3outeille HF staff commited on
Commit
c4511e6
·
verified ·
1 Parent(s): f6b3cf0

Upload llama-1B/8_GPUS/dp-2_tp-4_pp-1_mbz-128

Browse files
llama-1B/8_GPUS/dp-2_tp-4_pp-1_mbz-128/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-2_tp-4_pp-1_mbz-128/log.out
13
+ #SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-4_pp-1_mbz-128/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-2_tp-4_pp-1_mbz-128/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-2_tp-4_pp-1_mbz-128/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-2_tp-4_pp-1_mbz-128/status.txt
84
+ else
85
+ if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-4_pp-1_mbz-128/log.out; then
86
+ printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-4_pp-1_mbz-128/status.txt
87
+ elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-4_pp-1_mbz-128/log.out; then
88
+ printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-4_pp-1_mbz-128/status.txt
89
+ elif grep -q "Timeout at NCCL" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-4_pp-1_mbz-128/log.out; then
90
+ printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-4_pp-1_mbz-128/status.txt
91
+ else
92
+ printf "fail" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-4_pp-1_mbz-128/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-2_tp-4_pp-1_mbz-128 --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-2_tp-4_pp-1_mbz-128 --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-2_tp-4_pp-1_mbz-128 llama-1B/8_GPUS/dp-2_tp-4_pp-1_mbz-128 --commit-message "Upload llama-1B/8_GPUS/dp-2_tp-4_pp-1_mbz-128"
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-2_tp-4_pp-1_mbz-128/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: 2
49
+ expert_parallel_size: 1
50
+ pp: 1
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-2_tp-4_pp-1_mbz-128
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: 4
79
+ limit_test_batches: 0
80
+ limit_val_batches: 0
81
+ micro_batch_size: 128
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-2_tp-4_pp-1_mbz-128/log.out ADDED
@@ -0,0 +1,676 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ========================
2
+ START TIME: Wed Jul 3 23:00:24 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 23:00:32.681000 140000305452864 torch/distributed/run.py:757]
18
+ W0703 23:00:32.681000 140000305452864 torch/distributed/run.py:757] *****************************************
19
+ W0703 23:00:32.681000 140000305452864 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 23:00:32.681000 140000305452864 torch/distributed/run.py:757] *****************************************
21
+ [default0]:07/03/2024 23:00:54 [WARNING|DP=0|PP=0|TP=0|ip-26-0-169-86]: [Vocab Size Padding] Padded vocab (size: 50257) with 3 dummy tokens (new size: 50260)
22
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Config:
23
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Config(general=GeneralArgs(project='bench_cluster',
24
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: run='%date_%jobid',
25
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: seed=42,
26
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: step=None,
27
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: consumed_train_samples=None,
28
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: benchmark_csv_path=None,
29
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: ignore_sanity_checks=True),
30
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: parallelism=ParallelismArgs(dp=2,
31
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: pp=1,
32
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tp=4,
33
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: pp_engine=<nanotron.parallel.pipeline_parallel.engine.OneForwardOneBackwardPipelineEngine object at 0x7fa635b20820>,
34
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tp_mode=<TensorParallelLinearMode.REDUCE_SCATTER: 2>,
35
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tp_linear_async_communication=False,
36
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: expert_parallel_size=1),
37
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: model=ModelArgs(model_config=LlamaConfig(bos_token_id=1,
38
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: eos_token_id=2,
39
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: hidden_act='silu',
40
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: hidden_size=2048,
41
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: initializer_range=0.02,
42
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: intermediate_size=4096,
43
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: is_llama_config=True,
44
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: max_position_embeddings=4096,
45
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: num_attention_heads=32,
46
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: num_hidden_layers=24,
47
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: num_key_value_heads=32,
48
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: pad_token_id=None,
49
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: pretraining_tp=1,
50
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: rms_norm_eps=1e-05,
51
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: rope_scaling=None,
52
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: rope_theta=10000.0,
53
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tie_word_embeddings=True,
54
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: use_cache=True,
55
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: vocab_size=50260),
56
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: init_method=RandomInit(std=0.025),
57
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: dtype=torch.bfloat16,
58
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: make_vocab_size_divisible_by=1,
59
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: ddp_bucket_cap_mb=25),
60
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tokenizer=TokenizerArgs(tokenizer_name_or_path='openai-community/gpt2',
61
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tokenizer_revision=None,
62
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tokenizer_max_length=None),
63
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: checkpoints=CheckpointsArgs(checkpoints_path=Path('/dev/null'),
64
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: checkpoint_interval=100000,
65
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: save_initial_state=False,
66
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: resume_checkpoint_path=None,
