Hugging Face's logo

nanotron
/
old_bench

Model card Files Community
3outeille HF staff commited on
Commit
c82fb20
·
verified ·
1 Parent(s): dd97276

Upload llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8

Browse files
Files changed (4) hide show
  1. llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/bench.slurm +111 -0
  2. llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/config.yaml +90 -0
  3. llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/log.out +429 -0
  4. llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/status.txt +1 -0
llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/bench.slurm ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+
3
+ #SBATCH --job-name=bench_cluster
4
+ #SBATCH --time=02:00:00
5
+ #SBATCH --partition=hopper-prod
6
+ #SBATCH --nodes=1
7
+ #SBATCH --gres=gpu:8
8
+ #SBATCH --qos=normal
9
+ #SBATCH --ntasks-per-node=1
10
+ #SBATCH --cpus-per-task=96
11
+ #SBATCH --exclusive
12
+ #SBATCH --output=/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/log.out
13
+ #SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/log.out
14
+
15
+ # Function to update status based on squeue output
16
+ update_status() {
17
+ job_id=$1
18
+ status_file=$2
19
+ # For unknown reasons, it doenst update status for pending. It only works for running
20
+ while true; do
21
+ job_status=$(squeue --job $job_id --noheader --format=%T)
22
+ echo "Job status: $job_status"
23
+ if [ -z "$job_status" ]; then
24
+ # Job has finished or is not found
25
+ break
26
+ elif [ "$job_status" = "RUNNING" ]; then
27
+ printf "running" > $status_file
28
+ break
29
+ fi
30
+ sleep 10
31
+ done
32
+ }
33
+
34
+ # Misc initializations.
35
+ echo "========================"
36
+ echo "START TIME: $(date)"
37
+ source /fsx/ferdinandmom/miniforge3/etc/profile.d/conda.sh
38
+ conda activate /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster
39
+ echo python3 version = $(python3 --version)
40
+ echo "========================"
41
+
42
+ # Slurm stuff
43
+ export HOSTNAMES=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
44
+ export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
45
+ export MASTER_PORT=$((1024 + RANDOM % 64511))
46
+
47
+ export TMPDIR=/scratch
48
+ export HF_DATASETS_CACHE="/admin/home/ferdinand_mom/.cache"
49
+ export CUBLAS_WORKSPACE_CONFIG=":4096:8"
50
+ export CUDA_DEVICE_MAX_CONNECTIONS="1"
51
+
52
+ huggingface-cli login --token $HUGGINGFACE_TOKEN
53
+
54
+
55
+ NANOTRON_REPO="/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron"
56
+ CMD="$NANOTRON_REPO/run_train.py --config-file /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/config.yaml"
57
+
58
+ LAUNCHER="torchrun \
59
+ --nproc_per_node 8 \
60
+ --nnodes 1 \
61
+ --rdzv_endpoint ${MASTER_ADDR}:${MASTER_PORT} \
62
+ --rdzv_backend c10d \
63
+ --max_restarts 0 \
64
+ --tee 3 \
65
+ --node_rank ${SLURM_PROCID}"
66
+
67
+ # Checkout the bench_cluster branch
68
+ cd $NANOTRON_REPO
69
+ git checkout bench_cluster
70
+ cd ..
71
+ # Get the current job ID
72
+ job_id=${SLURM_JOB_ID}
73
+
74
+ # Update status to "pending" or "running" in the background
75
+ update_status $job_id /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/status.txt &
76
+
77
+ # Run the main command
78
+ srun -u $LAUNCHER $CMD
79
+ exit_status=$?
