Hugging Face's logo

nanotron
/
old_bench

Model card Files Community
3outeille HF staff commited on
Commit
3ff6bc1
•
1 Parent(s): dcfd885

Upload llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32

Browse files
Files changed (4) hide show
  1. llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32/bench.slurm +111 -0
  2. llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32/config.yaml +90 -0
  3. llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32/log.out +433 -0
  4. llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32/status.txt +1 -0
llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32/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-1_tp-4_pp-2_mbz-32/log.out
13
+ #SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32/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-1_tp-4_pp-2_mbz-32/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-1_tp-4_pp-2_mbz-32/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-1_tp-4_pp-2_mbz-32/status.txt
84
+ else
85
+ if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32/log.out; then
86
+ printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32/status.txt
87
+ elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32/log.out; then
88
+ printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32/status.txt
89
+ elif grep -q "Timeout at NCCL" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32/log.out; then
90
+ printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32/status.txt
91
+ else
92
+ printf "fail" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32/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-1_tp-4_pp-2_mbz-32 --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-1_tp-4_pp-2_mbz-32 --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-1_tp-4_pp-2_mbz-32 llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32 --commit-message "Upload llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32"
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-1_tp-4_pp-2_mbz-32/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: 1
49
+ expert_parallel_size: 1
50
+ pp: 2
51
+ pp_engine: 1f1b
52
+ tp: 4
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-1_tp-4_pp-2_mbz-32
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: 32
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-1_tp-4_pp-2_mbz-32/log.out ADDED
@@ -0,0 +1,433 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ========================
2
+ START TIME: Wed Jul 3 22:51:03 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
+ W0703 22:51:05.903000 140623413176128 torch/distributed/run.py:757]
18
+ W0703 22:51:05.903000 140623413176128 torch/distributed/run.py:757] *****************************************
19
+ W0703 22:51:05.903000 140623413176128 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
+ W0703 22:51:05.903000 140623413176128 torch/distributed/run.py:757] *****************************************
21
+ [default0]:07/03/2024 22:51:22 [WARNING|DP=0|PP=0|TP=0|ip-26-0-161-178]: [Vocab Size Padding] Padded vocab (size: 50257) with 3 dummy tokens (new size: 50260)
22
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: Config:
23
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: Config(general=GeneralArgs(project='bench_cluster',
24
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: run='%date_%jobid',
25
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: seed=42,
26
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: step=None,
27
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: consumed_train_samples=None,
28
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: benchmark_csv_path=None,
29
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: ignore_sanity_checks=True),
30
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: parallelism=ParallelismArgs(dp=1,
31
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: pp=2,
32
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: tp=4,
33
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: pp_engine=<nanotron.parallel.pipeline_parallel.engine.OneForwardOneBackwardPipelineEngine object at 0x7f9abd3a88e0>,
34
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: tp_mode=<TensorParallelLinearMode.REDUCE_SCATTER: 2>,
35
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: tp_linear_async_communication=False,
36
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: expert_parallel_size=1),
37
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: model=ModelArgs(model_config=LlamaConfig(bos_token_id=1,
38
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: eos_token_id=2,
39
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: hidden_act='silu',
40
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: hidden_size=2048,
41
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: initializer_range=0.02,
42
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: intermediate_size=4096,
43
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: is_llama_config=True,
44
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: max_position_embeddings=4096,
45
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: num_attention_heads=32,
46
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: num_hidden_layers=24,
47
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: num_key_value_heads=32,
48
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: pad_token_id=None,
49
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: pretraining_tp=1,
50
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: rms_norm_eps=1e-05,
51
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: rope_scaling=None,
52
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: rope_theta=10000.0,
53
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: tie_word_embeddings=True,
54
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: use_cache=True,
55
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: vocab_size=50260),
56
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: init_method=RandomInit(std=0.025),
57
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: dtype=torch.bfloat16,
58
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: make_vocab_size_divisible_by=1,
59
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: ddp_bucket_cap_mb=25),
60
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: tokenizer=TokenizerArgs(tokenizer_name_or_path='openai-community/gpt2',
61
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: tokenizer_revision=None,
62