67
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: checkpoints_path_is_shared_file_system=False),
68
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: logging=LoggingArgs(log_level='info',
69
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: log_level_replica='info',
70
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: iteration_step_info_interval=1),
71
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tokens=TokensArgs(sequence_length=4096,
72
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: train_steps=20,
73
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: micro_batch_size=128,
74
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: batch_accumulation_per_replica=4,
75
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: val_check_interval=-1,
76
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: limit_val_batches=0,
77
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: limit_test_batches=0),
78
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: optimizer=OptimizerArgs(optimizer_factory=AdamWOptimizerArgs(adam_eps=1e-08,
79
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: adam_beta1=0.9,
80
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: adam_beta2=0.95,
81
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: torch_adam_is_fused=True,
82
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: name='adamW'),
83
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: zero_stage=1,
84
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: weight_decay=0.01,
85
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: clip_grad=1.0,
86
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: accumulate_grad_in_fp32=True,
87
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: learning_rate_scheduler=LRSchedulerArgs(learning_rate=0.0001,
88
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: lr_warmup_steps=1,
89
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: lr_warmup_style='linear',
90
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: lr_decay_style='linear',
91
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: lr_decay_steps=19,
92
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: lr_decay_starting_step=None,
93
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: min_decay_lr=1e-05)),
94
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: data_stages=[DatasetStageArgs(name='Training Stage',
95
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: start_training_step=1,
96
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: data=DataArgs(dataset=PretrainDatasetsArgs(hf_dataset_or_datasets='roneneldan/TinyStories',
97
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: hf_dataset_splits='train',
98
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: hf_dataset_config_name=None,
99
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: dataset_processing_num_proc_per_process=64,
100
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: dataset_overwrite_cache=False,
101
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: text_column_name='text'),
102
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: seed=42,
103
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: num_loading_workers=0))],
104
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: profiler=ProfilerArgs(profiler_export_path=Path('/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-4_pp-1_mbz-128')),
105
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: lighteval=None)
106
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Model Config:
107
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: LlamaConfig(bos_token_id=1,
108
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: eos_token_id=2,
109
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: hidden_act='silu',
110
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: hidden_size=2048,
111
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: initializer_range=0.02,
112
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: intermediate_size=4096,
113
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: is_llama_config=True,
114
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: max_position_embeddings=4096,
115
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: num_attention_heads=32,
116
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: num_hidden_layers=24,
117
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: num_key_value_heads=32,
118
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: pad_token_id=None,
119
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: pretraining_tp=1,
120
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: rms_norm_eps=1e-05,
121
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: rope_scaling=None,
122
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: rope_theta=10000.0,
123
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tie_word_embeddings=True,
124
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: use_cache=True,
125
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: vocab_size=50260)
126
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Building model..
127
+ [default0]:07/03/2024 23:00:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Setting PP block ranks...
128
+ [default0]:07/03/2024 23:01:07 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Total number of parameters: 1.11G (2117.09MiB)
129
+ [default0]:07/03/2024 23:01:07 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Local number of parameters: 277M (529.27MiB)
130
+ [default0]:07/03/2024 23:01:07 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [After model building] Memory usage: 554.21MiB. Peak allocated: 606.24MiB Peak reserved: 608.00MiB
131
+ [default3]:07/03/2024 23:01:07 [INFO|DP=0|PP=0|TP=3|ip-26-0-169-86]: Local number of parameters: 277M (529.27MiB)
132
+ [default3]:07/03/2024 23:01:07 [INFO|DP=0|PP=0|TP=3|ip-26-0-169-86]: [After model building] Memory usage: 554.21MiB. Peak allocated: 606.24MiB Peak reserved: 608.00MiB
133
+ [default3]:07/03/2024 23:01:07 [INFO|DP=0|PP=0|TP=3|ip-26-0-169-86]: No checkpoint path provided.
134
+ [default2]:07/03/2024 23:01:07 [INFO|DP=0|PP=0|TP=2|ip-26-0-169-86]: Local number of parameters: 277M (529.27MiB)
135
+ [default2]:07/03/2024 23:01:07 [INFO|DP=0|PP=0|TP=2|ip-26-0-169-86]: [After model building] Memory usage: 554.21MiB. Peak allocated: 606.24MiB Peak reserved: 608.00MiB
136
+ [default2]:07/03/2024 23:01:07 [INFO|DP=0|PP=0|TP=2|ip-26-0-169-86]: No checkpoint path provided.