80
+
81
+ # Update status based on the exit status of `srun`
82
+ if [ $exit_status -eq 0 ]; then
83
+ printf "completed" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/status.txt
84
+ else
85
+ if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/log.out; then
86
+ printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/status.txt
87
+ elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/log.out; then
88
+ printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/status.txt
89
+ elif grep -q "Timeout at NCCL" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/log.out; then
90
+ printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/status.txt
91
+ else
92
+ printf "fail" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/status.txt
93
+ fi
94
+ fi
95
+
96
+ # Run the report script if the job completed successfully
97
+ if [ $exit_status -eq 0 ]; then
98
+ python /fsx/ferdinandmom/ferdinand-hf/bench_cluster/main.py report --inp_dir /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8 --is_logs
99
+ python /fsx/ferdinandmom/ferdinand-hf/bench_cluster/main.py report --inp_dir /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8 --is_profiler
100
+ fi
101
+
102
+
103
+ # Push to hub the folder using huggingface_cli
104
+ huggingface-cli upload nanotron/bench_cluster /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8 llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8 --commit-message "Upload llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8"
105
+
106
+ # Verify the upload
107
+ if [ $? -eq 0 ]; then
108
+ echo "Uploading to Huggingface Hub successful"
109
+ else
110
+ echo "Failed to upload to Huggingface Hub"
111
+ fi
llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/config.yaml ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ general:
2
+ project: bench_cluster
3
+ seed: 42
4
+ model:
5
+ ddp_bucket_cap_mb: 25
6
+ dtype: bfloat16
7
+ init_method:
8
+ std: 0.025
9
+ make_vocab_size_divisible_by: 1
10
+ model_config:
11
+ bos_token_id: 1
12
+ eos_token_id: 2
13
+ hidden_act: silu
14
+ hidden_size: 2048
15
+ initializer_range: 0.02
16
+ intermediate_size: 4096
17
+ is_llama_config: true
18
+ max_position_embeddings: 4096
19
+ num_attention_heads: 32
20
+ num_hidden_layers: 24
21
+ num_key_value_heads: 32
22
+ pad_token_id: null
23
+ pretraining_tp: 1
24
+ rms_norm_eps: 1.0e-05
25
+ rope_scaling: null
26
+ rope_theta: 10000.0
27
+ tie_word_embeddings: true
28
+ use_cache: true
29
+ vocab_size: 50257
30
+ optimizer:
31
+ accumulate_grad_in_fp32: true
32
+ clip_grad: 1.0
33
+ learning_rate_scheduler:
34
+ learning_rate: 0.0001
35
+ lr_decay_style: linear
36
+ lr_warmup_style: linear
37
+ lr_warmup_steps: 1
38
+ min_decay_lr: 1.0e-05
39
+ optimizer_factory:
40
+ adam_beta1: 0.9
41
+ adam_beta2: 0.95
42
+ adam_eps: 1.0e-08
43
+ name: adamW
44
+ torch_adam_is_fused: true
45
+ weight_decay: 0.01
46
+ zero_stage: 1
47
+ parallelism:
48
+ dp: 4
49
+ expert_parallel_size: 1
50
+ pp: 2
51
+ pp_engine: 1f1b
52
+ tp: 1
53
+ tp_linear_async_communication: false
54
+ tp_mode: REDUCE_SCATTER
55
+ profiler:
56
+ profiler_export_path: /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8
57
+ tokenizer:
58
+ tokenizer_max_length: null
59
+ tokenizer_name_or_path: openai-community/gpt2
60
+ tokenizer_revision: null
61
+ data_stages:
62
+ - name: Training Stage
63
+ start_training_step: 1
64
+ data:
65
+ dataset:
66
+ dataset_overwrite_cache: false
67
+ dataset_processing_num_proc_per_process: 64
68
+ hf_dataset_config_name: null
69
+ hf_dataset_or_datasets: roneneldan/TinyStories
70
+ hf_dataset_splits: train
71
+ text_column_name: text
72
+ num_loading_workers: 0
73
+ seed: 42
74
+ lighteval: null
75
+ tokens:
76
+ train_steps: 20
77
+ val_check_interval: -1
78
+ batch_accumulation_per_replica: 32
79
+ limit_test_batches: 0
80
+ limit_val_batches: 0
81
+ micro_batch_size: 8
82
+ sequence_length: 4096
83
+ logging:
84
+ iteration_step_info_interval: 1
85
+ log_level: info
86
+ log_level_replica: info
87
+ checkpoints:
88
+ checkpoint_interval: 100000
89
+ checkpoints_path: /dev/null
90
+ resume_checkpoint_path: null
llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/log.out ADDED
@@ -0,0 +1,429 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ========================
2
+ START TIME: Thu Jul 4 02:25:23 UTC 2024
3
+ python3 version = Python 3.10.14
4
+ ========================
5
+ 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.
6
+ Token is valid (permission: write).