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: tokenizer_max_length=None),
63
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: checkpoints=CheckpointsArgs(checkpoints_path=Path('/dev/null'),
64
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: checkpoint_interval=100000,
65
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: save_initial_state=False,
66
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: resume_checkpoint_path=None,
67
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: checkpoints_path_is_shared_file_system=False),
68
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: logging=LoggingArgs(log_level='info',
69
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: log_level_replica='info',
70
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: iteration_step_info_interval=1),
71
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: tokens=TokensArgs(sequence_length=4096,
72
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: train_steps=20,
73
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: micro_batch_size=32,
74
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: batch_accumulation_per_replica=32,
75
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: val_check_interval=-1,
76
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: limit_val_batches=0,
77
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: limit_test_batches=0),
78
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: optimizer=OptimizerArgs(optimizer_factory=AdamWOptimizerArgs(adam_eps=1e-08,
79
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: adam_beta1=0.9,
80
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: adam_beta2=0.95,
81
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: torch_adam_is_fused=True,
82
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: name='adamW'),
83
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: zero_stage=1,
84
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: weight_decay=0.01,
85
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: clip_grad=1.0,
86
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: accumulate_grad_in_fp32=True,
87
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: learning_rate_scheduler=LRSchedulerArgs(learning_rate=0.0001,
88
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: lr_warmup_steps=1,
89
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: lr_warmup_style='linear',
90
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: lr_decay_style='linear',
91
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: lr_decay_steps=19,
92
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: lr_decay_starting_step=None,
93
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: min_decay_lr=1e-05)),
94
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: data_stages=[DatasetStageArgs(name='Training Stage',
95
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: start_training_step=1,
96
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: data=DataArgs(dataset=PretrainDatasetsArgs(hf_dataset_or_datasets='roneneldan/TinyStories',
97
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: hf_dataset_splits='train',
98
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: hf_dataset_config_name=None,
99
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: dataset_processing_num_proc_per_process=64,
100
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: dataset_overwrite_cache=False,
101
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: text_column_name='text'),
102
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: seed=42,
103
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: num_loading_workers=0))],
104
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: profiler=ProfilerArgs(profiler_export_path=Path('/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-32')),
105
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: lighteval=None)
106
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: Model Config:
107
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: LlamaConfig(bos_token_id=1,
108
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: eos_token_id=2,
109
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: hidden_act='silu',
110
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: hidden_size=2048,
111
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: initializer_range=0.02,
112
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: intermediate_size=4096,
113
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: is_llama_config=True,
114
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: max_position_embeddings=4096,
115
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: num_attention_heads=32,
116
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: num_hidden_layers=24,
117
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: num_key_value_heads=32,
118
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: pad_token_id=None,
119
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: pretraining_tp=1,
120
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: rms_norm_eps=1e-05,
121
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: rope_scaling=None,
122
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: rope_theta=10000.0,
123
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: tie_word_embeddings=True,
124
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: use_cache=True,
125
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: vocab_size=50260)
126
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: Building model..
127
+ [default0]:07/03/2024 22:51:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: Setting PP block ranks...
128
+ [default1]:07/03/2024 22:51:35 [INFO|DP=0|PP=0|TP=1|ip-26-0-161-178]: Local number of parameters: 173M (329.19MiB)
129
+ [default1]:07/03/2024 22:51:35 [INFO|DP=0|PP=0|TP=1|ip-26-0-161-178]: [After model building] Memory usage: 344.13MiB. Peak allocated: 346.16MiB Peak reserved: 348.00MiB
130
+ [default1]:07/03/2024 22:51:35 [INFO|DP=0|PP=0|TP=1|ip-26-0-161-178]: No checkpoint path provided.
131
+ [default3]:07/03/2024 22:51:35 [INFO|DP=0|PP=0|TP=3|ip-26-0-161-178]: Local number of parameters: 173M (329.19MiB)
132
+ [default3]:07/03/2024 22:51:35 [INFO|DP=0|PP=0|TP=3|ip-26-0-161-178]: [After model building] Memory usage: 344.13MiB. Peak allocated: 346.16MiB Peak reserved: 348.00MiB
133
+ [default3]:07/03/2024 22:51:35 [INFO|DP=0|PP=0|TP=3|ip-26-0-161-178]: No checkpoint path provided.
