======================== START TIME: Wed Jul 3 23:09:08 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 23:09:16.686000 139898740016960 torch/distributed/run.py:757] W0703 23:09:16.686000 139898740016960 torch/distributed/run.py:757] ***************************************** W0703 23:09:16.686000 139898740016960 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 23:09:16.686000 139898740016960 torch/distributed/run.py:757] ***************************************** [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Config: [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Config(general=GeneralArgs(project='bench_cluster', [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: run='%date_%jobid', [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: seed=42, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: step=None, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: consumed_train_samples=None, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: benchmark_csv_path=None, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: ignore_sanity_checks=True), [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: parallelism=ParallelismArgs(dp=2, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: pp=4, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: tp=1, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: pp_engine=, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: tp_mode=, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: tp_linear_async_communication=False, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: expert_parallel_size=1), [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: model=ModelArgs(model_config=LlamaConfig(bos_token_id=1, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: eos_token_id=2, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: hidden_act='silu', [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: hidden_size=2048, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: initializer_range=0.02, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: intermediate_size=4096, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: is_llama_config=True, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: max_position_embeddings=4096, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: num_attention_heads=32, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: num_hidden_layers=24, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: num_key_value_heads=32, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: pad_token_id=None, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: pretraining_tp=1, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: rms_norm_eps=1e-05, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: rope_scaling=None, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: rope_theta=10000.0, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: tie_word_embeddings=True, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: use_cache=True, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: vocab_size=50257), [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: init_method=RandomInit(std=0.025), [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: dtype=torch.bfloat16, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: make_vocab_size_divisible_by=1, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: ddp_bucket_cap_mb=25), [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: tokenizer=TokenizerArgs(tokenizer_name_or_path='openai-community/gpt2', [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: tokenizer_revision=None, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: tokenizer_max_length=None), [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: checkpoints=CheckpointsArgs(checkpoints_path=Path('/dev/null'), [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: checkpoint_interval=100000, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: save_initial_state=False, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: resume_checkpoint_path=None, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: checkpoints_path_is_shared_file_system=False), [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: logging=LoggingArgs(log_level='info', [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: log_level_replica='info', [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: iteration_step_info_interval=1), [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: tokens=TokensArgs(sequence_length=4096, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: train_steps=20, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: micro_batch_size=1, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: batch_accumulation_per_replica=512, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: val_check_interval=-1, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: limit_val_batches=0, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: limit_test_batches=0), [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: optimizer=OptimizerArgs(optimizer_factory=AdamWOptimizerArgs(adam_eps=1e-08, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: adam_beta1=0.9, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: adam_beta2=0.95, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: torch_adam_is_fused=True, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: name='adamW'), [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: zero_stage=1, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: weight_decay=0.01, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: clip_grad=1.0, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: accumulate_grad_in_fp32=True, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: learning_rate_scheduler=LRSchedulerArgs(learning_rate=0.0001, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: lr_warmup_steps=1, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: lr_warmup_style='linear', [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: lr_decay_style='linear', [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: lr_decay_steps=19, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: lr_decay_starting_step=None, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: min_decay_lr=1e-05)), [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: data_stages=[DatasetStageArgs(name='Training Stage', [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: start_training_step=1, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: data=DataArgs(dataset=PretrainDatasetsArgs(hf_dataset_or_datasets='roneneldan/TinyStories', [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: hf_dataset_splits='train', [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: hf_dataset_config_name=None, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: dataset_processing_num_proc_per_process=64, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: dataset_overwrite_cache=False, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: text_column_name='text'), [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: seed=42, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: num_loading_workers=0))], [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: profiler=ProfilerArgs(profiler_export_path=Path('/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-1')), [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: lighteval=None) [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Model Config: [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: LlamaConfig(bos_token_id=1, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: eos_token_id=2, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: hidden_act='silu', [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: hidden_size=2048, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: initializer_range=0.02, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: intermediate_size=4096, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: is_llama_config=True, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: max_position_embeddings=4096, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: num_attention_heads=32, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: num_hidden_layers=24, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: num_key_value_heads=32, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: pad_token_id=None, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: pretraining_tp=1, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: rms_norm_eps=1e-05, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: rope_scaling=None, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: rope_theta=10000.0, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: tie_word_embeddings=True, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: use_cache=True, [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: vocab_size=50257) [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Building model.. [default0]:07/03/2024 23:09:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Setting PP block ranks... [default2]:07/03/2024 23:09:51 [INFO|DP=0|PP=1|TP=0|ip-26-0-174-36]: Local number of parameters: 294M (560.05MiB) [default0]:07/03/2024 23:09:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Total number of parameters: 1.21G (2312.82MiB) [default0]:07/03/2024 23:09:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Local number of parameters: 397M (756.37MiB) [default4]:07/03/2024 23:09:51 [INFO|DP=0|PP=2|TP=0|ip-26-0-174-36]: Local number of parameters: 252M (480.05MiB) [default4]:07/03/2024 23:09:51 [INFO|DP=0|PP=2|TP=0|ip-26-0-174-36]: [After model building] Memory usage: 486.06MiB. Peak allocated: 488.09MiB Peak reserved: 502.00MiB [default4]:07/03/2024 23:09:51 [INFO|DP=0|PP=2|TP=0|ip-26-0-174-36]: No checkpoint path provided. [default6]:07/03/2024 23:09:51 [INFO|DP=0|PP=3|TP=0|ip-26-0-174-36]: Local number of parameters: 271M (516.35MiB) [default6]:07/03/2024 23:09:51 [INFO|DP=0|PP=3|TP=0|ip-26-0-174-36]: [After model building] Memory usage: 520.36MiB. Peak allocated: 522.39MiB Peak reserved: 534.00MiB [default6]:07/03/2024 23:09:51 [INFO|DP=0|PP=3|TP=0|ip-26-0-174-36]: No checkpoint path provided. [default1]:07/03/2024 23:09:51 [INFO|DP=1|PP=0|TP=0|ip-26-0-174-36]: No checkpoint path provided. [default2]:07/03/2024 23:09:51 [INFO|DP=0|PP=1|TP=0|ip-26-0-174-36]: [After model building] Memory usage: 567.07MiB. Peak allocated: 569.10MiB Peak reserved: 594.00MiB [default2]:07/03/2024 23:09:51 [INFO|DP=0|PP=1|TP=0|ip-26-0-174-36]: No checkpoint path provided. [default0]:07/03/2024 23:09:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: [After model building] Memory usage: 763.38MiB. Peak allocated: 765.41MiB Peak reserved: 792.00MiB [default0]:07/03/2024 23:09:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: No checkpoint path provided. [default0]:07/03/2024 23:09:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Parametrizing model parameters using StandardParametrizator [default5]:07/03/2024 23:09:51 [INFO|DP=1|PP=2|TP=0|ip-26-0-174-36]: No checkpoint path provided. [default3]:07/03/2024 23:09:51 [INFO|DP=1|PP=1|TP=0|ip-26-0-174-36]: No checkpoint path provided. [default7]:07/03/2024 23:09:51 [INFO|DP=1|PP=3|TP=0|ip-26-0-174-36]: No checkpoint path provided. [default0]:07/03/2024 23:09:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: [Optimizer Building] Using LearningRateForSP as learning rate [default0]:07/03/2024 23:09:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: [ZeRO sharding] Size of optimizer params per rank: [default0]:07/03/2024 23:09:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: [ZeRO sharding] DP Rank 0 has 198M out of 397M (50.00%) params' optimizer states [default0]:07/03/2024 23:09:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: [ZeRO sharding] DP Rank 1 has 198M out of 397M (50.00%) params' optimizer states [default0]:07/03/2024 23:09:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: [Training Plan] Stage Training Stage has 19 remaining training steps and has consumed 0 samples [default0]:07/03/2024 23:09:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Using `datasets` library [default0]:07/03/2024 23:09:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Loading tokenizer from openai-community/gpt2 and transformers/hf_hub versions ('4.41.2', '0.23.4') [default0]:07/03/2024 23:09:55 [WARNING|DP=0|PP=0|TP=0|ip-26-0-174-36]: 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 23:09:57 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: [Training Plan] There are 1 training stages [default0]:07/03/2024 23:09:57 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: [Stage Training Stage] start from step 1 [default0]:07/03/2024 23:09:57 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: [default0]:07/03/2024 23:09:57 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: [Start training] datetime: 2024-07-03 23:09:57.268378 | mbs: 1 | grad_accum: 512 | global_batch_size: 1024 | sequence_length: 4096 | train_steps: 20 | start_iteration_step: 0 | consumed_train_samples: 0 [default0]:07/03/2024 23:09:57 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Resuming training from stage Training Stage, it has trained for 0 samples and has 19 remaining train steps [default0]:07/03/2024 23:09:57 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Memory usage: 3032.50MiB. Peak allocated 3032.50MiB. Peak reserved: 3064.00MiB [default1]:07/03/2024 23:09:57 [WARNING|DP=1|PP=0|TP=0|ip-26-0-174-36]: Repo card metadata block was not found. Setting CardData to empty. [default2]:07/03/2024 23:09:57 [WARNING|DP=0|PP=1|TP=0|ip-26-0-174-36]: Repo card metadata block was not found. Setting CardData to empty. [default5]:07/03/2024 23:09:57 [WARNING|DP=1|PP=2|TP=0|ip-26-0-174-36]: Repo card metadata block was not found. Setting CardData to empty. [default5]:Repo card metadata block was not found. Setting CardData to empty. [default4]:07/03/2024 23:09:57 [WARNING|DP=0|PP=2|TP=0|ip-26-0-174-36]: 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 23:09:57 [WARNING|DP=0|PP=3|TP=0|ip-26-0-174-36]: Repo card metadata block was not found. Setting CardData to empty. [default1]:Repo card metadata block was not found. Setting CardData to empty. [default3]:07/03/2024 23:09:57 [WARNING|DP=1|PP=1|TP=0|ip-26-0-174-36]: 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. [default2]:Repo card metadata block was not found. Setting CardData to empty. [default7]:Repo card metadata block was not found. Setting CardData to empty. [default7]:07/03/2024 23:09:57 [WARNING|DP=1|PP=3|TP=0|ip-26-0-174-36]: Repo card metadata block was not found. Setting CardData to empty. [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 [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 [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 [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]:/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 [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 [default0]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: Attempting to run cuBLAS, but there was no current CUDA context! Attempting to set the primary context... (Triggered internally at ../aten/src/ATen/cuda/CublasHandlePool.cpp:135.) [default0]: 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/distributed/distributed_c10d.py:2261: UserWarning: torch.distributed.all_reduce_coalesced