7
+ Your token has been saved to /admin/home/ferdinand_mom/.cache/huggingface/token
8
+ Login successful
9
+ Already on 'bench_cluster'
10
+ M examples/config_tiny_llama.py
11
+ M examples/config_tiny_llama.yaml
12
+ M examples/train_tiny_llama.sh
13
+ M src/nanotron/models/llama.py
14
+ M src/nanotron/trainer.py
15
+ Your branch is up to date with 'origin/bench_cluster'.
16
+ Job status: RUNNING
17
+ W0704 02:25:33.312000 140445158446912 torch/distributed/run.py:757]
18
+ W0704 02:25:33.312000 140445158446912 torch/distributed/run.py:757] *****************************************
19
+ W0704 02:25:33.312000 140445158446912 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.
20
+ W0704 02:25:33.312000 140445158446912 torch/distributed/run.py:757] *****************************************
21
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: Config:
22
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: Config(general=GeneralArgs(project='bench_cluster',
23
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: run='%date_%jobid',
24
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: seed=42,
25
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: step=None,
26
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: consumed_train_samples=None,
27
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: benchmark_csv_path=None,
28
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: ignore_sanity_checks=True),
29
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: parallelism=ParallelismArgs(dp=4,
30
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: pp=2,
31
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: tp=1,
32
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: pp_engine=<nanotron.parallel.pipeline_parallel.engine.OneForwardOneBackwardPipelineEngine object at 0x7f89861d0880>,
33
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: tp_mode=<TensorParallelLinearMode.REDUCE_SCATTER: 2>,
34
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: tp_linear_async_communication=False,
35
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: expert_parallel_size=1),
36
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: model=ModelArgs(model_config=LlamaConfig(bos_token_id=1,
37
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: eos_token_id=2,
38
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: hidden_act='silu',
39
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: hidden_size=2048,
40
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: initializer_range=0.02,
41
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: intermediate_size=4096,
42
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: is_llama_config=True,
43
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: max_position_embeddings=4096,
44
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: num_attention_heads=32,
45
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: num_hidden_layers=24,
46
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: num_key_value_heads=32,
47
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: pad_token_id=None,
48
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: pretraining_tp=1,
49
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: rms_norm_eps=1e-05,
50
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: rope_scaling=None,
51
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: rope_theta=10000.0,
52
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: tie_word_embeddings=True,
53
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: use_cache=True,
54
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: vocab_size=50257),
55
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: init_method=RandomInit(std=0.025),
56
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: dtype=torch.bfloat16,
57
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: make_vocab_size_divisible_by=1,
58
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: ddp_bucket_cap_mb=25),
59
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: tokenizer=TokenizerArgs(tokenizer_name_or_path='openai-community/gpt2',
60
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: tokenizer_revision=None,
61
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: tokenizer_max_length=None),
62
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: checkpoints=CheckpointsArgs(checkpoints_path=Path('/dev/null'),
63
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: checkpoint_interval=100000,
64
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: save_initial_state=False,
65
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: resume_checkpoint_path=None,
66
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: checkpoints_path_is_shared_file_system=False),
67
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: logging=LoggingArgs(log_level='info',
68
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: log_level_replica='info',
69
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: iteration_step_info_interval=1),
70
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: tokens=TokensArgs(sequence_length=4096,
71
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: train_steps=20,
72
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: micro_batch_size=8,
73
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: batch_accumulation_per_replica=32,
74
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: val_check_interval=-1,
75
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: limit_val_batches=0,
76
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: limit_test_batches=0),
77
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: optimizer=OptimizerArgs(optimizer_factory=AdamWOptimizerArgs(adam_eps=1e-08,
78
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: adam_beta1=0.9,
79
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: adam_beta2=0.95,
80
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: torch_adam_is_fused=True,
81
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: name='adamW'),
82
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: zero_stage=1,
83
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: weight_decay=0.01,
84
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: clip_grad=1.0,
85
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: accumulate_grad_in_fp32=True,
86
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: learning_rate_scheduler=LRSchedulerArgs(learning_rate=0.0001,
87
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: lr_warmup_steps=1,
88
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: lr_warmup_style='linear',
89
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: lr_decay_style='linear',
90
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: lr_decay_steps=19,
91
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: lr_decay_starting_step=None,
92
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: min_decay_lr=1e-05)),
93
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: data_stages=[DatasetStageArgs(name='Training Stage',
94
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: start_training_step=1,
95
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: data=DataArgs(dataset=PretrainDatasetsArgs(hf_dataset_or_datasets='roneneldan/TinyStories',
96
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: hf_dataset_splits='train',
97
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: hf_dataset_config_name=None,
98
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: dataset_processing_num_proc_per_process=64,
99
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: dataset_overwrite_cache=False,
100
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: text_column_name='text'),
101
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: seed=42,
102
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: num_loading_workers=0))],
103
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: profiler=ProfilerArgs(profiler_export_path=Path('/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8')),
104
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: lighteval=None)
105
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: Model Config:
106
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: LlamaConfig(bos_token_id=1,
107
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: eos_token_id=2,
108
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: hidden_act='silu',
109
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: hidden_size=2048,
110
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: initializer_range=0.02,
111
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: intermediate_size=4096,
112
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: is_llama_config=True,
113
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: max_position_embeddings=4096,
114
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: num_attention_heads=32,
115
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: num_hidden_layers=24,
116
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: num_key_value_heads=32,
117
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: pad_token_id=None,
118
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: pretraining_tp=1,
119
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: rms_norm_eps=1e-05,
120
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: rope_scaling=None,
121
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: rope_theta=10000.0,
122
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: tie_word_embeddings=True,
123
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: use_cache=True,
124
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: vocab_size=50257)
125
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: Building model..