137
+ [default1]:07/03/2024 23:01:07 [INFO|DP=0|PP=0|TP=1|ip-26-0-169-86]: Local number of parameters: 277M (529.27MiB)
138
+ [default1]:07/03/2024 23:01:07 [INFO|DP=0|PP=0|TP=1|ip-26-0-169-86]: [After model building] Memory usage: 554.21MiB. Peak allocated: 606.24MiB Peak reserved: 608.00MiB
139
+ [default1]:07/03/2024 23:01:07 [INFO|DP=0|PP=0|TP=1|ip-26-0-169-86]: No checkpoint path provided.
140
+ [default0]:07/03/2024 23:01:07 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: No checkpoint path provided.
141
+ [default0]:07/03/2024 23:01:07 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Parametrizing model parameters using StandardParametrizator
142
+ [default5]:07/03/2024 23:01:07 [INFO|DP=1|PP=0|TP=1|ip-26-0-169-86]: No checkpoint path provided.
143
+ [default7]:07/03/2024 23:01:07 [INFO|DP=1|PP=0|TP=3|ip-26-0-169-86]: No checkpoint path provided.
144
+ [default6]:07/03/2024 23:01:07 [INFO|DP=1|PP=0|TP=2|ip-26-0-169-86]: No checkpoint path provided.
145
+ [default4]:07/03/2024 23:01:07 [INFO|DP=1|PP=0|TP=0|ip-26-0-169-86]: No checkpoint path provided.
146
+ [default0]:07/03/2024 23:01:09 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [Optimizer Building] Using LearningRateForSP as learning rate
147
+ [default0]:07/03/2024 23:01:09 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [ZeRO sharding] Size of optimizer params per rank:
148
+ [default0]:07/03/2024 23:01:09 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [ZeRO sharding] DP Rank 0 has 139M out of 277M (50.00%) params' optimizer states
149
+ [default0]:07/03/2024 23:01:09 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [ZeRO sharding] DP Rank 1 has 139M out of 277M (50.00%) params' optimizer states
150
+ [default0]:07/03/2024 23:01:11 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [Training Plan] Stage Training Stage has 19 remaining training steps and has consumed 0 samples
151
+ [default0]:07/03/2024 23:01:11 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Using `datasets` library
152
+ [default0]:07/03/2024 23:01:11 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Loading tokenizer from openai-community/gpt2 and transformers/hf_hub versions ('4.41.2', '0.23.4')
153
+ [default0]:07/03/2024 23:01:11 [WARNING|DP=0|PP=0|TP=0|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty.
154
+ [default0]:Repo card metadata block was not found. Setting CardData to empty.
155
+ [default0]:07/03/2024 23:01:13 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [Training Plan] There are 1 training stages
156
+ [default0]:07/03/2024 23:01:13 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [Stage Training Stage] start from step 1
157
+ [default0]:07/03/2024 23:01:13 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]:
158
+ [default0]:07/03/2024 23:01:13 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [Start training] datetime: 2024-07-03 23:01:13.683845 | mbs: 128 | grad_accum: 4 | global_batch_size: 1024 | sequence_length: 4096 | train_steps: 20 | start_iteration_step: 0 | consumed_train_samples: 0
159
+ [default0]:07/03/2024 23:01:13 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Resuming training from stage Training Stage, it has trained for 0 samples and has 19 remaining train steps
160
+ [default0]:07/03/2024 23:01:13 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Memory usage: 2142.76MiB. Peak allocated 2142.76MiB. Peak reserved: 2198.00MiB
161
+ [default7]:07/03/2024 23:01:13 [WARNING|DP=1|PP=0|TP=3|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty.
162
+ [default6]:07/03/2024 23:01:13 [WARNING|DP=1|PP=0|TP=2|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty.
163
+ [default7]:Repo card metadata block was not found. Setting CardData to empty.
164
+ [default6]:Repo card metadata block was not found. Setting CardData to empty.
165
+ [default3]:07/03/2024 23:01:13 [WARNING|DP=0|PP=0|TP=3|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty.
166
+ [default2]:07/03/2024 23:01:13 [WARNING|DP=0|PP=0|TP=2|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty.
167
+ [default3]: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
+ [default1]:Repo card metadata block was not found. Setting CardData to empty.
170
+ [default1]:07/03/2024 23:01:13 [WARNING|DP=0|PP=0|TP=1|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty.
171
+ [default5]:Repo card metadata block was not found. Setting CardData to empty.
172
+ [default5]:07/03/2024 23:01:13 [WARNING|DP=1|PP=0|TP=1|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty.
173
+ [default4]:07/03/2024 23:01:13 [WARNING|DP=1|PP=0|TP=0|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty.