will be deprecated. If you must use it, please revisit our documentation later at https://pytorch.org/docs/master/distributed.html#collective-functions [default1]: warnings.warn( [default7]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py:2261: UserWarning: torch.distributed.all_reduce_coalesced will be deprecated. If you must use it, please revisit our documentation later at https://pytorch.org/docs/master/distributed.html#collective-functions [default7]: warnings.warn( [default6]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py:2261: UserWarning: torch.distributed.all_reduce_coalesced will be deprecated. If you must use it, please revisit our documentation later at https://pytorch.org/docs/master/distributed.html#collective-functions [default6]: warnings.warn( [default0]:07/03/2024 23:10:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Memory usage: 3100.03MiB. Peak allocated 11376.44MiB. Peak reserved: 11574.00MiB [default0]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py:2261: UserWarning: torch.distributed.all_reduce_coalesced will be deprecated. If you must use it, please revisit our documentation later at https://pytorch.org/docs/master/distributed.html#collective-functions [default0]: warnings.warn( [default0]:07/03/2024 23:10:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Memory usage: 4612.80MiB. Peak allocated 6503.72MiB. Peak reserved: 13670.00MiB [default6]:07/03/2024 23:10:40 [INFO|DP=0|PP=3|TP=0|ip-26-0-174-36]: iteration: 1 / 20 | consumed_tokens: 4.19M | elapsed_time_per_iteration_ms: 41.9K | tokens_per_sec: 100K | tokens_per_sec_per_gpu: 12.5K | global_batch_size: 1.02K | lm_loss: 11.1 | lr: 0.0001 | model_tflops_per_gpu: 114 | hardware_tflops_per_gpu: 114 | grad_norm: 25.1 | cuda_memory_allocated: 3.32G | cuda_max_memory_reserved: 6.37G | hd_total_memory_tb: 312G | hd_used_memory_tb: 65.8G | hd_free_memory_tb: 246G [default0]:07/03/2024 23:11:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Memory usage: 4612.80MiB. Peak allocated 12787.81MiB. Peak reserved: 13670.00MiB [default0]:07/03/2024 23:11:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Memory usage: 4612.80MiB. Peak allocated 6503.72MiB. Peak reserved: 13670.00MiB [default6]:07/03/2024 23:11:05 [INFO|DP=0|PP=3|TP=0|ip-26-0-174-36]: iteration: 2 / 20 | consumed_tokens: 8.39M | elapsed_time_per_iteration_ms: 25.3K | tokens_per_sec: 166K | tokens_per_sec_per_gpu: 20.7K | global_batch_size: 1.02K | lm_loss: 11.1 | lr: 9.53e-05 | model_tflops_per_gpu: 188 | hardware_tflops_per_gpu: 188 | grad_norm: 25.2 | cuda_memory_allocated: 3.32G | cuda_max_memory_reserved: 6.37G | hd_total_memory_tb: 312G | hd_used_memory_tb: 65.8G | hd_free_memory_tb: 246G [default0]:07/03/2024 23:11:30 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Memory usage: 4612.80MiB. Peak allocated 12787.81MiB. Peak reserved: 13670.00MiB [default0]:07/03/2024 23:11:30 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Memory usage: 4612.80MiB. Peak allocated 6503.72MiB. Peak reserved: 13670.00MiB [default0]:STAGE:2024-07-03 23:11:30 243244:243244 ActivityProfilerController.cpp:314] Completed Stage: Warm Up [default6]:07/03/2024 23:11:30 [INFO|DP=0|PP=3|TP=0|ip-26-0-174-36]: iteration: 3 / 20 | consumed_tokens: 12.6M | elapsed_time_per_iteration_ms: 25.2K | tokens_per_sec: 167K | tokens_per_sec_per_gpu: 20.8K | global_batch_size: 1.02K | lm_loss: 11.4 | lr: 9.05e-05 | model_tflops_per_gpu: 189 | hardware_tflops_per_gpu: 189 | grad_norm: 217 | cuda_memory_allocated: 3.32G | cuda_max_memory_reserved: 6.37G | hd_total_memory_tb: 312G | hd_used_memory_tb: 65.8G | hd_free_memory_tb: 246G [default0]:07/03/2024 23:11:58 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Memory usage: 4612.80MiB. Peak allocated 12787.81MiB. Peak reserved: 13670.00MiB [default0]:07/03/2024 23:11:58 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Memory usage: 4612.80MiB. Peak allocated 6503.72MiB. Peak reserved: 13670.00MiB [default6]:07/03/2024 23:11:58 [INFO|DP=0|PP=3|TP=0|ip-26-0-174-36]: iteration: 4 / 20 | consumed_tokens: 16.8M | elapsed_time_per_iteration_ms: 28.2K | tokens_per_sec: 149K | tokens_per_sec_per_gpu: 18.6K | global_batch_size: 1.02K | lm_loss: 13.8 | lr: 8.58e-05 | model_tflops_per_gpu: 169 | hardware_tflops_per_gpu: 169 | grad_norm: 22.5 | cuda_memory_allocated: 3.32G | cuda_max_memory_reserved: 6.37G | hd_total_memory_tb: 312G | hd_used_memory_tb: 65.8G | hd_free_memory_tb: 246G [default0]:07/03/2024 23:12:27 [INFO|DP=0|PP=0|TP=0|ip-26-0-174-36]: Memory usage: 4612.80MiB. Peak allocated 12787.81MiB. Peak reserved: 13670.00MiB [default6]:07/03/2024 23:12:27 [INFO|DP=0|PP=3|TP=0|ip-26-0-174-36]: iteration: 5 / 20 | consumed_tokens: 21M | elapsed_time_per_iteration_ms: 28.3K | tokens_per_sec: 148K | tokens_per_sec_per_gpu: 18.5K | global_batch_size: 1.02K | lm_loss: 9.98 | lr: 8.11e-05 | model_tflops_per_gpu: 168 | hardware_tflops_per_gpu: 168 | grad_norm: 16.5 [default6]:07/03/2024 23:12:55 [INFO|DP=0|PP=3|TP=0|ip-26-0-174-36]: iteration: 6 / 20 | consumed_tokens: 25.2M | elapsed_time_per_iteration_ms: 28.5K | tokens_per_sec: 147K | tokens_per_sec_per_gpu: 