126
+ [default0]:07/04/2024 02:25:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: Setting PP block ranks...
127
+ [default0]:07/04/2024 02:26:06 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: Total number of parameters: 1.21G (2312.82MiB)
128
+ [default0]:07/04/2024 02:26:06 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: Local number of parameters: 690M (1316.43MiB)
129
+ [default0]:07/04/2024 02:26:06 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: [After model building] Memory usage: 1330.44MiB. Peak allocated: 1332.47MiB Peak reserved: 1364.00MiB
130
+ [default0]:07/04/2024 02:26:06 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: No checkpoint path provided.
131
+ [default0]:07/04/2024 02:26:06 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: Parametrizing model parameters using StandardParametrizator
132
+ [default4]:07/04/2024 02:26:06 [INFO|DP=0|PP=1|TP=0|ip-26-0-171-88]: Local number of parameters: 522M (996.40MiB)
133
+ [default4]:07/04/2024 02:26:06 [INFO|DP=0|PP=1|TP=0|ip-26-0-171-88]: [After model building] Memory usage: 1006.41MiB. Peak allocated: 1008.44MiB Peak reserved: 1032.00MiB
134
+ [default4]:07/04/2024 02:26:06 [INFO|DP=0|PP=1|TP=0|ip-26-0-171-88]: No checkpoint path provided.
135
+ [default1]:07/04/2024 02:26:06 [INFO|DP=1|PP=0|TP=0|ip-26-0-171-88]: No checkpoint path provided.
136
+ [default2]:07/04/2024 02:26:06 [INFO|DP=2|PP=0|TP=0|ip-26-0-171-88]: No checkpoint path provided.
137
+ [default5]:07/04/2024 02:26:06 [INFO|DP=1|PP=1|TP=0|ip-26-0-171-88]: No checkpoint path provided.
138
+ [default6]:07/04/2024 02:26:06 [INFO|DP=2|PP=1|TP=0|ip-26-0-171-88]: No checkpoint path provided.
139
+ [default3]:07/04/2024 02:26:06 [INFO|DP=3|PP=0|TP=0|ip-26-0-171-88]: No checkpoint path provided.
140
+ [default7]:07/04/2024 02:26:06 [INFO|DP=3|PP=1|TP=0|ip-26-0-171-88]: No checkpoint path provided.
141
+ [default0]:07/04/2024 02:26:10 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: [Optimizer Building] Using LearningRateForSP as learning rate
142
+ [default0]:07/04/2024 02:26:10 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: [ZeRO sharding] Size of optimizer params per rank:
143
+ [default0]:07/04/2024 02:26:10 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: [ZeRO sharding] DP Rank 0 has 173M out of 690M (25.00%) params' optimizer states
144
+ [default0]:07/04/2024 02:26:10 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: [ZeRO sharding] DP Rank 1 has 173M out of 690M (25.00%) params' optimizer states
145
+ [default0]:07/04/2024 02:26:10 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: [ZeRO sharding] DP Rank 2 has 173M out of 690M (25.00%) params' optimizer states
146
+ [default0]:07/04/2024 02:26:10 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: [ZeRO sharding] DP Rank 3 has 173M out of 690M (25.00%) params' optimizer states
147
+ [default0]:07/04/2024 02:26:12 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: [Training Plan] Stage Training Stage has 19 remaining training steps and has consumed 0 samples
148
+ [default0]:07/04/2024 02:26:12 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: Using `datasets` library
149
+ [default0]:07/04/2024 02:26:12 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: Loading tokenizer from openai-community/gpt2 and transformers/hf_hub versions ('4.41.2', '0.23.4')
150
+ [default0]:07/04/2024 02:26:12 [WARNING|DP=0|PP=0|TP=0|ip-26-0-171-88]: Repo card metadata block was not found. Setting CardData to empty.
