975224 [2]: time : 2024-07-03_09:55:26 host : ip-26-0-172-73.ec2.internal rank : 58 (local_rank: 2) exitcode : -6 (pid: 975225) error_file: traceback : Signal 6 (SIGABRT) received by PID 975225 [3]: time : 2024-07-03_09:55:26 host : ip-26-0-172-73.ec2.internal rank : 59 (local_rank: 3) exitcode : -6 (pid: 975226) error_file: traceback : Signal 6 (SIGABRT) received by PID 975226 [4]: time : 2024-07-03_09:55:26 host : ip-26-0-172-73.ec2.internal rank : 62 (local_rank: 6) exitcode : -6 (pid: 975229) error_file: traceback : Signal 6 (SIGABRT) received by PID 975229 ------------------------------------------------------------ Root Cause (first observed failure): [0]: time : 2024-07-03_09:55:26 host : ip-26-0-172-73.ec2.internal rank : 56 (local_rank: 0) exitcode : -6 (pid: 975223) error_file: traceback : Signal 6 (SIGABRT) received by PID 975223 ============================================================ srun: error: ip-26-0-163-226: task 2: Exited with exit code 1 E0703 09:55:28.426000 140322518665024 torch/distributed/elastic/multiprocessing/api.py:826] failed (exitcode: -6) local_rank: 1 (pid: 1133781) of binary: /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/python3.10 W0703 09:55:28.439000 140322518665024 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-172-57.ec2.internal_1133710_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:28.468000 140322518665024 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-172-57.ec2.internal_1133710_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:28.477000 140322518665024 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-172-57.ec2.internal_1133710_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. Traceback (most recent call last): File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/torchrun", line 8, in sys.exit(main()) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper return f(*args, **kwargs) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 879, in main run(args) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 870, in run elastic_launch( File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 132, in __call__ return launch_agent(self._config, self._entrypoint, list(args)) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 263, in launch_agent raise ChildFailedError( torch.distributed.elastic.multiprocessing.errors.ChildFailedError: ============================================================ /fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py FAILED ------------------------------------------------------------ Failures: ------------------------------------------------------------ Root Cause (first observed failure): [0]: time : 2024-07-03_09:55:21 host : ip-26-0-172-57.ec2.internal rank : 49 (local_rank: 1) exitcode : -6 (pid: 1133781) error_file: traceback : Signal 6 (SIGABRT) received by PID 1133781 ============================================================ srun: error: ip-26-0-172-73: task 7: Exited with exit code 1 E0703 09:55:29.132000 140499797497664 torch/distributed/elastic/multiprocessing/api.py:826] failed (exitcode: -6) local_rank: 1 (pid: 2476242) of binary: /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/python3.10 W0703 09:55:29.148000 140499797497664 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-169-132.ec2.internal_2476171_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:29.176000 140499797497664 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-169-132.ec2.internal_2476171_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:29.184000 140499797497664 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-169-132.ec2.internal_2476171_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. Traceback (most recent call last): File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/torchrun", line 8, in sys.exit(main()) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper return f(*args, **kwargs) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 879, in main run(args) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 870, in run elastic_launch( File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 132, in __call__ return launch_agent(self._config, self._entrypoint, list(args)) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 263, in launch_agent raise ChildFailedError( torch.distributed.elastic.multiprocessing.errors.ChildFailedError: ============================================================ /fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py FAILED ------------------------------------------------------------ Failures: ------------------------------------------------------------ Root Cause (first observed failure): [0]: time : 2024-07-03_09:55:26 host : ip-26-0-169-132.ec2.internal rank : 33 (local_rank: 1) exitcode : -6 (pid: 2476242) error_file: traceback : Signal 6 (SIGABRT) received by PID 2476242 ============================================================ srun: error: ip-26-0-172-57: task 6: Exited with exit code 1 E0703 09:55:29.558000 140023290472256 torch/distributed/elastic/multiprocessing/api.py:826] failed (exitcode: -6) local_rank: 1 (pid: 843273) of binary: /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/python3.10 W0703 09:55:29.570000 140023290472256 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-163-220.ec2.internal_843202_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:29.600000 140023290472256 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-163-220.ec2.internal_843202_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:29.608000 140023290472256 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-163-220.ec2.internal_843202_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. Traceback (most recent call last): File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/torchrun", line 8, in sys.exit(main()) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper return f(*args, **kwargs) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 879, in main run(args) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 870, in run elastic_launch( File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 132, in __call__ return launch_agent(self._config, self._entrypoint, list(args)) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 263, in launch_agent raise