174
+ [default4]:Repo card metadata block was not found. Setting CardData to empty.
175
+ [default2]:[rank2]: Traceback (most recent call last):
176
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
177
+ [default2]:[rank2]: trainer.train(dataloader)
178
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
179
+ [default2]:[rank2]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
180
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
181
+ [default2]:[rank2]: outputs = self.pipeline_engine.train_batch_iter(
182
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
183
+ [default2]:[rank2]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
184
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
185
+ [default2]:[rank2]: output = model(**micro_batch)
186
+ [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
187
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
188
+ [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
189
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
190
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
191
+ [default2]:[rank2]: sharded_logits = self.model(
192
+ [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
193
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
194
+ [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
195
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
196
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
197
+ [default2]:[rank2]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
198
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
199
+ [default2]:[rank2]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
200
+ [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
201
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
202
+ [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
203
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
204
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 151, in forward
205
+ [default2]:[rank2]: output = self.pp_block(**new_kwargs)
206
+ [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
207
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
208
+ [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
209
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
210
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 637, in forward
211
+ [default2]:[rank2]: hidden_states = self.mlp(hidden_states=hidden_states)["hidden_states"]
212
+ [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
213
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
214
+ [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
215
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
216
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 172, in forward
217
+ [default2]:[rank2]: hidden_states = self.down_proj(self.split_silu_mul(merged_states))
218
+ [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
219
+ [default2]:[rank2]: return self._call_impl(*args, **kwargs)
220
+ [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
221
+ [default2]:[rank2]: return forward_call(*args, **kwargs)
222
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/nn.py", line 159, in forward
223
+ [default2]:[rank2]: return row_linear(
224
+ [default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/functional.py", line 474, in row_linear
225
+ [default2]:[rank2]: out = F.linear(input, weight, bias)
226
+ [default2]:[rank2]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.00 GiB. GPU  has a total capacity of 79.33 GiB of which 1.74 GiB is free. Including non-PyTorch memory, this process has 77.58 GiB memory in use. Of the allocated memory 65.72 GiB is allocated by PyTorch, and 410.23 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)
227
+ [default1]:[rank1]: Traceback (most recent call last):
228
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
229
+ [default1]:[rank1]: trainer.train(dataloader)
230
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
231
+ [default1]:[rank1]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
232
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
233
+ [default1]:[rank1]: outputs = self.pipeline_engine.train_batch_iter(
234
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
235
+ [default1]:[rank1]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
236
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
237
+ [default1]:[rank1]: output = model(**micro_batch)
238
+ [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
239
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
240
+ [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
241
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
242
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
243
+ [default1]:[rank1]: sharded_logits = self.model(
244
+ [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
245
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
246
+ [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
247
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
248
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
249
+ [default1]:[rank1]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
250
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
251
+ [default1]:[rank1]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
252
+ [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
253
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
254
+ [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
255
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
256
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 151, in forward
257
+ [default1]:[rank1]: output = self.pp_block(**new_kwargs)
258
+ [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
259
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
260
+ [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
261
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
262
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 637, in forward
263
+ [default1]:[rank1]: hidden_states = self.mlp(hidden_states=hidden_states)["hidden_states"]
264
+ [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
265
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
266
+ [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
267
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
268
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 172, in forward
269
+ [default1]:[rank1]: hidden_states = self.down_proj(self.split_silu_mul(merged_states))
270
+ [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
271
+ [default1]:[rank1]: return self._call_impl(*args, **kwargs)
272
+ [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
273