151
+ [default0]:Repo card metadata block was not found. Setting CardData to empty.
152
+ [default0]:07/04/2024 02:26:14 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: [Training Plan] There are 1 training stages
153
+ [default0]:07/04/2024 02:26:14 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: [Stage Training Stage] start from step 1
154
+ [default0]:07/04/2024 02:26:14 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]:
155
+ [default0]:07/04/2024 02:26:14 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: [Start training] datetime: 2024-07-04 02:26:14.736561 | mbs: 8 | grad_accum: 32 | global_batch_size: 1024 | sequence_length: 4096 | train_steps: 20 | start_iteration_step: 0 | consumed_train_samples: 0
156
+ [default0]:07/04/2024 02:26:14 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: Resuming training from stage Training Stage, it has trained for 0 samples and has 19 remaining train steps
157
+ [default0]:07/04/2024 02:26:14 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-88]: Memory usage: 4621.51MiB. Peak allocated 4621.51MiB. Peak reserved: 4658.00MiB
158
+ [default1]:07/04/2024 02:26:14 [WARNING|DP=1|PP=0|TP=0|ip-26-0-171-88]: Repo card metadata block was not found. Setting CardData to empty.
159
+ [default4]:07/04/2024 02:26:14 [WARNING|DP=0|PP=1|TP=0|ip-26-0-171-88]: Repo card metadata block was not found. Setting CardData to empty.
160
+ [default5]:07/04/2024 02:26:14 [WARNING|DP=1|PP=1|TP=0|ip-26-0-171-88]: Repo card metadata block was not found. Setting CardData to empty.
161
+ [default1]:Repo card metadata block was not found. Setting CardData to empty.
162
+ [default4]:Repo card metadata block was not found. Setting CardData to empty.
163
+ [default7]:07/04/2024 02:26:14 [WARNING|DP=3|PP=1|TP=0|ip-26-0-171-88]: Repo card metadata block was not found. Setting CardData to empty.
164
+ [default5]:Repo card metadata block was not found. Setting CardData to empty.
165
+ [default7]:Repo card metadata block was not found. Setting CardData to empty.
166
+ [default2]:07/04/2024 02:26:15 [WARNING|DP=2|PP=0|TP=0|ip-26-0-171-88]: Repo card metadata block was not found. Setting CardData to empty.
167
+ [default3]:07/04/2024 02:26:14 [WARNING|DP=3|PP=0|TP=0|ip-26-0-171-88]: Repo card metadata block was not found. Setting CardData to empty.
168
+ [default2]:Repo card metadata block was not found. Setting CardData to empty.
169
+ [default3]:Repo card metadata block was not found. Setting CardData to empty.
170
+ [default6]:07/04/2024 02:26:15 [WARNING|DP=2|PP=1|TP=0|ip-26-0-171-88]: Repo card metadata block was not found. Setting CardData to empty.
171
+ [default6]:Repo card metadata block was not found. Setting CardData to empty.