18.4K | global_batch_size: 1.02K | lm_loss: 10.9 | lr: 7.63e-05 | model_tflops_per_gpu: 167 | hardware_tflops_per_gpu: 167 | grad_norm: 93.9 [default0]:STAGE:2024-07-03 23:14:23 243244:243244 ActivityProfilerController.cpp:320] Completed Stage: Collection [default0]:STAGE:2024-07-03 23:14:30 243244:243244 ActivityProfilerController.cpp:324] Completed Stage: Post Processing [default6]:[rank6]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=27657, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600025 milliseconds before timing out. [default2]:[rank2]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=55305, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600091 milliseconds before timing out. [default4]:[rank4]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=55305, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600032 milliseconds before timing out. [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 278, in train_batch_iter [default6]:[rank6]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default6]:[rank6]: output = model(**micro_batch) [default6]:[rank6]: 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]:[rank6]: return self._call_impl(*args, **kwargs) [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default6]:[rank6]: return forward_call(*args, **kwargs) [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default6]:[rank6]: sharded_logits = self.model( [default6]:[rank6]: 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]:[rank6]: return self._call_impl(*args, **kwargs) [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default6]:[rank6]: return forward_call(*args, **kwargs) [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default6]:[rank6]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default6]:[rank6]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default6]:[rank6]: 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]:[rank6]: return self._call_impl(*args, **kwargs) [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default6]:[rank6]: return forward_call(*args, **kwargs) [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default6]:[rank6]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default6]:[rank6]: pipeline_state.run_communication() [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default6]:[rank6]: recv_activation_tensor = recv_activation() [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default6]:[rank6]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default6]:[rank6]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default6]:[rank6]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 267, in _recv_meta [default6]:[rank6]: self.second_metadata = torch.empty(second_metadata_num_elements, dtype=torch.long, device=self.device) [default6]:[rank6]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate more than 1EB memory. [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( [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 [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 252, in train_batch_iter [default4]:[rank4]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default4]:[rank4]: output = model(**micro_batch) [default4]:[rank4]: 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]:[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 252, in train_batch_iter [default2]:[rank2]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model) [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward [default4]:[rank4]: return self._call_impl(*args, **kwargs) [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default4]:[rank4]: return forward_call(*args, **kwargs) [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default4]:[rank4]: sharded_logits = self.model( [default2]:[rank2]: output = model(**micro_batch) [default4]:[rank4]: 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]:[rank4]: return self._call_impl(*args, **kwargs) [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default4]:[rank4]: return forward_call(*args, **kwargs) [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default2]:[rank2]: 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]:[rank2]: return self._call_impl(*args, **kwargs) [default2]:[rank2]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default4]:[rank4]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default2]:[rank2]: return forward_call(*args, **kwargs) [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default4]:[rank4]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default4]:[rank4]: 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]:[rank4]: return self._call_impl(*args, **kwargs) [default2]:[rank2]: sharded_logits = self.model( [default2]:[rank2]: 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]:[rank2]: return self._call_impl(*args, **kwargs) [default2]:[rank2]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default2]:[rank2]: return forward_call(*args, **kwargs) [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default4]:[rank4]: return