+ [default1]:[rank1]: return forward_call(*args, **kwargs)
274
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/nn.py", line 159, in forward
275
+ [default1]:[rank1]: return row_linear(
276
+ [default1]:[rank1]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/functional.py", line 474, in row_linear
277
+ [default1]:[rank1]: out = F.linear(input, weight, bias)
278
+ [default1]:[rank1]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.00 GiB. GPU  has a total capacity of 79.33 GiB of which 1.74 GiB is free. Including non-PyTorch memory, this process has 77.58 GiB memory in use. Of the allocated memory 65.72 GiB is allocated by PyTorch, and 410.23 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)
279
+ [default0]:[rank0]: Traceback (most recent call last):
280
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
281
+ [default0]:[rank0]: trainer.train(dataloader)
282
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
283
+ [default0]:[rank0]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
284
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
285
+ [default0]:[rank0]: outputs = self.pipeline_engine.train_batch_iter(
286
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
287
+ [default0]:[rank0]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
288
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
289
+ [default0]:[rank0]: output = model(**micro_batch)
290
+ [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
291
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
292
+ [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
293
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
294
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
295
+ [default0]:[rank0]: sharded_logits = self.model(
296
+ [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
297
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
298
+ [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
299
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
300
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
301
+ [default0]:[rank0]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
302
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
303
+ [default0]:[rank0]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
304
+ [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
305
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
306
+ [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
307
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
308
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 151, in forward
309
+ [default0]:[rank0]: output = self.pp_block(**new_kwargs)
310
+ [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
311
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
312
+ [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
313
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
314
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 637, in forward
315
+ [default0]:[rank0]: hidden_states = self.mlp(hidden_states=hidden_states)["hidden_states"]
316
+ [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
317
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
318
+ [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
319
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
320
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 172, in forward
321
+ [default0]:[rank0]: hidden_states = self.down_proj(self.split_silu_mul(merged_states))
322
+ [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
323
+ [default0]:[rank0]: return self._call_impl(*args, **kwargs)
324
+ [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
325
+ [default0]:[rank0]: return forward_call(*args, **kwargs)
326
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/nn.py", line 159, in forward
327
+ [default0]:[rank0]: return row_linear(
328
+ [default0]:[rank0]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/functional.py", line 474, in row_linear
329
+ [default0]:[rank0]: out = F.linear(input, weight, bias)
330
+ [default0]:[rank0]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.00 GiB. GPU
331
+ [default3]:[rank3]: Traceback (most recent call last):
332
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
333
+ [default3]:[rank3]: trainer.train(dataloader)
334
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
335
+ [default3]:[rank3]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
336
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
337
+ [default3]:[rank3]: outputs = self.pipeline_engine.train_batch_iter(
338
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
339
+ [default3]:[rank3]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
340
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
341
+ [default3]:[rank3]: output = model(**micro_batch)
342
+ [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
343
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
344
+ [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
345
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
346
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
347
+ [default3]:[rank3]: sharded_logits = self.model(
348
+ [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
349
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
350
+ [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
351
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
352
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
353
+ [default3]:[rank3]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
354
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
355
+ [default3]:[rank3]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
356
+ [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
357
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
358
+ [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
359
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
360
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 151, in forward
361
+ [default3]:[rank3]: output = self.pp_block(**new_kwargs)
362
+ [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
363
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
364
+ [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
365
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
366
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 637, in forward
367
+ [default3]:[rank3]: hidden_states = self.mlp(hidden_states=hidden_states)["hidden_states"]
368
+ [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
369
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
370
+ [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
371
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
372