134
+ [default4]:07/03/2024 22:51:35 [INFO|DP=0|PP=1|TP=0|ip-26-0-161-178]: Local number of parameters: 131M (249.16MiB)
135
+ [default4]:07/03/2024 22:51:35 [INFO|DP=0|PP=1|TP=0|ip-26-0-161-178]: [After model building] Memory usage: 260.10MiB. Peak allocated: 262.13MiB Peak reserved: 264.00MiB
136
+ [default4]:07/03/2024 22:51:35 [INFO|DP=0|PP=1|TP=0|ip-26-0-161-178]: No checkpoint path provided.
137
+ [default7]:07/03/2024 22:51:35 [INFO|DP=0|PP=1|TP=3|ip-26-0-161-178]: Local number of parameters: 131M (249.16MiB)
138
+ [default7]:07/03/2024 22:51:35 [INFO|DP=0|PP=1|TP=3|ip-26-0-161-178]: [After model building] Memory usage: 260.10MiB. Peak allocated: 262.13MiB Peak reserved: 264.00MiB
139
+ [default7]:07/03/2024 22:51:35 [INFO|DP=0|PP=1|TP=3|ip-26-0-161-178]: No checkpoint path provided.
140
+ [default2]:07/03/2024 22:51:35 [INFO|DP=0|PP=0|TP=2|ip-26-0-161-178]: Local number of parameters: 173M (329.19MiB)
141
+ [default2]:07/03/2024 22:51:35 [INFO|DP=0|PP=0|TP=2|ip-26-0-161-178]: [After model building] Memory usage: 344.13MiB. Peak allocated: 346.16MiB Peak reserved: 348.00MiB
142
+ [default2]:07/03/2024 22:51:35 [INFO|DP=0|PP=0|TP=2|ip-26-0-161-178]: No checkpoint path provided.
143
+ [default0]:07/03/2024 22:51:35 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: Total number of parameters: 1.21G (2313.42MiB)
144
+ [default0]:07/03/2024 22:51:35 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: Local number of parameters: 173M (329.19MiB)
145
+ [default0]:07/03/2024 22:51:35 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: [After model building] Memory usage: 344.13MiB. Peak allocated: 346.16MiB Peak reserved: 348.00MiB
146
+ [default0]:07/03/2024 22:51:35 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: No checkpoint path provided.
147
+ [default0]:07/03/2024 22:51:35 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: Parametrizing model parameters using StandardParametrizator
148
+ [default5]:07/03/2024 22:51:35 [INFO|DP=0|PP=1|TP=1|ip-26-0-161-178]: Local number of parameters: 131M (249.16MiB)
149
+ [default5]:07/03/2024 22:51:35 [INFO|DP=0|PP=1|TP=1|ip-26-0-161-178]: [After model building] Memory usage: 260.10MiB. Peak allocated: 262.13MiB Peak reserved: 264.00MiB
150
+ [default5]:07/03/2024 22:51:35 [INFO|DP=0|PP=1|TP=1|ip-26-0-161-178]: No checkpoint path provided.
151
+ [default6]:07/03/2024 22:51:35 [INFO|DP=0|PP=1|TP=2|ip-26-0-161-178]: Local number of parameters: 131M (249.16MiB)
152
+ [default6]:07/03/2024 22:51:35 [INFO|DP=0|PP=1|TP=2|ip-26-0-161-178]: [After model building] Memory usage: 260.10MiB. Peak allocated: 262.13MiB Peak reserved: 264.00MiB
153
+ [default6]:07/03/2024 22:51:35 [INFO|DP=0|PP=1|TP=2|ip-26-0-161-178]: No checkpoint path provided.
154
+ [default0]:07/03/2024 22:51:36 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: [Optimizer Building] Using LearningRateForSP as learning rate
155
+ [default0]:07/03/2024 22:51:36 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: [ZeRO sharding] Size of optimizer params per rank:
156
+ [default0]:07/03/2024 22:51:36 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: [ZeRO sharding] DP Rank 0 has 173M out of 173M (100.00%) params' optimizer states
157
+ [default0]:07/03/2024 22:51:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: [Training Plan] Stage Training Stage has 19 remaining training steps and has consumed 0 samples
158
+ [default0]:07/03/2024 22:51:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: Using `datasets` library
159
+ [default0]:07/03/2024 22:51:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: Loading tokenizer from openai-community/gpt2 and transformers/hf_hub versions ('4.41.2', '0.23.4')
160
+ [default0]:07/03/2024 22:51:38 [WARNING|DP=0|PP=0|TP=0|ip-26-0-161-178]: Repo card metadata block was not found. Setting CardData to empty.