172
+ [default0]:[rank0]: Traceback (most recent call last):
173
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
174
+ [default0]:[rank0]: trainer.train(dataloader)
175
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
176
+ [default0]:[rank0]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
177
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
178
+ [default0]:[rank0]: outputs = self.pipeline_engine.train_batch_iter(
179
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
180
+ [default0]:[rank0]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
181
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
182
+ [default0]:[rank0]: output = model(**micro_batch)
183
+ [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
184
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
185
+ [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
186
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
187
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
188
+ [default0]:[rank0]: sharded_logits = self.model(
189
+ [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
190
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
191
+ [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
192
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
193
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
194
+ [default0]:[rank0]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
195
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
196
+ [default0]:[rank0]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
197
+ [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
198
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
199
+ [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
200
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
201
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 151, in forward
202
+ [default0]:[rank0]: output = self.pp_block(**new_kwargs)
203
+ [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
204
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
205
+ [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
206
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
207
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 631, in forward
208
+ [default0]:[rank0]: output = self.attn(hidden_states=hidden_states, sequence_mask=sequence_mask)
209
+ [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
210
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
211
+ [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
212
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
213
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 360, in forward
214
+ [default0]:[rank0]: qkv_states = self.qkv_proj(
215
+ [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
216
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
217
+ [default0]:[rank0]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
218
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
219
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/nn.py", line 87, in forward
220
+ [default0]:[rank0]: return column_linear(
221
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/functional.py", line 359, in column_linear
222
+ [default0]:[rank0]: return F.linear(input, weight, bias)
223
+ [default0]:[rank0]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 384.00 MiB. GPU
224
+ [default2]:[rank2]: Traceback (most recent call last):
225
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
226
+ [default2]:[rank2]: trainer.train(dataloader)
227
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
228
+ [default2]:[rank2]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
229
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
230
+ [default2]:[rank2]: outputs = self.pipeline_engine.train_batch_iter(
231
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
232
+ [default2]:[rank2]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
233
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
234
+ [default2]:[rank2]: output = model(**micro_batch)
235
+ [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
236
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
237
+ [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
238
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
239
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
240
+ [default2]:[rank2]: sharded_logits = self.model(
241
+ [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
242
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
243
+ [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
244
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
245
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
246
+ [default2]:[rank2]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
247
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
248
+ [default2]:[rank2]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
249
+ [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
250
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
251
+ [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
252
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
253
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 151, in forward
254
+ [default2]:[rank2]: output = self.pp_block(**new_kwargs)
255
+ [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
256
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
257
+ [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
258
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
259
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 631, in forward
260
+ [default2]:[rank2]: output = self.attn(hidden_states=hidden_states, sequence_mask=sequence_mask)
261
+ [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
262
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
263
+ [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
264
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
265
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 360, in forward
266
+ [default2]:[rank2]: qkv_states = self.qkv_proj(
267
+ [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
268
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
269
+ [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
270
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
271
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/nn.py", line 87, in forward
272
+ [default2]:[rank2]: return column_linear(
273
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/functional.py", line 359, in column_linear
274
+ [default2]:[rank2]: return F.linear(input, weight, bias)
275
+ [default2]:[rank2]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 384.00 MiB. GPU  has a total capacity of 79.33 GiB of which 229.94 MiB is free. Including non-PyTorch memory, this process has 79.09 GiB memory in use. Of the allocated memory 66.66 GiB is allocated by PyTorch, and 144.27 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
276
+ [default1]:[rank1]: Traceback (most recent call last):
277
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
278
+ [default1]:[rank1]: trainer.train(dataloader)
279
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
280
+ [default1]:[rank1]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
281
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
282
+ [default1]:[rank1]: outputs = self.pipeline_engine.train_batch_iter(
283
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
284
+ [default1]:[rank1]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
285
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
286
+ [default1]:[rank1]: output = model(**micro_batch)
287
+ [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
288
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
289
+ [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
290
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
291
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
292
+ [default1]:[rank1]: sharded_logits = self.model(
293
+ [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
294
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
295
+ [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
296
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
297
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
298
+ [default1]:[rank1]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
299
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
300