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 172, in forward
373
+ [default3]:[rank3]: hidden_states = self.down_proj(self.split_silu_mul(merged_states))
374
+ [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
375
+ [default3]:[rank3]: return self._call_impl(*args, **kwargs)
376
+ [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
377
+ [default3]:[rank3]: return forward_call(*args, **kwargs)
378
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/nn.py", line 159, in forward
379
+ [default3]:[rank3]: return row_linear(
380
+ [default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/functional.py", line 474, in row_linear
381
+ [default3]:[rank3]: out = F.linear(input, weight, bias)
382
+ [default3]:[rank3]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.00 GiB. GPU  has a total capacity of 79.33 GiB of which 1.97 GiB is free. Including non-PyTorch memory, this process has 77.35 GiB memory in use. Of the allocated memory 65.72 GiB is allocated by PyTorch, and 410.23 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)
383
+ [default5]:[rank5]: Traceback (most recent call last):
384
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
385
+ [default5]:[rank5]: trainer.train(dataloader)
386
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
387
+ [default5]:[rank5]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
388
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
389
+ [default5]:[rank5]: outputs = self.pipeline_engine.train_batch_iter(
390
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
391
+ [default5]:[rank5]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
392
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
393
+ [default5]:[rank5]: output = model(**micro_batch)
394
+ [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
395
+ [default5]:[rank5]: return self._call_impl(*args, **kwargs)
396
+ [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
397
+ [default5]:[rank5]: return forward_call(*args, **kwargs)
398
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
399
+ [default5]:[rank5]: sharded_logits = self.model(
400
+ [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
401
+ [default5]:[rank5]: return self._call_impl(*args, **kwargs)
402
+ [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
403
+ [default5]:[rank5]: return forward_call(*args, **kwargs)
404
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
405
+ [default5]:[rank5]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
406
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
407
+ [default5]:[rank5]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
408
+ [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
409
+ [default5]:[rank5]: return self._call_impl(*args, **kwargs)
410
+ [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
411
+ [default5]:[rank5]: return forward_call(*args, **kwargs)
412
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 151, in forward
413
+ [default5]:[rank5]: output = self.pp_block(**new_kwargs)
414
+ [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
415
+ [default5]:[rank5]: return self._call_impl(*args, **kwargs)
416
+ [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
417
+ [default5]:[rank5]: return forward_call(*args, **kwargs)
418
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 637, in forward
419
+ [default5]:[rank5]: hidden_states = self.mlp(hidden_states=hidden_states)["hidden_states"]
420
+ [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
421
+ [default5]:[rank5]: return self._call_impl(*args, **kwargs)
422
+ [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
423
+ [default5]:[rank5]: return forward_call(*args, **kwargs)
424
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 172, in forward
425
+ [default5]:[rank5]: hidden_states = self.down_proj(self.split_silu_mul(merged_states))
426
+ [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
427
+ [default5]:[rank5]: return self._call_impl(*args, **kwargs)
428
+ [default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
429
+ [default5]:[rank5]: return forward_call(*args, **kwargs)
430
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/nn.py", line 159, in forward
431
+ [default5]:[rank5]: return row_linear(
432
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/functional.py", line 474, in row_linear
433
+ [default5]:[rank5]: out = F.linear(input, weight, bias)
434
+ [default5]:[rank5]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.00 GiB. GPU  has a total capacity of 79.33 GiB of which 1.74 GiB is free. Including non-PyTorch memory, this process has 77.58 GiB memory in use. Of the allocated memory 65.72 GiB is allocated by PyTorch, and 410.23 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)
435
+ [default7]:[rank7]: Traceback (most recent call last):
436
+ [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
437
+ [default7]:[rank7]: trainer.train(dataloader)
438
+ [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
439
+ [default7]:[rank7]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
440
+ [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
441
+ [default7]:[rank7]: outputs = self.pipeline_engine.train_batch_iter(
442
+ [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
443
+ [default7]:[rank7]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
444
+ [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
445
+ [default7]:[rank7]: output = model(**micro_batch)
446
+ [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
447
+ [default7]:[rank7]: return self._call_impl(*args, **kwargs)
448
+ [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
449
+ [default7]:[rank7]: return forward_call(*args, **kwargs)
450
+ [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
451
+ [default7]:[rank7]: sharded_logits = self.model(
452
+ [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
453
+ [default7]:[rank7]: return self._call_impl(*args, **kwargs)
454
+ [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
455
+ [default7]:[rank7]: return forward_call(*args, **kwargs)
456
+ [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
457
+ [default7]:[rank7]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
458
+ [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
459
+ [default7]:[rank7]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
460
+ [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
461
+ [default7]:[rank7]: return self._call_impl(*args, **kwargs)
462
+ [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
463
+ [default7]:[rank7]: return forward_call(*args, **kwargs)
464
+ [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 151, in forward
465