ChildFailedError( torch.distributed.elastic.multiprocessing.errors.ChildFailedError: ============================================================ /fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py FAILED ------------------------------------------------------------ Failures: ------------------------------------------------------------ Root Cause (first observed failure): [0]: time : 2024-07-03_09:55:26 host : ip-26-0-163-220.ec2.internal rank : 9 (local_rank: 1) exitcode : -6 (pid: 843273) error_file: traceback : Signal 6 (SIGABRT) received by PID 843273 ============================================================ srun: error: ip-26-0-169-132: task 5: Exited with exit code 1 W0703 09:55:30.335000 140554625292096 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-168-238.ec2.internal_1932497_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:30.355000 140554625292096 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-168-238.ec2.internal_1932497_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. Traceback (most recent call last): File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/rendezvous/c10d_rendezvous_backend.py", line 113, in _call_store return getattr(self._store, store_op)(*args, **kwargs) torch.distributed.DistNetworkError: Broken pipe The above exception was the direct cause of the following exception: Traceback (most recent call last): File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/torchrun", line 8, in sys.exit(main()) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper return f(*args, **kwargs) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 879, in main run(args) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 870, in run elastic_launch( File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 132, in __call__ return launch_agent(self._config, self._entrypoint, list(args)) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 254, in launch_agent result = agent.run() File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/metrics/api.py", line 123, in wrapper result = f(*args, **kwargs) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/agent/server/api.py", line 733, in run result = self._invoke_run(role) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/agent/server/api.py", line 908, in _invoke_run num_nodes_waiting = rdzv_handler.num_nodes_waiting() File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/rendezvous/dynamic_rendezvous.py", line 1174, in num_nodes_waiting self._state_holder.sync() File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/rendezvous/dynamic_rendezvous.py", line 419, in sync get_response = self._backend.get_state() File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/rendezvous/c10d_rendezvous_backend.py", line 73, in get_state base64_state: bytes = self._call_store("get", self._key) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/rendezvous/c10d_rendezvous_backend.py", line 115, in _call_store raise RendezvousConnectionError( torch.distributed.elastic.rendezvous.api.RendezvousConnectionError: The connection to the C10d store has failed. See inner exception for details. W0703 09:55:30.529000 139730091611904 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1252] The node 'ip-26-0-169-86.ec2.internal_1905143_0' has failed to send a keep-alive heartbeat to the rendezvous 'none' due to an error of type RendezvousConnectionError. srun: error: ip-26-0-163-220: task 1: Exited with exit code 1 W0703 09:55:31.070000 139735752345408 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-169-86.ec2.internal_1905143_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. W0703 09:55:31.081000 139735752345408 torch/distributed/elastic/rendezvous/dynamic_rendezvous.py:1203] The node 'ip-26-0-169-86.ec2.internal_1905143_0' has failed to shutdown the rendezvous 'none' due to an error of type RendezvousConnectionError. Traceback (most recent call last): File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/rendezvous/c10d_rendezvous_backend.py", line 113, in _call_store return getattr(self._store, store_op)(*args, **kwargs) torch.distributed.DistNetworkError: Broken pipe The above exception was the direct cause of the following exception: Traceback (most recent call last): File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/torchrun", line 8, in sys.exit(main()) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper return f(*args, **kwargs) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 879, in main run(args) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 870, in run elastic_launch( File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 132, in __call__ return launch_agent(self._config, self._entrypoint, list(args)) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 254, in launch_agent result = agent.run() File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/metrics/api.py", line 123, in wrapper result = f(*args, **kwargs) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/agent/server/api.py", line 733, in run result = self._invoke_run(role) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/agent/server/api.py", line 908, in _invoke_run num_nodes_waiting = rdzv_handler.num_nodes_waiting() File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/rendezvous/dynamic_rendezvous.py", line 1174, in num_nodes_waiting self._state_holder.sync() File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/rendezvous/dynamic_rendezvous.py", line 419, in sync get_response = self._backend.get_state() File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/rendezvous/c10d_rendezvous_backend.py", line 73, in get_state base64_state: bytes = self._call_store("get", self._key) File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/rendezvous/c10d_rendezvous_backend.py", line 115, in _call_store raise RendezvousConnectionError( torch.distributed.elastic.rendezvous.api.RendezvousConnectionError: The connection to the C10d store has failed. See inner exception for details. srun: error: ip-26-0-168-238: task 3: Exited with exit code 1 srun: error: ip-26-0-169-86: task 4: Exited with exit code 1 Consider using `hf_transfer` for faster uploads. This solution comes with some limitations. See https://huggingface.co/docs/huggingface_hub/hf_transfer for more details.