161
+ [default0]:Repo card metadata block was not found. Setting CardData to empty.
162
+ [default0]:07/03/2024 22:51:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: [Training Plan] There are 1 training stages
163
+ [default0]:07/03/2024 22:51:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: [Stage Training Stage] start from step 1
164
+ [default0]:07/03/2024 22:51:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]:
165
+ [default0]:07/03/2024 22:51:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: [Start training] datetime: 2024-07-03 22:51:38.760448 | mbs: 32 | grad_accum: 32 | global_batch_size: 1024 | sequence_length: 4096 | train_steps: 20 | start_iteration_step: 0 | consumed_train_samples: 0
166
+ [default0]:07/03/2024 22:51:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: Resuming training from stage Training Stage, it has trained for 0 samples and has 19 remaining train steps
167
+ [default0]:07/03/2024 22:51:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-161-178]: Memory usage: 1660.89MiB. Peak allocated 1660.89MiB. Peak reserved: 1668.00MiB
168
+ [default2]:Repo card metadata block was not found. Setting CardData to empty.
169
+ [default3]:07/03/2024 22:51:38 [WARNING|DP=0|PP=0|TP=3|ip-26-0-161-178]: Repo card metadata block was not found. Setting CardData to empty.
170
+ [default7]:07/03/2024 22:51:38 [WARNING|DP=0|PP=1|TP=3|ip-26-0-161-178]: Repo card metadata block was not found. Setting CardData to empty.
171
+ [default1]:Repo card metadata block was not found. Setting CardData to empty.
172
+ [default1]:07/03/2024 22:51:38 [WARNING|DP=0|PP=0|TP=1|ip-26-0-161-178]: Repo card metadata block was not found. Setting CardData to empty.
173
+ [default2]:07/03/2024 22:51:38 [WARNING|DP=0|PP=0|TP=2|ip-26-0-161-178]: Repo card metadata block was not found. Setting CardData to empty.
174
+ [default5]:07/03/2024 22:51:38 [WARNING|DP=0|PP=1|TP=1|ip-26-0-161-178]: Repo card metadata block was not found. Setting CardData to empty.
175
+ [default3]:Repo card metadata block was not found. Setting CardData to empty.
176
+ [default7]:Repo card metadata block was not found. Setting CardData to empty.
177
+ [default5]:Repo card metadata block was not found. Setting CardData to empty.
178
+ [default4]:07/03/2024 22:51:39 [WARNING|DP=0|PP=1|TP=0|ip-26-0-161-178]: Repo card metadata block was not found. Setting CardData to empty.
179
+ [default4]:Repo card metadata block was not found. Setting CardData to empty.
180
+ [default6]:07/03/2024 22:51:39 [WARNING|DP=0|PP=1|TP=2|ip-26-0-161-178]: Repo card metadata block was not found. Setting CardData to empty.
181
+ [default6]:Repo card metadata block was not found. Setting CardData to empty.