+ [default7]:[rank7]: output = self.pp_block(**new_kwargs)
466
+ [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
467
+ [default7]:[rank7]: return self._call_impl(*args, **kwargs)
468
+ [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
469
+ [default7]:[rank7]: return forward_call(*args, **kwargs)
470
+ [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 637, in forward
471
+ [default7]:[rank7]: hidden_states = self.mlp(hidden_states=hidden_states)["hidden_states"]
472
+ [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
473
+ [default7]:[rank7]: return self._call_impl(*args, **kwargs)
474
+ [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
475
+ [default7]:[rank7]: return forward_call(*args, **kwargs)
476
+ [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 172, in forward
477
+ [default7]:[rank7]: hidden_states = self.down_proj(self.split_silu_mul(merged_states))
478
+ [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
479
+ [default7]:[rank7]: return self._call_impl(*args, **kwargs)
480
+ [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
481
+ [default7]:[rank7]: return forward_call(*args, **kwargs)
482
+ [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/nn.py", line 159, in forward
483
+ [default7]:[rank7]: return row_linear(
484
+ [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/functional.py", line 479, in row_linear
485
+ [default7]:[rank7]: out = differentiable_reduce_scatter_sum(out, group=group)
486
+ [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/distributed_differentiable_primitives.py", line 145, in differentiable_reduce_scatter_sum
487
+ [default7]:[rank7]: return DifferentiableReduceScatterSum.apply(tensor, group)
488
+ [default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/function.py", line 598, in apply
489
+ [default7]:[rank7]: return super().apply(*args, **kwargs) # type: ignore[misc]
490
+ [default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/distributed_differentiable_primitives.py", line 111, in forward
491
+ [default7]:[rank7]: sharded_tensor = torch.empty(
492
+ [default7]:[rank7]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 512.00 MiB. GPU  has a total capacity of 79.33 GiB of which 449.94 MiB is free. Including non-PyTorch memory, this process has 78.88 GiB memory in use. Of the allocated memory 67.72 GiB is allocated by PyTorch, and 410.23 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)
493
+ [default6]:[rank6]: Traceback (most recent call last):
494
+ [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
495
+ [default6]:[rank6]: trainer.train(dataloader)
496
+ [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
497
+ [default6]:[rank6]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
498
+ [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
499
+ [default6]:[rank6]: outputs = self.pipeline_engine.train_batch_iter(
500
+ [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
501
+ [default6]:[rank6]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
502
+ [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
503
+ [default6]:[rank6]: output = model(**micro_batch)
504
+ [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
505
+ [default6]:[rank6]: return self._call_impl(*args, **kwargs)
506
+ [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
507
+ [default6]:[rank6]: return forward_call(*args, **kwargs)
508
+ [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
509
+ [default6]:[rank6]: sharded_logits = self.model(
510
+ [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
511
+ [default6]:[rank6]: return self._call_impl(*args, **kwargs)
512
+ [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
513
+ [default6]:[rank6]: return forward_call(*args, **kwargs)
514
+ [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
515
+ [default6]:[rank6]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
516
+ [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
517
+ [default6]:[rank6]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
518
+ [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
519
+ [default6]:[rank6]: return self._call_impl(*args, **kwargs)
520
+ [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
521
+ [default6]:[rank6]: return forward_call(*args, **kwargs)
522
+ [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 151, in forward
523
+ [default6]:[rank6]: output = self.pp_block(**new_kwargs)
524
+ [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
525
+ [default6]:[rank6]: return self._call_impl(*args, **kwargs)
526
+ [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
527
+ [default6]:[rank6]: return forward_call(*args, **kwargs)
528
+ [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 637, in forward
529
+ [default6]:[rank6]: hidden_states = self.mlp(hidden_states=hidden_states)["hidden_states"]
530
+ [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
531
+ [default6]:[rank6]: return self._call_impl(*args, **kwargs)
532
+ [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
533
+ [default6]:[rank6]: return forward_call(*args, **kwargs)
534
+ [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 172, in forward
535
+ [default6]:[rank6]: hidden_states = self.down_proj(self.split_silu_mul(merged_states))
536
+ [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
537
+ [default6]:[rank6]: return self._call_impl(*args, **kwargs)
538
+ [default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
539
+ [default6]:[rank6]: return forward_call(*args, **kwargs)
540
+ [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/nn.py", line 159, in forward
541
+ [default6]:[rank6]: return row_linear(
542
+ [default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/functional.py", line 474, in row_linear
543
+ [default6]:[rank6]: out = F.linear(input, weight, bias)
544
+ [default6]:[rank6]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.00 GiB. GPU  has a total capacity of 79.33 GiB of which 1.74 GiB is free. Including non-PyTorch memory, this process has 77.58 GiB memory in use. Of the allocated memory 65.72 GiB is allocated by PyTorch, and 410.23 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)
545
+ [default4]:[rank4]: Traceback (most recent call last):
546
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
547
+ [default4]:[rank4]: trainer.train(dataloader)
548
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
549
+ [default4]:[rank4]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
550
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
551
+ [default4]:[rank4]: outputs = self.pipeline_engine.train_batch_iter(
552
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
553