forward_call(*args, **kwargs) [default2]:[rank2]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0] [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default4]:[rank4]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states [default2]:[rank2]: hidden_encoder_states = encoder_block(**hidden_encoder_states) [default2]:[rank2]: 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]:[rank2]: return self._call_impl(*args, **kwargs) [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default4]:[rank4]: pipeline_state.run_communication() [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default4]:[rank4]: recv_activation_tensor = recv_activation() [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default2]:[rank2]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl [default4]:[rank4]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default4]:[rank4]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default2]:[rank2]: return forward_call(*args, **kwargs) [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward [default2]:[rank2]: new_kwargs[name] = recv_from_pipeline_state_buffer( [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer [default4]:[rank4]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 267, in _recv_meta [default4]:[rank4]: self.second_metadata = torch.empty(second_metadata_num_elements, dtype=torch.long, device=self.device) [default4]:[rank4]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate more than 1EB memory. [default2]:[rank2]: pipeline_state.run_communication() [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication [default2]:[rank2]: recv_activation_tensor = recv_activation() [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__ [default2]:[rank2]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0] [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors [default2]:[rank2]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag) [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors [default2]:[rank2]: meta = self._recv_meta(from_rank=from_rank, tag=tag) [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 267, in _recv_meta [default2]:[rank2]: self.second_metadata = torch.empty(second_metadata_num_elements, dtype=torch.long, device=self.device) [default2]:[rank2]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate more than 1EB memory. [default6]:[rank6]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 27657, last enqueued NCCL work: 27657, last completed NCCL work: 27656. [default6]:[rank6]:[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]:[rank6]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down. [default6]:[rank6]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=27657, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600025 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 (0x7fc01c23b897 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 (0x7fc01d514c62 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 (0x7fc01d519a80 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 (0x7fc01d51adcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7fc068fb3e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7fc06dffa609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7fc06ddc5353 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=27657, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600025 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 (0x7fc01c23b897 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 (0x7fc01d514c62 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 (0x7fc01d519a80 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 (0x7fc01d51adcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #4: + 0xd3e95 (0x7fc068fb3e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #5: + 0x8609 (0x7fc06dffa609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #6: clone + 0x43 (0x7fc06ddc5353 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 (0x7fc01c23b897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default6]:frame #1: + 0xe32119 (0x7fc01d19e119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default6]:frame #2: + 0xd3e95 (0x7fc068fb3e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default6]:frame #3: + 0x8609 (0x7fc06dffa609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default6]:frame #4: clone + 0x43 (0x7fc06ddc5353 in /lib/x86_64-linux-gnu/libc.so.6) [default6]: [default4]:[rank4]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 2] Timeout at NCCL work: 55305, last enqueued NCCL work: 55305, last completed NCCL work: 55304. [default4]:[rank4]:[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]:[rank4]:[E ProcessGroupNCCL.cpp:583] [Rank 2] To avoid data inconsistency, we are taking the entire process down. [default4]:[rank4]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 2] Process group watchdog thread terminated with exception: [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=55305, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600032 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 (0x7f0979431897 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 (0x7f097a70ac62 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 (0x7f097a70fa80 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 (0x7f097a710dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f09c61a9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f09cb1f0609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f09cafbb353 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=55305, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600032 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 (0x7f0979431897 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 (0x7f097a70ac62 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 (0x7f097a70fa80 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 (0x7f097a710dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #4: + 0xd3e95 (0x7f09c61a9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default4]:frame #5: + 0x8609 (0x7f09cb1f0609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #6: clone + 0x43 (0x7f09cafbb353 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 (0x7f0979431897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default4]:frame #1: + 0xe32119 (0x7f097a394119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default4]:frame #2: + 0xd3e95 (0x7f09c61a9e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:[rank2]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 55305, last enqueued NCCL work: 55305, last completed NCCL work: 55304. [default2]:[rank2]:[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]:frame #3: + 0x8609 (0x7f09cb1f0609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default4]:frame #4: clone + 0x43 (0x7f09cafbb353 in /lib/x86_64-linux-gnu/libc.so.6) [default4]: [default2]:[rank2]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down. [default2]:[rank2]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=55305, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600091 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 (0x7fec3ebb1897 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 (0x7fec3fe8ac62 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 (0x7fec3fe8fa80 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 (0x7fec3fe90dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7fec8b929e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #5: + 0x8609 (0x7fec90970609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7fec9073b353 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=55305, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600091 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 (0x7fec3ebb1897 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 (0x7fec3fe8ac62 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 (0x7fec3fe8fa80 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 (0x7fec3fe90dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #4: + 0xd3e95 (0x7fec8b929e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #5: + 0x8609 (0x7fec90970609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #6: clone + 0x43 (0x7fec9073b353 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 (0x7fec3ebb1897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so) [default2]:frame #1: + 0xe32119 (0x7fec3fb14119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so) [default2]:frame #2: + 0xd3e95 (0x7fec8b929e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6) [default2]:frame #3: + 0x8609 (0x7fec90970609 in /lib/x86_64-linux-gnu/libpthread.so.0) [default2]:frame #4: clone + 0x43 (0x7fec9073b353 in /lib/x86_64-linux-gnu/libc.so.6) [default2]: W0703 23:23:02.743000 139898740016960 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 243244 closing signal SIGTERM W0703 23:23:02.743000 139898740016960 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 243245 closing signal SIGTERM W0703 23:23:02.743000 139898740016960 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 243247 closing signal SIGTERM W0703 23:23:02.743000 139898740016960 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 243249 closing signal SIGTERM W0703 23:23:02.744000 139898740016960 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 243251 closing signal SIGTERM E0703 23:23:06.831000 139898740016960 torch/distributed/elastic/multiprocessing/api.py:826] failed (exitcode: -6) local_rank: 2 (pid: 243246) 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: [1]: time : 2024-07-03_23:23:02 host : ip-26-0-174-36.ec2.internal rank : 4 (local_rank: 4) exitcode : -6 (pid: 243248) error_file: traceback : Signal 6 (SIGABRT) received by PID 243248 [2]: time : 2024-07-03_23:23:02 host : ip-26-0-174-36.ec2.internal rank : 6 (local_rank: 6) exitcode : -6 (pid: 243250) error_file: traceback : Signal 6 (SIGABRT) received by PID 243250 ------------------------------------------------------------ Root Cause (first observed failure): [0]: time : 2024-07-03_23:23:02 host : ip-26-0-174-36.ec2.internal rank : 2 (local_rank: 2) exitcode : -6 (pid: 243246) error_file: traceback : Signal 6 (SIGABRT) received by PID 243246 ============================================================ srun: error: ip-26-0-174-36: task 0: 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. ip-26-0-174-36_243244.1720048960776918215.pt.trace.json.tmp: 0%| | 0.00/2.93G [00:00