182
+ [default2]:[rank2]: Traceback (most recent call last):
183
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
184
+ [default2]:[rank2]: trainer.train(dataloader)
185
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
186
+ [default2]:[rank2]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
187
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
188
+ [default2]:[rank2]: outputs = self.pipeline_engine.train_batch_iter(
189
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
190
+ [default2]:[rank2]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
191
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
192
+ [default2]:[rank2]: output = model(**micro_batch)
193
+ [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
194
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
195
+ [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
196
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
197
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
198
+ [default2]:[rank2]: sharded_logits = self.model(
199
+ [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
200
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
201
+ [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
202
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
203
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
204
+ [default2]:[rank2]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
205
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
206
+ [default2]:[rank2]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
207
+ [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
208
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
209
+ [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
210
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
211
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 151, in forward
212
+ [default2]:[rank2]: output = self.pp_block(**new_kwargs)
213
+ [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
214
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
215
+ [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
216
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
217
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 637, in forward
218
+ [default2]:[rank2]: hidden_states = self.mlp(hidden_states=hidden_states)["hidden_states"]
219
+ [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
220
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
221
+ [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
222
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
223
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 171, in forward
224
+ [default2]:[rank2]: merged_states = self.gate_up_proj(hidden_states)
225
+ [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
226
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
227
+ [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
228
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
229
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/nn.py", line 87, in forward
230
+ [default2]:[rank2]: return column_linear(
231
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/functional.py", line 359, in column_linear
232
+ [default2]:[rank2]: return F.linear(input, weight, bias)
233
+ [default2]:[rank2]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 512.00 MiB. GPU  has a total capacity of 79.33 GiB of which 137.94 MiB is free. Including non-PyTorch memory, this process has 79.18 GiB memory in use. Of the allocated memory 66.63 GiB is allocated by PyTorch, and 271.33 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)
234
+ [default0]:[rank0]: Traceback (most recent call last):
235
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
236
+ [default0]:[rank0]: trainer.train(dataloader)
237
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
238
+ [default0]:[rank0]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
239
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
240
+ [default0]:[rank0]: outputs = self.pipeline_engine.train_batch_iter(
241
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
242
+ [default0]:[rank0]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
243
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
244
+ [default0]:[rank0]: output = model(**micro_batch)
245
+ [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
246
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
247
+ [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
248
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
249
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
250
+ [default0]:[rank0]: sharded_logits = self.model(
251
+ [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
252
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
253
+ [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
254
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
255
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
256
+ [default0]:[rank0]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
257
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
258
+ [default0]:[rank0]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
259
+ [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
260
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
261
+ [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
262
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
263
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 151, in forward
264
+ [default0]:[rank0]: output = self.pp_block(**new_kwargs)
265
+ [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
266
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
267
+ [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
268
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
269
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 637, in forward
270
+ [default0]:[rank0]: hidden_states = self.mlp(hidden_states=hidden_states)["hidden_states"]
271
+ [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
272
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
273
+ [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
274
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
275
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 171, in forward
276
+ [default0]:[rank0]: merged_states = self.gate_up_proj(hidden_states)
277
+ [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
278
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
279
+ [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
280
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
281
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/nn.py", line 87, in forward
282
+ [default0]:[rank0]: return column_linear(
283
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/functional.py", line 359, in column_linear
284
+ [default0]:[rank0]: return F.linear(input, weight, bias)
285
+ [default0]:[rank0]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 512.00 MiB. GPU
286
+ [default3]:[rank3]: Traceback (most recent call last):
287
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
288
+ [default3]:[rank3]: trainer.train(dataloader)
289
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
290
+ [default3]:[rank3]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
291
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
292
+ [default3]:[rank3]: outputs = self.pipeline_engine.train_batch_iter(
293
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
294
+ [default3]:[rank3]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
295
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
296
+ [default3]:[rank3]: output = model(**micro_batch)
297
+ [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
298
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
299
+ [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
300
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
301
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
302
+ [default3]:[rank3]: sharded_logits = self.model(
303
+ [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
304
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
305
+ [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
306
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
307
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
308
+ [default3]:[rank3]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
309
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
310
+ [default3]:[rank3]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