+ [default4]:[rank4]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
554
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
555
+ [default4]:[rank4]: output = model(**micro_batch)
556
+ [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
557
+ [default4]:[rank4]: return self._call_impl(*args, **kwargs)
558
+ [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
559
+ [default4]:[rank4]: return forward_call(*args, **kwargs)
560
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
561
+ [default4]:[rank4]: sharded_logits = self.model(
562
+ [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
563
+ [default4]:[rank4]: return self._call_impl(*args, **kwargs)
564
+ [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
565
+ [default4]:[rank4]: return forward_call(*args, **kwargs)
566
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
567
+ [default4]:[rank4]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
568
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
569
+ [default4]:[rank4]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
570
+ [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
571
+ [default4]:[rank4]: return self._call_impl(*args, **kwargs)
572
+ [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
573
+ [default4]:[rank4]: return forward_call(*args, **kwargs)
574
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 151, in forward
575
+ [default4]:[rank4]: output = self.pp_block(**new_kwargs)
576
+ [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
577
+ [default4]:[rank4]: return self._call_impl(*args, **kwargs)
578
+ [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
579
+ [default4]:[rank4]: return forward_call(*args, **kwargs)
580
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 637, in forward
581
+ [default4]:[rank4]: hidden_states = self.mlp(hidden_states=hidden_states)["hidden_states"]
582
+ [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
583
+ [default4]:[rank4]: return self._call_impl(*args, **kwargs)
584
+ [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
585
+ [default4]:[rank4]: return forward_call(*args, **kwargs)
586
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 172, in forward
587
+ [default4]:[rank4]: hidden_states = self.down_proj(self.split_silu_mul(merged_states))
588
+ [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
589
+ [default4]:[rank4]: return self._call_impl(*args, **kwargs)
590
+ [default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
591
+ [default4]:[rank4]: return forward_call(*args, **kwargs)
592
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/nn.py", line 159, in forward
593
+ [default4]:[rank4]: return row_linear(
594
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/tensor_parallel/functional.py", line 474, in row_linear
595
+ [default4]:[rank4]: out = F.linear(input, weight, bias)
596
+ [default4]:[rank4]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.00 GiB. GPU  has a total capacity of 79.33 GiB of which 1.78 GiB is free. Including non-PyTorch memory, this process has 77.54 GiB memory in use. Of the allocated memory 65.72 GiB is allocated by PyTorch, and 410.23 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)
597
+ E0703 23:01:28.066000 140000305452864 torch/distributed/elastic/multiprocessing/api.py:826] failed (exitcode: 1) local_rank: 0 (pid: 2026681) of binary: /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/python3.10
598
+ Traceback (most recent call last):
599
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/torchrun", line 8, in <module>
600
+ sys.exit(main())
601
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper
602
+ return f(*args, **kwargs)
603
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 879, in main
604
+ run(args)
605
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 870, in run
606
+ elastic_launch(
607
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 132, in __call__
608
+ return launch_agent(self._config, self._entrypoint, list(args))
609
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 263, in launch_agent
610
+ raise ChildFailedError(
611
+ torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
612
+ ============================================================
613
+ /fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py FAILED
614
+ ------------------------------------------------------------
615
+ Failures:
616
+ [1]:
617
+ time : 2024-07-03_23:01:27
618
+ host : ip-26-0-169-86.ec2.internal
619
+ rank : 1 (local_rank: 1)
620
+ exitcode : 1 (pid: 2026682)
621
+ error_file: <N/A>
622
+ traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
623
+ [2]:
624
+ time : 2024-07-03_23:01:27
625
+ host : ip-26-0-169-86.ec2.internal
626
+ rank : 2 (local_rank: 2)
627
+ exitcode : 1 (pid: 2026683)
628
+ error_file: <N/A>
629
+ traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
630
+ [3]:
631
+ time : 2024-07-03_23:01:27
632
+ host : ip-26-0-169-86.ec2.internal
633
+ rank : 3 (local_rank: 3)
634
+ exitcode : 1 (pid: 2026684)
635
+ error_file: <N/A>
636
+ traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
637
+ [4]:
638
+ time : 2024-07-03_23:01:27
639
+ host : ip-26-0-169-86.ec2.internal
640
+ rank : 4 (local_rank: 4)
641
+ exitcode : 1 (pid: 2026685)
642
+ error_file: <N/A>
643
+ traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
644
+ [5]:
645
+ time : 2024-07-03_23:01:27
646
+ host : ip-26-0-169-86.ec2.internal
647
+ rank : 5 (local_rank: 5)
648
+ exitcode : 1 (pid: 2026686)
649
+ error_file: <N/A>
650
+ traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
651
+ [6]:
652
+ time : 2024-07-03_23:01:27
653
+ host : ip-26-0-169-86.ec2.internal
654
+ rank : 6 (local_rank: 6)
655
+ exitcode : 1 (pid: 2026687)
656
+ error_file: <N/A>
657
+ traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
658
+ [7]:
659
+ time : 2024-07-03_23:01:27
660
+ host : ip-26-0-169-86.ec2.internal
661
+ rank : 7 (local_rank: 7)
662
+ exitcode : 1 (pid: 2026688)
663
+ error_file: <N/A>
664
+ traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
665
+ ------------------------------------------------------------
666
+ Root Cause (first observed failure):
667
+ [0]:
668
+ time : 2024-07-03_23:01:27
669
+ host : ip-26-0-169-86.ec2.internal
670
+ rank : 0 (local_rank: 0)
671
+ exitcode : 1 (pid: 2026681)
672
+ error_file: <N/A>
673
+ traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
674
+ ============================================================
675
+ srun: error: ip-26-0-169-86: task 0: Exited with exit code 1
676
+ 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-2_tp-4_pp-1_mbz-128/status.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ oom