+ [default1]:[rank1]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
301
+ [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
302
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
303
+ [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
304
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
305
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 151, in forward
306
+ [default1]:[rank1]: output = self.pp_block(**new_kwargs)
307
+ [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
308
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
309
+ [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
310
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
311
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 631, in forward
312
+ [default1]:[rank1]: output = self.attn(hidden_states=hidden_states, sequence_mask=sequence_mask)
313
+ [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
314
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
315
+ [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
316
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
317
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 360, in forward
318
+ [default1]:[rank1]: qkv_states = self.qkv_proj(
319
+ [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
320
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
321
+ [default1]:[rank1]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
322
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
323
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/nn.py", line 87, in forward
324
+ [default1]:[rank1]: return column_linear(
325
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/functional.py", line 359, in column_linear
326
+ [default1]:[rank1]: return F.linear(input, weight, bias)
327
+ [default1]:[rank1]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 384.00 MiB. GPU  has a total capacity of 79.33 GiB of which 229.94 MiB is free. Including non-PyTorch memory, this process has 79.09 GiB memory in use. Of the allocated memory 66.66 GiB is allocated by PyTorch, and 144.27 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
328
+ [default3]:[rank3]: Traceback (most recent call last):
329
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
330
+ [default3]:[rank3]: trainer.train(dataloader)
331
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
332
+ [default3]:[rank3]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
333
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
334
+ [default3]:[rank3]: outputs = self.pipeline_engine.train_batch_iter(
335
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
336
+ [default3]:[rank3]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
337
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
338
+ [default3]:[rank3]: output = model(**micro_batch)
339
+ [default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
340
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
341
+ [default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
342
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
343
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
344
+ [default3]:[rank3]: sharded_logits = self.model(
345
+ [default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
346
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
347
+ [default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
348
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
349
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
350
+ [default3]:[rank3]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
351
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
352
+ [default3]:[rank3]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
353
+ [default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
354
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
355
+ [default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
356
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
357
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 151, in forward
358
+ [default3]:[rank3]: output = self.pp_block(**new_kwargs)
359
+ [default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
360
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
361
+ [default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
362
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
363
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 631, in forward
364
+ [default3]:[rank3]: output = self.attn(hidden_states=hidden_states, sequence_mask=sequence_mask)
365
+ [default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
366
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
367
+ [default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
368
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
369
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 389, in forward
370
+ [default3]:[rank3]: .contiguous()
371
+ [default3]:[rank3]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 384.00 MiB. GPU  has a total capacity of 79.33 GiB of which 85.94 MiB is free. Including non-PyTorch memory, this process has 79.23 GiB memory in use. Of the allocated memory 67.03 GiB is allocated by PyTorch, and 144.27 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
372
+ [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.)
373
+ [default4]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
374
+ W0704 02:26:23.479000 140445158446912 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1142885 closing signal SIGTERM
375
+ W0704 02:26:23.480000 140445158446912 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1142886 closing signal SIGTERM
376
+ W0704 02:26:23.480000 140445158446912 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1142887 closing signal SIGTERM
377
+ W0704 02:26:23.481000 140445158446912 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1142888 closing signal SIGTERM
378
+ E0704 02:26:24.897000 140445158446912 torch/distributed/elastic/multiprocessing/api.py:826] failed (exitcode: 1) local_rank: 0 (pid: 1142881) of binary: /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/python3.10
379
+ Traceback (most recent call last):
380
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/torchrun", line 8, in <module>
381
+ sys.exit(main())
382
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper
383
+ return f(*args, **kwargs)
384
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 879, in main
385
+ run(args)
386
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 870, in run
387
+ elastic_launch(
388
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 132, in __call__
389
+ return launch_agent(self._config, self._entrypoint, list(args))
390
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 263, in launch_agent
391
+ raise ChildFailedError(
392
+ torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
393
+ ============================================================
394
+ /fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py FAILED
395
+ ------------------------------------------------------------
396
+ Failures:
397
+ [1]:
398
+ time : 2024-07-04_02:26:23
399
+ host : ip-26-0-171-88.ec2.internal
400
+ rank : 1 (local_rank: 1)
401
+ exitcode : 1 (pid: 1142882)
402
+ error_file: <N/A>
403
+ traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
404
+ [2]:
405
+ time : 2024-07-04_02:26:23
406
+ host : ip-26-0-171-88.ec2.internal
407
+ rank : 2 (local_rank: 2)
408
+ exitcode : 1 (pid: 1142883)
409
+ error_file: <N/A>
410
+ traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
411
+ [3]:
412
+ time : 2024-07-04_02:26:23
413
+ host : ip-26-0-171-88.ec2.internal
414
+ rank : 3 (local_rank: 3)
415
+ exitcode : 1 (pid: 1142884)
416
+ error_file: <N/A>
417
+ traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
418
+ ------------------------------------------------------------
419
+ Root Cause (first observed failure):
420
+ [0]:
421
+ time : 2024-07-04_02:26:23
422
+ host : ip-26-0-171-88.ec2.internal
423
+ rank : 0 (local_rank: 0)
424
+ exitcode : 1 (pid: 1142881)
425
+ error_file: <N/A>
426
+ traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
427
+ ============================================================
428
+ srun: error: ip-26-0-171-88: task 0: Exited with exit code 1
429
+ 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.
llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-8/status.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ oom