311
+ [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
312
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
313
+ [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
314
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
315
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 151, in forward
316
+ [default3]:[rank3]: output = self.pp_block(**new_kwargs)
317
+ [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
318
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
319
+ [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
320
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
321
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 637, in forward
322
+ [default3]:[rank3]: hidden_states = self.mlp(hidden_states=hidden_states)["hidden_states"]
323
+ [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
324
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
325
+ [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
326
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
327
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 171, in forward
328
+ [default3]:[rank3]: merged_states = self.gate_up_proj(hidden_states)
329
+ [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
330
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
331
+ [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
332
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
333
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/nn.py", line 87, in forward
334
+ [default3]:[rank3]: return column_linear(
335
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/functional.py", line 359, in column_linear
336
+ [default3]:[rank3]: return F.linear(input, weight, bias)
337
+ [default3]:[rank3]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 512.00 MiB. GPU  has a total capacity of 79.33 GiB of which 377.94 MiB is free. Including non-PyTorch memory, this process has 78.95 GiB memory in use. Of the allocated memory 66.63 GiB is allocated by PyTorch, and 271.33 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)
338
+ [default1]:[rank1]: Traceback (most recent call last):
339
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
340
+ [default1]:[rank1]: trainer.train(dataloader)
341
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
342
+ [default1]:[rank1]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
343
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
344
+ [default1]:[rank1]: outputs = self.pipeline_engine.train_batch_iter(
345
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
346
+ [default1]:[rank1]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
347
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
348
+ [default1]:[rank1]: output = model(**micro_batch)
349
+ [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
350
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
351
+ [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
352
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
353
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
354
+ [default1]:[rank1]: sharded_logits = self.model(
355
+ [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
356
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
357
+ [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
358
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
359
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
360
+ [default1]:[rank1]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
361
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
362
+ [default1]:[rank1]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
363
+ [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
364
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
365
+ [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
366
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
367
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 151, in forward
368
+ [default1]:[rank1]: output = self.pp_block(**new_kwargs)
369
+ [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
370
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
371
+ [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
372
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
373
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 637, in forward
374
+ [default1]:[rank1]: hidden_states = self.mlp(hidden_states=hidden_states)["hidden_states"]
375
+ [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
376
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
377
+ [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
378
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
379
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 171, in forward
380
+ [default1]:[rank1]: merged_states = self.gate_up_proj(hidden_states)
381
+ [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
382
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
383
+ [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
384
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
385
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/nn.py", line 87, in forward
386
+ [default1]:[rank1]: return column_linear(
387
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/functional.py", line 359, in column_linear
388
+ [default1]:[rank1]: return F.linear(input, weight, bias)
389
+ [default1]:[rank1]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 512.00 MiB. GPU  has a total capacity of 79.33 GiB of which 137.94 MiB is free. Including non-PyTorch memory, this process has 79.18 GiB memory in use. Of the allocated memory 66.63 GiB is allocated by PyTorch, and 271.33 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)
390
+ W0703 22:51:46.184000 140623413176128 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1034415 closing signal SIGTERM
391
+ W0703 22:51:46.184000 140623413176128 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1034416 closing signal SIGTERM
392
+ W0703 22:51:46.184000 140623413176128 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1034417 closing signal SIGTERM
393
+ W0703 22:51:46.184000 140623413176128 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1034418 closing signal SIGTERM
394
+ W0703 22:51:46.184000 140623413176128 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1034419 closing signal SIGTERM
395
+ W0703 22:51:46.184000 140623413176128 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1034420 closing signal SIGTERM
396
+ E0703 22:51:47.999000 140623413176128 torch/distributed/elastic/multiprocessing/api.py:826] failed (exitcode: 1) local_rank: 0 (pid: 1034413) of binary: /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/python3.10
397
+ Traceback (most recent call last):
398
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/torchrun", line 8, in <module>
399
+ sys.exit(main())
400
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper
401
+ return f(*args, **kwargs)
402
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 879, in main
403
+ run(args)
404
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 870, in run
405
+ elastic_launch(
406
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 132, in __call__
407
+ return launch_agent(self._config, self._entrypoint, list(args))
408
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 263, in launch_agent
409
+ raise ChildFailedError(
410
+ torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
411
+ ============================================================
412
+ /fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py FAILED
413
+ ------------------------------------------------------------
414
+ Failures:
415
+ [1]:
416
+ time : 2024-07-03_22:51:46
417
+ host : ip-26-0-161-178.ec2.internal
418
+ rank : 1 (local_rank: 1)
419
+ exitcode : 1 (pid: 1034414)
420
+ error_file: <N/A>
421
+ traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
422
+ ------------------------------------------------------------
423
+ Root Cause (first observed failure):
424
+ [0]:
425
+ time : 2024-07-03_22:51:46
426
+ host : ip-26-0-161-178.ec2.internal
427
+ rank : 0 (local_rank: 0)
428
+ exitcode : 1 (pid: 1034413)
429
+ error_file: <N/A>
430
+ traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
431
+ ============================================================
432
+ srun: error: ip-26-0-161-178: task 0: Exited with exit code 1
433
+ 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-1_tp-4_pp-2_mbz-32/status.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ oom